I added an R9700 32GB to my 10+ year old desktop that had a 980 4GB card in it, for a grand total of $1350 or so. The payoff compared to what I was using with GHCP was 33 months, but when GHCP announced their price increase, it basically became a 3 month payoff at minimum (so yes, GHCP did a 10x price increase for non-parallel agentic workflows)
I can easily run Qwen3.6 35B-A3B with Q5_K_M with a 260k+ context window with some vram to spare. It easily runs probably 80tps. It took me quite a while to find the
Compared to GHCP Claude Sonnet 4.5 or 4.6, I have full parity. The wall clock time is faster for agentic workflows, and rule following is about on par.
With either, doing something kind of novel or obscure takes more hand holding compared to just generate a GUI or crud app. For example, trying to build an actual program that performs a complicated process correctly requires quite a bit of hand holding to get it to properly help.
Sure, it isn't Opus or something, but I think with the right harness, it probably can get close. I think most of the issues these days is the harnesses are lacking.
Lots of comments are expressing skepticism about compatibility but it's pretty cool how Nvidia has the clout to convince a bunch of game publishers and creative apps to release Arm versions. Popular games like League of Legends as well as stuff like Adobe Photoshop and Premiere are getting native Arm ports.
> Over 100 Windows software providers such as Adobe, Blackmagic Design, Blender, CapCut, ComfyUI and OTOY, and game developers such as KRAFTON, NetEase, Remedy Entertainment, Riot Games and XBOX are embracing the new RTX Spark platform. [...] NVIDIA is partnering with Adobe to rearchitect Adobe Premiere and Photoshop for RTX Spark. [0]
> Gaming on Arm is finally coming of age thanks to the NVIDIA partnership. Native anti-cheat solutions from Epic and BattlEye are fully supported on the RTX Spark platform. Major developers are jumping on board, with Riot Games bringing League of Legends and Valorant natively to the architecture, alongside KRAFTON bringing PUBG Battlegrounds. [1]
Also, Nintento Switch is an Nvidia/Arm gaming device so many game publishers already have some experience with the combo.
Quite a few of those already have arm ports for windows, and have since the 1st gen Snapdragon X Elite. I have the surface laptop 7 with that chip, and I remember it being made a big deal when photoshop & lightroom were ported. I believe Blender also had an arm build for windows a while ago too as did Davinci Resolve (as of 2024 I believe).
The big news is more so on the games side, which is probably where Nvidia had some pull.
I'm curious what "rearchitect for RTS Spark" means in practice though. Sounds like its less convincing them to make an arm build for windows, but they are maybe taking advantage of some hardware specific features? If so, what does that mean for the Snapdragon X series I wonder?
This thread is almost 1:1 identical to when Apple released their own silicon. This has the potential to be a worthy competitor for the Windows ecosystem, precisely because of NVidia’s moat as the grandparent pointed out.
Microsoft pulls in their weight as well, so this seems like it has a decent chance of getting industry support.
If you can get desktop RTX 5070 performance, oodles of (v)RAM, and minimal power usage out of a thin and light mobile device it's a win. This is change. If you can afford it.
Yes, there is a chance but it could also turn into another Itanium. Just because it is a superior product and backed by giants, doesn't necessarily guarantee success.
That said, Apple still deserves a lot of credit. They had a 5+ year edge, especially around the vision of tightly integrating the NPU and unified memory.
Like gaming consoles they calculated that unified memory will be cheaper for them in the long run. The funny thing is that while it gave them a unintended edge on local A.I, the "cheaper" calculation, didn't work out so well for them.
Looks like just rebranded DGX in laptop form, the biggest miss is the weak memory speed, 1/2 of the M5 laptop memory speed, and 1/3 of the M3 ultra that is now years old...
I'm not sure that's such a bad thing. It's not going to challenge the Apple M5, but if you're specifically looking for something in the "not-Mac" market, having a laptop-sized version of the DGX is probably going to be pretty successful.
Some competition for Apple in this space and competition for Intel and AMD is great.
But I really do question how well Windows on Arm is really going to work out long term.
For Apple it worked because they were able to force the issue. If you wanted a new Mac it was going to be Arm and we all knew eventually (this year or is it next year?) Intel support would drop. Over time we have seen M series exclusive features.
Developers were forced to update or abandon Mac which gave users a great experience (with some early growing pains).
This is something that Windows will never be able too do. They will always be stuck maintaining an emulator and a likely large subset of apps only supporting one over the other. (also does this work the other way around with an Arm only app working on x86?)
This seems like a repeat of when it was not uncommon for games to only support Intel or AMD or NVIDIA or AMD. But worse since they are not both x86. Sure at least we have emulation but just like with Rosetta2 it shouldn't ever be the long term solution.
For Apple it worked because they waited until they had a really, really good ARM ISA CPU (combined with arguably sandbagging their x86 offering for a few years prior but I digress).
Qualcomm is also working on a really good ARM ISA CPU with their acquisition of NuVia and subsequent Oryon architecture.
Meanwhile this is just using off-the-shelf ARM CPUs in a MediaTek SoC with blackwell bolted to the side of it. ARM's CPUs so far have been subpar for laptop-class chips. Hence why neither Apple nor Qualcomm are using them.
> tbh, I always read this as Intel doing some sales magic here.
Possibly, but Apple choosing a new, thicker chassis the same generation that they introduce their more power efficient replacement is certainly a thing. Even if Intel failed to achieve the TDP they told Apple, Apple also seems to no longer believe the thinness they were doing was viable for that TDP anyway.
Intel's product offering certainly wasn't as compelling towards the end there, but it also looked almost uniquely bad in Apple's chassis vs everyone else's
That's surely one thing, Apple went all-in on ARM, for Microsoft it's still a kinda "reduced experience".
But the bigger problem in my opinion: How much of the Windows userbase actually sticks to Windows because of its backwards-compatibility?
--> What would happen if they break this model and the OS is only judged based on its user experience and available applications...?
I'm not sure it would stand any chance to compete in the B2C space. If I think about it, there's not a single new feature in Windows of the last ~20 years I particularly care about.
Without backwards compatibility, there's barely any ecosystem. MacOS on the other hand is full of ecosystem features, improving collaboration, connectivity, handoff across devices, etc.
> MacOS on the other hand is full of ecosystem features, improving collaboration, connectivity, handoff across devices, etc.
True, but if you're only in the ecosystem as a mac user, in many ways it's felt like a mixed bag. I still wildly prefer mac over other operating systems, but if upgrades had a price, I think those sales would mostly go to iPhone users. Even at free, I'm yet to find a compelling reason to install Tahoe, and will probably just continue waiting until the next one.
Kinda underwhelming. I was hoping to see that they improved their memory bandwidth to move toward competing with the M5 Max. But this is more akin to the Strix Halo.
From what I'm reading it's probably the same chip that's used in the DGX Spark, the memory bandwidth at 300MB/s is equivalent to an M5 Pro, however you can't get an M5 Pro with 128GB of RAM. Apple pushes you to the biggest M5 Max chip, which at the 14 inch form factor, costs you $5099. You can get an ASUS GB10 machine with 2TB storage for $4000, so I guess the RTX Spark laptops will be more than that due to battery and screen, etc.
Perhaps the next generation of the spark will improve on the bandwidth and RAM size numbers. Yes it's a lot like a Strix Halo, but this has CUDA, which will be of interest to developers who want that.
I was looking for AMD AI Max+ 395 laptops recently, and the only ones I've found were 13 inch models, which seems odd from a heat dumping standpoint. I'm looking for 16 inches, I guess the 13 inch form factor would make it easy for commutes where you're taking it to dock to a large monitor at work or home, but no 14 inch screens?
I've tried the Z13 Flow and I actually like the form factor except for the folio keyboard. I especially like that, since it's a tablet, it vents hot air out the top instead of into your lap/table. But the whole driver situation was very weird and things would randomly stop working. That may have improved since I tried one ~1 year ago.
There are still a *lot* of sharp edges with the Spark: compatibility, overstated performance, power consumption/heat generation, etc. It's one thing to have that situation on a box explicitly aimed at developers and quite another with an actual consumer-focused laptop.
I don't think there's any incentive for Nvidia to make this a Windows-only device, so most likely it will be fully supported on Linux, just like their GPUs are.
What trouble? If you want a GPU that works on Linux, let alone FreeBSD, you buy nVidia, install their drivers and get on with your life (and sure, maybe you can't use Wayland, but why would you want to?). I'm all for open-source in theory, but in practice the AMD drivers cause far more trouble than the nVidia ones ever do.
Depends. It is the typical Nvidia problem. Everything is a black box but when it all works it is the best option available. But when it breaks, you hate them with a passion.
Sorry, but when it comes to chipsets, they're not even close.
The DGX Spark uses a GB10 with 128 GB unified LPDDR5X memory, while the DGX Station has a GB300 with 496 GB LPDDR5X (CPU) + 252 GB HBM3e (GPU) memory.
It's like Little League versus Major League, which is why the latter costs about 20 times more than the former.
The fact that both run Linux is just because they're part of the same DGX family.
Doesn't it come with Nvidia's blend of Ubuntu with a custom kernel? Do other distros work as well as "DGX OS" or are nvidia's kernel changes pretty important to have?
This plus the price different had me buy an AMD Strix Halo board last week. It seems the work with vLLM and training models could make the Spark worth the price difference, but before today's news I had the same thought about support and did not want to lock myself into a cool paperweight, especially with 128gb of RAM on the line. AMD is x86 and I can repurpose that or run Linux forever.
I've not noticed much in it that is NVIDIA specific.
But I would say that as an Ubuntu and Debian user for decades I have no incentive to use anything else on it and I'm just pleased to have a Linux on Aarch64 machine that is well supported for a change.
afaict, they have their own package repo mirrors and a few dedicated packages for nvidia stuff
tbh, I was rather unimpressed with the out-of-box experience for an "ai" computer, you couldn't even run a model locally with the common tools people use (no llama-cpp, ollama, vllm, etc). No huggingface CLI eiher, like come on!
I need to update that because I have a nice vllm setup on there now with 4 models running, but should be able to get anyone else going without having to muddle about as I did.
Honestly this looks like Microsoft must have thrown a pile of money at them to not mention it, as it's just too obviously the main question.
No one seriously cares about this running Windows. We want Steam and CUDA/Ollama, and Windows just gets in the way. nVidia are simply not that oblivious, but I have to admit in their position I'd have considered the Microsoft involvement more trouble than it's worth, which is among the many reasons I'm not a billionaire.
Maybe they think the RAM market is so terrible it will kill the whole initiative regardless.
I’ve read all the stuff about how llama.cpp is much faster and better than ollama, and i believe it - but good god llama.cpp isn’t user friendly.
You’d think in an era where “code is free” there would be an easier story around running local ai than compiling llama.cpp by hand and then spending hours researching flags - only for it to crash from an oom error every ten prompts or so.
You're supposed to use a cheap ChatGPT subscription to run optimization loops over llama.cpp flags with a self-contained reproducible benchmark script and just let it burn for hours/days until it is fully optimized ))))
Valve did that little more than a decade ago, the original Steam Machines. It didn't take, and despite the success of the Deck and current techy trends, Linux does not have the % to make the ROI worthwhile if it isn't simple for developers. Proton is a wedge in the door that will help Linux get there.
A potential change in Valve's culture/management aside, "let valve do the work" is a feature, not a bug. Studio spends all their budget targeting one platform (which still has ~90+% of the PC gaming market), and get Linux support for free.
Windows' monopoly on game dev isn't just market share either, since game dev isn't just code. You still need Photoshop, Maya, etc. and in smaller studies there's typically a crossover where some devs are doing art as well. Visual Studio's C++ debugger is still one of the best, and the tooling elsewhere hasn't caught up yet (compared to DX + PIX).
Then you also have to solve distribution and handling the fragmented display & audio stack. It's gotten a lot better, but its still a factor.
I'm fine with most of the work going into Wine/Proton. A stable ABI for Linux is a boon, if it happens to be Win32 then so be it.
In truth if AMD or nVidia put their mind to having decent profiling tooling on Linux, and the AI wave suggests they will have no option, then this could readily become a thing.
There are two new things being announced here: the GB10 chip being put into laptops, and GB10 running Windows. GB10 running Linux is not news, it's a product that's been shipping since last fall.
I think this is the first time an ARM windows device gets marketed for gaming. Would be interesting to see what kind of performance hit games have on the x86 to ARM translation layer.
Rosetta on Mac was obviously impressive. There was also impressive Arm->Intel translation in the mobile ecosystem at one time.
One reason it works surprisingly well on modern systems is how much is offloaded to the GPU. You aren't going to get great power optimization or anything without it being truly native though.
There are games which are CPU limited though, and it will be interesting how those do. Curiously those also tend to be in engines with Arm support already.
There was a presentation from Valve about their Dex compatibility layer. They did something that seems so obvious in retrospect.
When you lay out the software stack it is essentially OS > Game code > APIs. Both the OS and APIs are native code, it is only that middle point that needs the real work.
This is why x86 to ARM doesn't have such a heavy performance cost. So games can be CPU heavy but if it is heavy at the API end, that isnt a huge issue.
For Apple use of Rosetta 2 was only temporary as they moved whole lineup to ARM. MS would not abandon x64 anytime soon. So I'm guessing they will try hard to convince developers to release for both architectures.
I'm surprised they released this thing. Brand perception is probably a lot more important to Nvidia than whatever sales they could get from this thing, and if it's basically just DGX Spark, it's likely to underwhelm.
I've heard there's still a large backlog of both software problems, and hardware problems with the platform. The software problems could be fixed with time, but they'll still give a shitty first impression. I'd have thought Nvidia would just bury this and try again with a successor run of silicon with a new design.
This thing seems practically destined to just be a repeat of the Snapdragon laptop debacle.
DGX Spark runs Linux, and nobody is going to install Windows on that machine. This laptop got it backwards.
If someone decides to run Ollama for local inference with this laptop, they fit perfectly into the "has too much money to waste" bracket, which is addressed by a few other comments in the discussion.
I am wary of those ARM-based Windows machines because I am unsure how good the ongoing driver support for those SoCs will be. Will they even outlive the Windows version they currently ship with?
Looking at devices like the NVIDIA Shield gives me some hope that NVIDIA will be better than Qualcomm here. I just hope this is not a case where the OEM has to purchase X years of driver support from the chip vendor beforehand, and that NVIDIA will provide support directly itself.
This seems to be an attempt to compete with people running local models on Apple hardware—even though those local Mac Mini setups aren't really powerful.
I expect we'll get there in a few years, so perhaps this is Nvidia taking an early step in that direction.
In that case, this goes against Anthropic and OpenAI's business models. Which is a double whammy after Jensen Huang's recent comment about how agentic coding will only increase demand for software engineers, not reduce it.
So it also feels like a part of a budding shift in the competitive tension between the various parts of the AI supply chain.
Local AI was/is bound to happen, eventually. It'd be smart of Nvidia to get ahead of it.
Non-techy consumers may never do it, but at some point businesses are going to start asking when do they stop paying per token and start running models themselves. Right now the hardware is cost prohibitive, but I doubt that'll always be the case. Eventually the hardware will get cheaper and more available, and Nvidia seems to be betting on that.
They don't care where inference happens, so long as it happens on Nvidia hardware.
IMO it's only a matter of time before "self-hosting local AI" is as complicated as installing an app and clicking a download button.
And when that happens, the pitch to non-techy users is "Free ChatGPT you can use offline with zero privacy risk". Once hardware accessibility and LLM efficiency advance to the point that this becomes feasible, I suspect it'll result in a much bigger hit to the cloud AI market than many expect.
That workflow has been around for awhile now. I'm sure there are others but LM Studio has a model browser in app that effectively simplifies things to hitting download and hitting launch. The complexity tends to be in that there's a lot of models to choose from and also knowing how to set up whatever tool you're using with a local model. None of it's particularly hard, unless you start trying to customize settings.
I think the bigger hang up is that they're still slower and less capable than the frontier models, especially at the hardware specs most home users are likely to have.
LM Studio Link is brilliant, outside their central login/auth requirement. Tailscale is the backbone, I think, so it makes sense but I'm sure a method with wireguard could exist and enable similar performance.
the current dielmma for me is how do I install a model on a remote LM Studio device without bypassing Lm Studio to SSH or remote in?
> lms link [servername] get model ?
> lms get [servername] model ?
> lms get model --link [servername] ?
Maybe I need to read the docs again but I swear the only way is remote or go to that device and download via the GUI, ssh in and use the local cli.
Maybe can copy/paste from one device's downloads dir to the server? Maybe I need to try hosting models on my NAS and see if I can download from device 1 then run on device 2 without install/setup?
Why is it only a matter of time? The AI-as-a-service companies are going to continue to improve their products by improving both the part that could be reproduced in a self-hosted setup, but also the “secret sauce” they put on top of that to make it a better product. There is no incentive for this “secret sauce” to be something that can be reproduced for self-hosting, is there?
What secret sauce? We already have open source tooling for tool use, web browsing, and code execution/computer use. Open weight models will win in the end.
AIaaS might keep an edge with multi-modal agentic workflows, but for 80% of general use cases, no "secret sauce" needed, the open weight models are already there, and tooling is constantly getting better.
The bottleneck is the cost of local hardware right now.
The "secret sauce" is vendor lock-in. A textbook case is the vmware broadcom situation. Vmware was cheap so corporations found little reason to use open source. Broadcom made vmware expensive but now those corporations are finding out that it is a lot of work (aka expensive) to switch infrastructure.
I think a major incentive could be to sell hardware. If Apple is able to get their hands on a local LLM capable of covering a significant % of what people use ChatGPT for, the pitch they can offer is:
"Free, private, offline ChatGPT so long as your laptop has X GB of RAM"
Beyond that, I wouldn't underestimate the incentive of "because I can". The "secret sauce" you refer to is effectively just a DB & a while loop that feeds text to a bunch of tensors. If an indie dev decides they want to release something that dismantles the OpenAI & Anthropic moats, there really isn't all that big of a technical barrier stopping them.
LLM inference decode is heavily dependent on memory speed, not just having lots of memory. You can't say "X amount of ram" because the memory bandwidth on an M1 is 68.3 GB/s versus the 614 GB/s of an M5 Max, or a 4090's 1.01 TB/s over GDDR6X.
This basically creates a bottleneck at the oldest/cheapest Apple Silicon machines, which are already crippled for context prefill.
But honestly, obsoleting a huge number of otherwise great Apple Silicon machines is something Apple would moment consider a major "pro" of building a compelling local AI stack.
With how much speculation around the difficult time Apple has had getting people to upgrade from M1, I'm sure they'd jump at such an opportunity.
I'm from the times when you had to purchase a separate chip to perform floating point math. It was called a math co-processor. [1]
After a few generations (and over a decade) that was indistinguishable from the CPU chip itself.
It's a long hyperbole, I know, but I think local inference is inevitable; and the big fishes know it.
Will that be a complex technical setup? An appliance? An additional chip in your motherboard? So transparent it's burned right into the CPU? Those are just implementation details. We're probably just one generational breakthrough away from it.
I don't believe Anthropic and OpenAI are any more fearful of local AI than Google or Microsoft are of people hosting their own email.
Local AI capabilities are growing at a rapid pace, but so is hosted AI. While you can do a surprising amount of useful work with a model occupying a few to a few hundred gigs of VRAM, the hosted models are going to be way ahead for a long time.
- v4.5: 1x cost, 100% quality, 100% speed but maybe sometimes 80% speed because of load
- v4.6: 3x cost, 105% quality, 80% speed most of the time depends
- v4.7: 9x cost, 115% quality, 90% speed most of the time
Then people will either stick with v4.5 for everything it can do and, if knowledgeable, use v4.7+ for critical or specific tasks.
But if we add the option of:
LocalLLM: one time hardware + electricity cost, good enough quality for 90% of work, good enough speed for 90% of work, no vendor lock in/sudden cost spikes...
Then there is an edge to running it yourself unless you can burn investor cash to get to the next level.
I think the recent headlines on org token spend plus my own experience just today (June 1) with the new Copilot Pro limits is going to push those with the compute to run locally.
As of about 1pm today I did something to hit 47% of my entire June premium requests (copilot Pro, not converted).
As of 2pm I'm using Gemma 4 E4B on a 12gb GPU (with large context window) off my desktop to power VS Code with Copilot on my laptop. I'm going to build an AMD Strix Halo system next week when parts arrive so I can queue up a few models in parallel or work with something I need that much RAM for.
I'm not lifting the earth with my LLM setup. Gemma 4 E4B is solid for accelerating my current projects. and it's costing me pennies more per hour vs blowing half my Copilot Pro plan in a distracted morning.
I'm at a vendor conference this weekend that is showing off their Agent/Agentic workflows. Nobody can tell me how they balance the cost long term. Hopefully whoever the vendor is paying for their cloud LLM token usage doesn't spike cost in a year (or the vendor themselves) after companies convert and are trapped VMware style with these agent processes. You can bring your own (cloud) model subscription. I need to find out if we can point it back to our own local LLM endpoint and try local models for the same processes. Even if it takes 5x longer, it could be cheaper and more secure.
I fear the same thing, but still am unsure why or how :)
Google/Microsoft and hosting your own email is a byproduct of how difficult (socially, not technically) hosting your own email has become - mostly because SMTP protocol is inherently broken by spam and patched by social construct (trusted nodes, abuse@, 3+ DNS entries and counting, etc). Purely technical solutions, such HashCash etc, got discontinued in exchange for social ones. Central providers made (sometimes in exchange for, sometimes as excuse of, spam protection) self-hosting socially hard.
Now, I wonder if, and how, once Anthropic and OpenAI need to demonstrate profitability, could hamstring local AI. Which has been /so far/ very valuable for me in doing things that hosted providers don't want liability for, and align against (even if totally lawful and fair use!).
It's not even anything new, it's basically the mobile version of the DGX Spark. The two chips (N1X/GB10) are pretty similar in terms of architecture and specs. I don't get why this seems to be getting so much attention now.
But I like it. It's a copy of Apple's SoC design philosophy, same as AMD's Strix Halo, which I always thought was really cool both for laptops and home PCs. NVidia's traditional consumer cards pull way too much power and are too noisy to comfortably put them in a living or office environment.
That said, Apple's vertical integration is a massive competitive advantage here, IMO. Nvidia's reliance on Microsoft & Windows for software support likely makes competing w/ Apple an uphill battle.
If/when Local AI gets good enough to compete with Cloud AI on most inference workloads, Apple starts to look like Nvidia's biggest competitor.
While this is admittedly a dream scenario, the biggest downside would be Apple effectively having a monopoly in "Agent-ready" consumer electronics. Hopefully local AI both becomes the norm, and there is sufficient competition among the consumer platforms.
Side-note: I would love to see an "RTX Spark" Framework 13 mainboard at some point.
I don't understand this stance. Microsoft is reliant on Nvidia, they don't have a good ARM SOC to ship with without them. They will bend over backwards to accommodate these SOCs on Windows, and probably don't have much work to do in the first place.
Apple's vertical integration has led to a Siri overhaul that took half a decade to roll out, and it won't even run locally. They built an NPU coprocessor that's basically dark silicon for expensive inference, and then shipped MLX to stop Tensorflow and Pytorch from replacing Apple's role in the stack entirely. Mac owners are pleading for signed CUDA drivers for the PCIe or Thunderbolt in their $5,000+ Mac Pros. Apple's ecosystem is pure liability for AI, they're not moving any product for datacenter inference and can't even sell the hardware to themselves: https://9to5mac.com/2026/03/02/some-apple-ai-servers-are-rep...
Nvidia's profit margins are safe. Even if the RTX Spark is a completely failed product, Apple is not encroaching on the markets that Nvidia dominates.
Fair points all around. Ultimately it all comes down to execution.
In theory, Apple SHOULD have an advantage given they have everything they need in house and can all pull in a unified direction. In practice, it's not always the case that all the teams in a large corporation are all that much better at pulling in the same direction than multiple different corporations in a partnership. And all this will be moot if Local LLMs never catch up to cloud LLMs in terms of quality.
Regardless, it'll be very interesting to see how Nvidia's partnerships with Microsoft & hardware OEMs play out. If the AI inference compute share shifts appreciably to local consumer hardware, I'll want to see strong competition.
I'd argue that Apple had the upper hand, but they folded super early. They abandoned OpenCL, which was the most promising CUDA competitor with industry-wide buy in from dozens of companies. Then they transitioned to an ecosystem-first mindset prevented Apple from cooperating to take down Nvidia, and their locked-down software stopped the industry's first high-speed ARM servers from reaching their audience. Nvidia capitalized on both opportunities to the tune of trillions in valuation.
Without Khronos involved, I don't think that Apple has the buy-in to create a real industry-scale CUDA alternative. At this point, it might just be most profitable to support CUDA in macOS and give the people what they want.
You can do a lot with existing devices in a medium to decent gaming PC (or probably phone/laptop, I haven't tried.) I think HN tends to skew toward only thinking of LLM as useful for coding, but they are very useful for many non-coding things, and existing local LLMs are quite capable. I imagine it won't be long before apps with LLM-based features will try to run locally first and fall back to cloud LLMs just to save token costs. Actually I'd be surprised if some apps aren't doing this already.
definitely! it has the advantage that it can run CUDA kernels but on the other hand it has lower memory bandwidth and probably loses a token/s fight for many LLMs.
> In that case, this goes against Anthropic and OpenAI's business models. Which is a double whammy after Jensen Huang's recent comment about how agentic coding will only increase demand for software engineers, not reduce it.
The writing is on the wall, neither Anthropic nor OpenAI are anywhere near close to sustainability and if one or, worse, both fail the entire demand bubble for NVDA crashes.
It's smart to set up alternative destination markets while they can do so in peace.
The RTX GPU laptops run very hot. Even though they are pound for pound better, it’s just runs too hot for local llm usage for me at least. Prefer Macs for this. A lot of AMD cards also run cooler. I wonder if undervting would help with smaller models and heat.
I mean the GB10 is pretty efficient for the power it has, but imho is nowhere near the power efficiency of Apple Silicon (it was never intended to be a chip used for mobile devices). I guess this is kind of the movement Apple did with the A12Z and the Mini but... the other way around?
I think its gonna be another failure as we are used to see with the PC market these days.
It is great for inference for single user/single session. it is not replacement for graphical accelerator, that run several concurrent inference sessions in parallel.
Basically the same tradeoff as macmini with unified memory.
It's probably more that LLM inference speed comes from having a large amount of fast RAM. And fast RAM is brutally expensive right now.
At this point, your cost-efficient options include used 3090s, "frankenrigs" using recycled data center cards, and a handful of "workstation" class cards, where the originally high margins and the long enterprise purchasing cycles have kept prices from going up too fast.
In contrast, a lot of these "personal" AI systems are basically a GPU-like core wired to larger amounts of slow RAM. Which is still semi-affordable. Generally speaking, they make for OK chatbots but extremely slow coding agents. Whereas you can run a modestly useful coding agent at reasonable speed on a 3090.
So yeah, a lot of these systems are bit scammy. But not because it's a secret conspiracy to protect data center cards. Rather, there simply isn't enough fast RAM in the entire world. So they'll flog you disappointly slow RAM instead.
TL;dr: Might be useful for some use cases, but benchmark very carefully.
For anyone curious to know how this will fare against Macbooks, at least in CPU perf: DGX Spark has the exact same GPU and CPU as the top RTX Spark laptops will, so you can just directly compare from that.
Of course, DGX Spark is a miniPC, so laptops will likely be slower due to power limits/throttling.
I think DGX Spark has poor memory bandwidth because these laptops were the plan all along. NVIDIA didn't want to commit to the extra costs of a 512-bit memory bus for their first laptop SoC, so they went with the more modest 256-bit bus, same as AMD did for Strix Halo.
The GB10 itself is pretty good and I love using mine for broad Linux development. But it's too expensive for consumer level pricing, and even for the "prosumer" the price is pretty stiff. Even if they dropped the CX-7 and halfed the RAM and shipped a smaller hard drive, would it be below, say, $2500 USD? I guess we'll see, but this variant is coming out pretty late so maybe it's just best to wait for the 2nd generation.
This feels like getting a foot in the door to ensure Apple doesn't entirely eat Nvidia's lunch if AI inference workloads start to shift from cloud to local.
With MLX, Apple is building an answer to CUDA, and if people start switching from ChatGPT & Claude to some app that runs on their M5, suddenly Apple starts to look like Nvidia's biggest competitor.
If Nvidia doesn't have a pathway towards getting hardware into the hands of consumers, it could be a really difficult road ahead for them.
Apple seems to still own the creative space. If those tools are able to run local models for any AI workflows suddenly anthropic/etc could lose a massive segment. Or at least demonstrate to others wanting a slice of the cloud AI profits it can be done.
I'm here for it. Local models can do a lot of what I need at almost no cost, plus the fun of making them work better or building a new system to handle that aspect of my home lab. A Strix Halo system may not be amazingly fast but at 128gb of RAM it can keep up with most open models worth exploring.
Based on June 1 Copilot Pro plan premium token burn and cost, unless you REALLY know how to use cloud AI efficiently and are tooled up to do so a local LLM on hardware you may already own is very appetizing.
I converted a lot of work today to a 6.5gb local LLM on a 12gb GPU and no, it's not as good. But it is 'free' or at least feels that way, especially when I need to redo something and my copilot premium request % doesn't change.
We'll need to wait for the benchmarks, but this looks great! Windows 11 ARM64 is already amazing, and if these really are an upgrade from the Qualcomm chips we're going to have even better laptops on the market.
With 128GB ram, the price tag would be pretty high. And lots of application does not work Windows on Arm. Even Microsoft provides something like Rosetta 2 for windows, still x86 architecture would be the most popular one for Windows for a looong time.
Saying that I think this is product is kinda dead on arrival.
+ battery too. I've wondered if a mini pc with battery would make for a good form factor. I often move between places where I have a desk with a screen but still use a laptop because I want to just suspend and resume. If a mini pc had a small battery just to hold its RAM while suspended I could move between places and just plug in a single USB-C cable and have my full workstation up and running. The thermals could be better than in a laptop and having a built-in UPS better than with a desktop. But last time I checked no one packaged things like that.
There's the Khadas Mind series of mini pcs. They have a proprietary docking interface though. Agree that it would be great if this form-factor was more common.
They didn't say that Mediatek made the cpu sores. Grace is NVidia's own cpu arm cores. I bet that Mediatek made other parts of SoC necessary for a notebook
Well, MediaTek actually said they made most of the SoC in fact. But the actual CPU cores themselves are all but certainly off-the-shelf Cortex parts, since MediaTek doesn't have a custom core design at all afaik.
NVIDIA hasn't done custom CPU cores for anything they've yet branded "Grace". The original Grace data center CPU (paired with the Hopper data center GPU) used ARM Neoverse V2 cores. The "GB10" chip shipped in DGX Spark and announced here for RTX Spark uses Cortex X925 and Cortex A725 CPU cores.
Physically, NVIDIA did the GPU chiplet and Mediatek did the other chiplet that has the CPU, DRAM controller, and IO.
GB300 is nominally "available" in desktop form factor workstations priced around $100k. That's a few orders of magnitude away from the ordinary desktop PC market that consumers participate in.
I really hope these take off and succeed and they support Linux. Qualcomm is seriously holding back the Linux ARM adoption with their continuous missteps.
Unfortunately in the current market 32GB of ddr5 seems to run about $400 as 2x16gb DIMMS, and even more for 1x32GB DIMM (higher density chips are more expensive). So $600 really isn't much over market price, especially considering strix halo uses 8000MHz ram instead of the typical 6000 found in consumer dimms.
I didn't see this in the article but elsewhere I've seen the memory bandwidth quoted as 600GB/s [1]. For comparison:
- 5090/6000 Pro: 1792GB/s
- 5080:: 960GB/s
- 5070Ti: 892GB/s
- M3 Ultra: 819GB/s
- DGX Spark: 273GB/s (less than an M5 Pro at 307GB/s)
Memory bandwidth isn't everything but it will cap inference rate pretty heavily. Also, the M3 Ultra is for an almost 2 year old Mac Studio. It's widely expected that it'll be refreshed in Q3 with a likely M5 or M4 Ultra with >1000GB/s. I really hope Apple realizes what a market opportunity Apple has here.
The above shows just how good value the 5090 really is. It basically a RTX 6000 Pro with less RAM (and ~12% fewer CUDA units), which is a ~$10k card, for 20-30% of the price. This also demonstrates how NVidia uses VRAM for market segmentation. As an aside, the true data center cards (eg B100, H100) use HBM memory at ~3.2TB/s.
> tl;dr - For software development, Qwen3.6 27B, 5090 gives you ~3x speed over M5 Max, letting you plow through code, while M5 Max gives you ~4x memory, letting you use higher quantization and bigger context. Which would you choose and why?
I've read a number of things from which the consensus seems to be that yes you can run a larger model and/or have more context with a 128GB+ Mac but the performance gap is still massive and with current hardware we're still talking about inference rates that matter. By this I mean there's a big difference between 10tok/s vs 30. Once we get to t apoint where it's 100 vs 300, it won't be as big of a deal, a bit like FPS in games.
Oh and there are similar concerns with the DGX Spark [2].
But probably worth clarifying it's not a typical "MediaTek CPU" some might assume by that. It has Nvidia's customized ARM CPU implementation + their GPU.
I’m getting more and more convinced that we will end up running LLMs in our personal computers. Which makes me wonder where Anthropic/OpenAIs moats will come from.
1. in order to run LLMs, especially the best ones, you need complicated devices which are expensive
2. if you buy one for your personal use, you are probably not going to utilize it all the time and it will be idle a lot
It seems to me that it will always be more economical that the LLM-running devices are in a datacenter where it is easier to make sure they are always utilized
If a model is substantially better than most humans at most tasks, the human isn't going to be able to perceive the difference between Claude Opus 7.7 and 8.7. Humans at some point aren't going to be able to perceive the difference on benchmarks either, because they are going to get wildly abstract.
AI vendors are really going to struggle to shift tokens far beyond the frontier of human capabilities. It's reasonable (not guaranteed) to assume that, if the trend of frontier models (doubling capabilities on benchmarks every n months) holds, then the same trend will hold for local models, and those local models will meet and exceed the perception frontier. This would mean a human cannot tell the difference between Mistral-Open-2030 and Claude Opus 2030.
That's a bunch of "ifs", but there's nothing exceptional about those "ifs". They're basically the scenario if nothing changes between now and ~2030 with regards to capabilities trend attainment.
The trend over the past three decades of personal computing has been for devices to become exponentially more powerful regardless of the actual computing needs of users. The excess computing power has famously been requested by projects such as SETI@Home and Folding@Home, and been exploited by bad actors for crypto mining. The most basic laptop today used only for web browsing and word processing would be a powerful workstation 20 years ago, when the most basic laptop was also used only for web browsing and word processing (and arguably for more things, as it was all mostly local software).
There is no ceiling to the power of consumer hardware. If it's cheap enough, it will be bought.
I think you missed the point of my message. Web browsing still happens by connecting to data centres, so why are consumer laptops so much more (unnecessarily) powerful today than they were 20 years ago? All the more so given that, at that time, you were running MS Office locally rather than using Office 365 or Google Docs remotely.
Even two or three years people were pointing out "The ChatGPT subscriptions you can buy with $2000 give you much more compute than whatever home setup you come up with" on r/LocalLLM. I did my own elementary school maths and came to the same conclusion.
Yet till this day people still boast how their beefy M4 Pro/Max machine with 32+GB RAM (which is not at all a "normal person's setup" and costs $2000+) runs LLMs smoothly, and "that's the future".
Someone needs to re-learn basic maths and take a walk around Best Buy to understand what "consumer laptop" looks like.
If there end up being useful workflows where you keep stuff running in the background or overnight that's one advantage, compared to a data center that might cut off your access during peak hours or etc.
Think of it like having a graphics card at home versus using a cloud gaming stream? Technically subscribing to GeForce is much cheaper up front than getting a card, but people still do that. So will the audience of people running agents at home be as large as PC gaming? I think that's kind of plausible.
>2. if you buy one for your personal use, you are probably not going to utilize it all the time and it will be idle a lot
I think consumers are primed for that type of behaviour though. I have an iPhone on my desk. It has something like 2-3tflops CPU+GPU, which is double that of the largest super computer on earth when Jurassic Park came out, and is probably more computing power than existed on earth when I was born in the 80s.
I use this device for around 1hr per day to write text messages.
It's inevitable. What might be a prosumer device today priced at 4000$ will be a regular consumer device in 10 years and models only get better.
Local models today are fine for a lot of mundane tasks and will continue to be so. The use cases where paying for frontier models is worth it, will continue to shrink for folks not doing frontier work.
While I agree with that in principle, it is very worrisome that the prices of personal computers, especially of any personal computer that is not a big desktop, have been increasing continuously.
The price of a mini-PC with Intel Panther Lake is at least double in comparison with the price of a mini-PC with Arrow Lake H having similar specifications, and I am talking about barebones, before adding DRAM and SSDs, whose prices have risen even more.
The rise in prices is somewhat obfuscated by the confusing names of CPUs, i.e. some old and new CPUs may seem to be at similar prices and they have similar names, but the new CPU actually corresponds to a lower segment of the market, by having e.g. a smaller GPU and a lower clock frequency, while the CPU model that really corresponds to the old is named such that it seems to belong to the class corresponding to its present price.
As a concrete example of this obfuscation, which may confuse the buyers of laptops or mini-PCs, I have an ASUS 15 Pro with "Core Ultra 5 225H". If I would buy an ASUS 16 Pro now, the corresponding CPU model, the cheapest which is not worse than what I have, would be "Core Ultra X7 358H".
The best open weight LLMs don't run on this computer, or almost any consumer grade computer. Even the memory requirement for Gemma 4 is out of reach for most consumers (by which I mean those who are not on HN). Unless there is some magic that would make high quality LLMs consume no more than 8GB RAM which makes them usable on a 16GB laptop (which is the norm these days), "local LLM for personal computing" is mostly just a myth.
We're hitting the atomic limits of what's possible with minimum feature size in silicon. It's also very hard to remove 1 kW of heat from a laptop, let alone do it quietly or on battery.
i think a lot of that is for government and enterprise use. even for personal computers themselves (i.e.: laptops) they're usually loss leaders, they don't turn profit. You can run a server (and many do) on laptops, but that didn't replace cloud services or server hosting. You can't store enormous amounts of data on your laptop/phone for the llm to use, or access tools the app dev wouldn't want exposed on untrusted devices.
The whole replacing people angle is just the short term use case the more ghoulish executives are thinking about. In practice, lots of lots of new use cases have been made possible by LLMs. A lot of which can be done locally. But whatever capacity you have locally, they can have more of and for cheaper, and they manage the model instead of you doing it yourself. I think you put it nicely though, their moat will be thinned, and I doubt they'll be as profitable as their funding suggests, but at the same time the demand for them won't go away either. I don't know if OpenAI and Anthropic will be viable, but I'm nearly certain Deepseek is.
The tipping point will be power usage, if a local llm can run the same workload for less power that would be a game changer. Nvidia might get decimated, but even Google and others have moved on from GPUs already, they have faster and more power efficient TPUs. Add to that network bandwidth and availability issues, their moat remains. Also consider that even for graphics capabilities, user devices just don't have a consistent spec to make things like widespread 3d graphics and webgl usage viable. Someone's cheap android phone will never run a local llm reliably,same as it won't a 3d game. even if they have a high-end iphone, network providers aren't always performant as they are in western countries, and then there are people that won't want to install your app or local software, and then browser based exposure of the capability to sites which will have similar hardware spec issues, OS instabilities, competing tabs,etc...
No thunderbolt is a big no for me. Its one of the greatest feature of MacbookPro that makes it dockable and expandable as a desktop with a good thunderbolt dock.
Well, it was only a matter of time, since both AMD and now Intel are now switching to APUs. Nvidia could either cede the desktop GPU market to them, going all-in into AI datacenter chips, or it could challenge them.
Maybe the Nth time's the charm and Microsoft+Nvidia will manage to make Windows on ARM a viable platform.
Looks like the MSI one might be a 2-in-1, if it has good stylus support I might have a good candidate for an upgrade, thought my ~3-4 year old Galaxy Book is holding up alright for now.
can these do training or only inference? currently working on learning machine learning and I'd love to have a physical machine I could aim to build real workloads on in a few years.
It’s possible (likely, even) to have a chip fast enough for inference, but not fast enough or with enough memory to do meaningful training runs. Like the current DGX spark.
Unified RAM means its soldered to the mainboard, right?
I'm not sure if I like this. Sure for a laptop this might be not a big problem but if this ARM ecosystem is a success it will spread to desktop computers and I fear we could lose the existing modularity.
I have no idea how powerful or power efficient these guys are, but this seems to be the first step in a bigger push towards Windows on ARM (without loosing gaming).
I think more announcements will follow soon from other companies.
It's worth noting that Nvidia power management on Linux has been absymal. There also aren't any of the usual power management options to see how much power things are using, which is quite atypical for a modern system.
Nvidia really threw stuff over the wall with the DGX Spark release. They don't seem to really care. I sort of think they'll spend a little more time on Windows, where there's no pesky upstreaming to do and they can just do whatever, but man, it's such typical hubris from Nvidia to build such an expensive box with good chips but make it basically unsupportable and roasty hot all the time.
You also generally have to run an ever more stale two year old Ubuntu derived DGX OS to get anywhere, with bespoke kernel and drivers all. None of it is well supported, none of it just works like a comparable PC or even well behaved arm system would.
As for other ARM, there were rumors AMD Sound Wave is/was going to be a ~10W arm APU, but there hasn't been much said about it lately. Honestly given the ram crunch, it's maybe just not worth trying to build a system with a cheap core, if the rest of your costs are going to stay so stratospheric.
https://www.techpowerup.com/341848/amd-sound-wave-arm-powere...
What is this product anyway? Is it a general purpose CPU or is it specifically designed for MS Windows? Nvidia stepping back from the open source?
"Introducing the NVIDIA RTX Spark™ Superchip. The fusion of NVIDIA AI and RTX graphics in a single chip redefines Windows PCs and delivers amazing creating, AI development, and gaming—on the slimmest, most beautiful RTX laptops ever and small, ultra-efficient desktops."
Its nvidia attempt to gain additional market share and expected as well. If the whole ecosystem is around nvidia and its the easiest way of running stuff, Nvidia offering more enterprise infrastrcuture allows companies to just buy directly nvidia.
Nvidia is also very very rich and pushes the boundaries of stuff. They stoped waiting for industry standards. You can see this in there network stuff. All nvidia.
Next logical step (at least now, not something i thought about) was there CPU for their GPU racks/clusters/systems.
Now they have everything anyway, RTX Spark is just logical.
I don't think its specificly targeted at Apple at all.
Apple has like 10-15% market share and just because some IT nerds buy themselves a mac mini doesn't mean much.
Plenty of them actually just run openclaw without local models. Something which surprised me quite a lot.
But i have two 4090 at home. They consume a lot of power and i had to research the proper Mainboardmodel and had to mod one 4090 to use water cooling because they run too hot.
There Spark setup was at 3k, way to expensive for normal people. If they can get this down and sell more, great for their ecosystem (strengthening it) and getting more money from people.
It does surprise me though that they have enough capacity for this chip and not just putting everyting in Rubin but perhaps the build out has slowed down a little or they start to diverse already for economic savety
All the news articles in my feed mentioned Nvidia reinventing personal computing which is laughable given the specs are worse than the m series. I’m guessing they saw how well Apple devices were selling and rushed to get something similar out so they can ride the hype train and have something to fall back on if ai DC spend slows down.
There's a lot of companies trying to support datacenter systems like GH and Rubin that don't have dev hardware remotely resembling it. M-series isn't a good option, speaking from the personal experience of currently using one for this exact purpose.
I wouldn't say it's Nvidia stepping back from open source... if anything this is doubling down on it, as one of the selling points of this is the 128GB of unified memory which will allow for hosting local models (i.e, nvidia's new open model they just released). I guess it's pretty cool, I'm a big supporter of local LLMs/open weight models so seems enticing to me, although I'm not sure this will be super applicable to a lot of regular consumers. Seems like a pretty niche product.
I really like this, but I think the reason Apple Silicon took off was that Apple sort of forced devs to support ARM. Not sure if Microsoft can do the same for Windows…
Developers weren’t really “forced” to support ARM. They simply recognized that all future Macs would be ARM, whereas most new PCs would continue to run on x86. So the incentive to adopt ARM was much weaker on the PC side.
After nvidia's many years of neglecting Linux, paired with direct Microsoft's involvement? Are we going to trust them, to allow installing Linux in these easily?
It ships with DGX OS 7, which includes Ubuntu's 24.04 repos. It is not using mainline Ubuntu, and if you want to run Ubuntu 26.04, you'll have to do some work.
The thing I think is really funny is that if this takes off, frontier model companies and datacenters will end up holding the bag, and as per usual after the last few tech hype cycles, NVIDIA will still be selling.
Eventually a lot of inference will get right-sized into something you affordably run yourself.
> "Our goal is to deliver unmetered intelligence to every home and every desk with Windows," said Satya Nadella, chairman and head of Microsoft.
Then:
> However, Ian Fogg, Research Director at industry analyst firm FDM CCS Insight said the change was "likely to come with a significant price tag" and Nvidia would be targeting "those looking for workstation-class performance".
First, make it possible. Then, expand the market. The early adopters help pay R&D for later efforts. Every desk is a good goal, even if not hit by the first doodad.
It just feels too much like what they said about Apple II and early Windows. A play at nostalgia instead putting real thought into it.
I was an engineer at both MS and Apple, and wholeheartedly agree with you.
My question is, what happens to the people who use RTX cards for gaming? This new solution isn't meant for that. Do they need an "AI accelerator" and a gaming-centric GPU?
ARM64+GPU sure seems like the future. I'm still using my M1 and even that can handle models well, has decent graphics, M5 is a beast, and M6 must surely go even bigger on LLM compute. Now Microsoft has a compelling ARM64+GPU future too.
Strix halo's 8060S gpu is very weak, and is roughly equivalent to a 4060 laptop GPU, whereas GB10's gpu is equivalent to a desktop 5070. For LLM throughput, tok/s is similar due to bottleneck by memory bandwidth, but the GB10 has 3x faster prefill. People have also been able to squeeze out much better performance on GB10 using NVFP4 and other improvements in the months after the DGX Spark launch, so don't be misled by early lackluster benchmarks. For the RTX Spark, which also targets gaming and creative applications, the 3x faster GPU is quite nice.
I feel like the shape of the market right now for "home lab" inference is:
The sparks are good if your ultimate plan is to spend even more on NVidia hardware in future to run your dev setups at usable speeds. Or, you're developing for a work cluster.
If you mainly want to run local models at acceptable speeds portably, buy a mac with lots of RAM. If you’re happy with non-portable / racked, buy 3090s (dense) or mac studios (MoEs). Buy newer cards if you are restricted on power or slots. If you are rich, buy a6000 blackwells.
The only Question is is it worth suffering hip and x86? I suspect a lot of folks might like a machine that mimics their GB300 But costs less than a dgx.
Also I heard the tensor core instructions on the dgx are gimped and you’re better off with a rtx pro x000. Is that the same with these machines?
Is CUDA really a lead for long? Aren’t all the latest competitive approaches avoiding all the standard software stacks and writing deeply customized software that is very directly tied to whatever hardware they use?
And is it really a way to lock in people? With AI coding tools, isn’t it trivial to write software on top of CUDA and rewrite it to target some other hardware?
It all sounds good on paper. But I have trouble believing Windows can be a good platform for this. Microsoft has lost all trust after inserting ads into windows, slowly removing power user features, and exploiting every dark pattern they can. And for years, the ARM based Windows laptops have been useless due to app compatibility issues. Why would this change now? Is it priced to be a lot cheaper than Apple’s laptops? Or is this a niche product for AI developers basically?
Anecdotally Windows ARM works fine for me, although to be honest most of my work is command line + browser anyway. WSL works like a treat. Steam installs and most lower end games also play fine on my ARM laptop too. Games that require kernel anticheat don't work.
I think they make a great "second device" where you have something meatier to fall back to if something doesn't quite work right. I'm not sure if it's ready to take on the "main device" role just yet. But it's a far far better experience than the Surface RT days.
The "gaming" take is a strange one indeed for an ARM platform. Hopefully they (Microsoft or Nvidia?) put some real effort into the translation layer. They claim modern AAA games, but it is possible they strongarmed the developers to make them an ARM build for a few select titles...
It's clear gaming was not a major concern, it's just "good enough" for someone running AI models and occasionally wants to play some games, not made to primarily play games.
Yep. I noticed the press releases talk about all the partners they have. It seems like a desperate attempt to manufacture a consensus to invest in this new hardware instead of leaving it sort of abandoned like the other Windows ARM stuff. But the problem is that these attempts end up having a few very visible apps working on the architecture and others not actually doing anything substantial.
Sure the graphics capabilities are probably very good. But if you’re a game developer who has traditionally built on Windows on x86 chips, would you want to invest in this new chip or invest in making games for the Apple ecosystem? Aren’t there more new customers to reach in the Apple world than this new Nvidia world?
> But if you’re a game developer who has traditionally built on Windows on x86 chips, would you want to invest in this new chip or invest in making games for the Apple ecosystem?
Windows and the new chip. Higher developer productivity and higher chances of a substantial audience.
Who cares about Windows, the goal is to run local AI models similar to AMD Strix Halo and Apple Silicon machines. The OS is honestly a distant last concern as long as the models work well, as you could put Linux on these too, but not sure how well wake lock works.
I would never trust Microsoft. Their next drama is revoking Office 2019 perpetual licenses https://www.youtube.com/watch?v=KRnno9VIZx0. It never ends with them because they know they have you by the balls.
This may finally be the chip family ARM on Windows has always needed. Qualcomm's chips have always been dogs with slow off-the-shelf ARM CPU cores that have pathetic single-threaded performance compared to x86 AMD/Intel or ARM Apple Silicon designs.
For reference, this is just a single benchmark, but as an idea of each vendor's top mobile CPU single-threaded performance:
Geekbench Single Thread Score:
- DGX Spark (same CPU as RTX Spark): 3125
- Snapdragon X1 Elite: 2950
- Snapdragon X2 Elite Extreme: 4050
- AMD Ryzen 9 9955HX: 3225
- Intel Core Ultra 9 290HX Plus: 3175
- Apple M5 Max: 4350
I'm happy to be wrong about Qualcomm's latest X2 chip performance, even if it is shipping in only a single product so far. Their previous best was the lowest in this list.
Qualcomm Snapdragon x1 and upcoming x2 use their Oryon core and have much faster single-thread performance than Intel/Amd and this nvidia soc that uses off-the-shelf arm cores
That wasn't true of the X1, but apparently the X2 (which is only in a single device so far) does appear to finally be fast. The first Windows ARM CPU to be faster than any of its x86 rivals. Competitive with Apple Silicon single-thread performance even.
I was disappointed to see that the RTX Spark has the ARM cores from the DGX Spark. I was hoping it had their new in-house developed cores that Nvidia is starting to use on their latest gen server parts. They look really fast. That said, if RTX Spark has CPU performance like the DGX Spark, it will be almost as fast as the top AMD/Intel parts.
It won't, the top tier RTX Spark has the same exact CPU and GPU as DGX Spark, so you can check DGX Spark CPU benchmarks to see how it fares. Spoiler: it's about M3 Max level. And they're only coming this fall.
>Lenovo, HP, Dell and Apple accounted for almost 75% of the world's PC market in the first three months of this year, according to research firm Gartner.
I added an R9700 32GB to my 10+ year old desktop that had a 980 4GB card in it, for a grand total of $1350 or so. The payoff compared to what I was using with GHCP was 33 months, but when GHCP announced their price increase, it basically became a 3 month payoff at minimum (so yes, GHCP did a 10x price increase for non-parallel agentic workflows)
I can easily run Qwen3.6 35B-A3B with Q5_K_M with a 260k+ context window with some vram to spare. It easily runs probably 80tps. It took me quite a while to find the
Compared to GHCP Claude Sonnet 4.5 or 4.6, I have full parity. The wall clock time is faster for agentic workflows, and rule following is about on par.
With either, doing something kind of novel or obscure takes more hand holding compared to just generate a GUI or crud app. For example, trying to build an actual program that performs a complicated process correctly requires quite a bit of hand holding to get it to properly help.
Sure, it isn't Opus or something, but I think with the right harness, it probably can get close. I think most of the issues these days is the harnesses are lacking.
Lots of comments are expressing skepticism about compatibility but it's pretty cool how Nvidia has the clout to convince a bunch of game publishers and creative apps to release Arm versions. Popular games like League of Legends as well as stuff like Adobe Photoshop and Premiere are getting native Arm ports.
> Over 100 Windows software providers such as Adobe, Blackmagic Design, Blender, CapCut, ComfyUI and OTOY, and game developers such as KRAFTON, NetEase, Remedy Entertainment, Riot Games and XBOX are embracing the new RTX Spark platform. [...] NVIDIA is partnering with Adobe to rearchitect Adobe Premiere and Photoshop for RTX Spark. [0]
> Gaming on Arm is finally coming of age thanks to the NVIDIA partnership. Native anti-cheat solutions from Epic and BattlEye are fully supported on the RTX Spark platform. Major developers are jumping on board, with Riot Games bringing League of Legends and Valorant natively to the architecture, alongside KRAFTON bringing PUBG Battlegrounds. [1]
Also, Nintento Switch is an Nvidia/Arm gaming device so many game publishers already have some experience with the combo.
[0] https://nvidianews.nvidia.com/news/nvidia-microsoft-windows-...
[1] https://www.windowslatest.com/2026/06/01/microsoft-builds-it...
Quite a few of those already have arm ports for windows, and have since the 1st gen Snapdragon X Elite. I have the surface laptop 7 with that chip, and I remember it being made a big deal when photoshop & lightroom were ported. I believe Blender also had an arm build for windows a while ago too as did Davinci Resolve (as of 2024 I believe).
The big news is more so on the games side, which is probably where Nvidia had some pull.
I'm curious what "rearchitect for RTS Spark" means in practice though. Sounds like its less convincing them to make an arm build for windows, but they are maybe taking advantage of some hardware specific features? If so, what does that mean for the Snapdragon X series I wonder?
Press releases are easy. Delivering on promises, not so much.
This thread is almost 1:1 identical to when Apple released their own silicon. This has the potential to be a worthy competitor for the Windows ecosystem, precisely because of NVidia’s moat as the grandparent pointed out.
Microsoft pulls in their weight as well, so this seems like it has a decent chance of getting industry support.
If you can get desktop RTX 5070 performance, oodles of (v)RAM, and minimal power usage out of a thin and light mobile device it's a win. This is change. If you can afford it.
Yes, there is a chance but it could also turn into another Itanium. Just because it is a superior product and backed by giants, doesn't necessarily guarantee success.
Not sure how it's comparable to Itanium at all? ARM is not a new architecture. It's not even a new architecture for Windows.
That said, Apple still deserves a lot of credit. They had a 5+ year edge, especially around the vision of tightly integrating the NPU and unified memory.
Like gaming consoles they calculated that unified memory will be cheaper for them in the long run. The funny thing is that while it gave them a unintended edge on local A.I, the "cheaper" calculation, didn't work out so well for them.
In what way hasn't it worked out for Apple? Some of their products are totally sold out.
They are limiting sku's on everything but highest end. The ramocalips is hitting high end RAM, especially hard.
Will this push even more games into Linux?
You know most of them already have arm versions...
Somewhere, a monkey’s paw must have curled its finger.
Apple and Steam have been successfully applying pressure for years. Who's willfully staying behind at this point?
Looks like just rebranded DGX in laptop form, the biggest miss is the weak memory speed, 1/2 of the M5 laptop memory speed, and 1/3 of the M3 ultra that is now years old...
I'm not sure that's such a bad thing. It's not going to challenge the Apple M5, but if you're specifically looking for something in the "not-Mac" market, having a laptop-sized version of the DGX is probably going to be pretty successful.
Some competition for Apple in this space and competition for Intel and AMD is great.
But I really do question how well Windows on Arm is really going to work out long term.
For Apple it worked because they were able to force the issue. If you wanted a new Mac it was going to be Arm and we all knew eventually (this year or is it next year?) Intel support would drop. Over time we have seen M series exclusive features.
Developers were forced to update or abandon Mac which gave users a great experience (with some early growing pains).
This is something that Windows will never be able too do. They will always be stuck maintaining an emulator and a likely large subset of apps only supporting one over the other. (also does this work the other way around with an Arm only app working on x86?)
This seems like a repeat of when it was not uncommon for games to only support Intel or AMD or NVIDIA or AMD. But worse since they are not both x86. Sure at least we have emulation but just like with Rosetta2 it shouldn't ever be the long term solution.
For Apple it worked because they waited until they had a really, really good ARM ISA CPU (combined with arguably sandbagging their x86 offering for a few years prior but I digress).
Qualcomm is also working on a really good ARM ISA CPU with their acquisition of NuVia and subsequent Oryon architecture.
Meanwhile this is just using off-the-shelf ARM CPUs in a MediaTek SoC with blackwell bolted to the side of it. ARM's CPUs so far have been subpar for laptop-class chips. Hence why neither Apple nor Qualcomm are using them.
"Sandbagging their x86 offering" is a new one. There's no winning.
> arguably sandbagging their x86 offering
tbh, I always read this as Intel doing some sales magic here.
Apple: "Hey, we're making a product that has a 15w thermal envelope, do you have anything?"
Intel: "Yes!"
(Unspoken: their products will throttle down to fit, in fact, they will try to run always at 99ºC so you always get the best performance! FEATURE!)
Apple: "uhhhh..."
Consumers: "HEH IS IT EVEN A PRO DEVICE IF IT DOESN"T HAVE <INTEL MARKETING BRAND TERM>?"
Apple: "UHHHH... Guess we'll do it ourselves"
> tbh, I always read this as Intel doing some sales magic here.
Possibly, but Apple choosing a new, thicker chassis the same generation that they introduce their more power efficient replacement is certainly a thing. Even if Intel failed to achieve the TDP they told Apple, Apple also seems to no longer believe the thinness they were doing was viable for that TDP anyway.
Intel's product offering certainly wasn't as compelling towards the end there, but it also looked almost uniquely bad in Apple's chassis vs everyone else's
That's surely one thing, Apple went all-in on ARM, for Microsoft it's still a kinda "reduced experience".
But the bigger problem in my opinion: How much of the Windows userbase actually sticks to Windows because of its backwards-compatibility?
--> What would happen if they break this model and the OS is only judged based on its user experience and available applications...?
I'm not sure it would stand any chance to compete in the B2C space. If I think about it, there's not a single new feature in Windows of the last ~20 years I particularly care about.
Without backwards compatibility, there's barely any ecosystem. MacOS on the other hand is full of ecosystem features, improving collaboration, connectivity, handoff across devices, etc.
> MacOS on the other hand is full of ecosystem features, improving collaboration, connectivity, handoff across devices, etc.
True, but if you're only in the ecosystem as a mac user, in many ways it's felt like a mixed bag. I still wildly prefer mac over other operating systems, but if upgrades had a price, I think those sales would mostly go to iPhone users. Even at free, I'm yet to find a compelling reason to install Tahoe, and will probably just continue waiting until the next one.
I feel like making universal binaries a thing, and pushing for it to be standard is one viable path.
They already kind of are with ARM64EC, however Windows ecosystem isn't macOS, unless there is market pressure, most shops will keep doing x86/x64.
Microslop doesn’t want people to be able to run their binaries elsewhere, it’s the only reason people buy their product.
They also buy it, because to this day most people cannot buy GNU/Linux powered laptops on the stores they usually buy their computers from.
They only know Apple, Windows and Chromebooks.
Kinda underwhelming. I was hoping to see that they improved their memory bandwidth to move toward competing with the M5 Max. But this is more akin to the Strix Halo.
From what I'm reading it's probably the same chip that's used in the DGX Spark, the memory bandwidth at 300MB/s is equivalent to an M5 Pro, however you can't get an M5 Pro with 128GB of RAM. Apple pushes you to the biggest M5 Max chip, which at the 14 inch form factor, costs you $5099. You can get an ASUS GB10 machine with 2TB storage for $4000, so I guess the RTX Spark laptops will be more than that due to battery and screen, etc.
Perhaps the next generation of the spark will improve on the bandwidth and RAM size numbers. Yes it's a lot like a Strix Halo, but this has CUDA, which will be of interest to developers who want that.
I was looking for AMD AI Max+ 395 laptops recently, and the only ones I've found were 13 inch models, which seems odd from a heat dumping standpoint. I'm looking for 16 inches, I guess the 13 inch form factor would make it easy for commutes where you're taking it to dock to a large monitor at work or home, but no 14 inch screens?
I've tried the Z13 Flow and I actually like the form factor except for the folio keyboard. I especially like that, since it's a tablet, it vents hot air out the top instead of into your lap/table. But the whole driver situation was very weird and things would randomly stop working. That may have improved since I tried one ~1 year ago.
128 GB memory is also lame. I'm hankering for a windows equivalent of the mac studio that came with 512 GB.
There are still a *lot* of sharp edges with the Spark: compatibility, overstated performance, power consumption/heat generation, etc. It's one thing to have that situation on a box explicitly aimed at developers and quite another with an actual consumer-focused laptop.
Can it work with Linux? That's all I care about.
An nvidia engineer in a discord server said it should work fine
I don't think there's any incentive for Nvidia to make this a Windows-only device, so most likely it will be fully supported on Linux, just like their GPUs are.
> just like their GPUs are
So with proprietary blobs that give you more trouble that they're worth?
Those blobs are worth $5T; show some respect.
Kids these days, amirite?
What trouble? If you want a GPU that works on Linux, let alone FreeBSD, you buy nVidia, install their drivers and get on with your life (and sure, maybe you can't use Wayland, but why would you want to?). I'm all for open-source in theory, but in practice the AMD drivers cause far more trouble than the nVidia ones ever do.
Depends. It is the typical Nvidia problem. Everything is a black box but when it all works it is the best option available. But when it breaks, you hate them with a passion.
They worked with Microsoft to make this commercially viable. That’s possibly reason enough.
I wouldn't trust it to have good upstream support. It's Nvidia. So not really interested.
Sort of. It's the same chipset as in the DGX Spark & DGX Station, which run Ubuntu (NVIDIA's flavor).
Sorry, but when it comes to chipsets, they're not even close. The DGX Spark uses a GB10 with 128 GB unified LPDDR5X memory, while the DGX Station has a GB300 with 496 GB LPDDR5X (CPU) + 252 GB HBM3e (GPU) memory. It's like Little League versus Major League, which is why the latter costs about 20 times more than the former. The fact that both run Linux is just because they're part of the same DGX family.
I took "same" to mean "compatible driver and software stack" not "compute perf", rightly or wrongly.
DGX Spark comes with linux out of the box, it would be hard to imagine this device is not also compatible
Doesn't it come with Nvidia's blend of Ubuntu with a custom kernel? Do other distros work as well as "DGX OS" or are nvidia's kernel changes pretty important to have?
Hopefully better than support on their Jetson or orin boards, where compiling anything is hard because of the outdated stack.
This plus the price different had me buy an AMD Strix Halo board last week. It seems the work with vLLM and training models could make the Spark worth the price difference, but before today's news I had the same thought about support and did not want to lock myself into a cool paperweight, especially with 128gb of RAM on the line. AMD is x86 and I can repurpose that or run Linux forever.
I've not noticed much in it that is NVIDIA specific.
But I would say that as an Ubuntu and Debian user for decades I have no incentive to use anything else on it and I'm just pleased to have a Linux on Aarch64 machine that is well supported for a change.
For some value of "well supported" - NVIDIA's own internal catalogs (libraries, NIMs, etc) are still spotty on aarch64 coverage.
afaict, they have their own package repo mirrors and a few dedicated packages for nvidia stuff
tbh, I was rather unimpressed with the out-of-box experience for an "ai" computer, you couldn't even run a model locally with the common tools people use (no llama-cpp, ollama, vllm, etc). No huggingface CLI eiher, like come on!
I did put together my eventual setup in a repo https://github.com/verdverm/sparky
I need to update that because I have a nice vllm setup on there now with 4 models running, but should be able to get anyone else going without having to muddle about as I did.
Honestly this looks like Microsoft must have thrown a pile of money at them to not mention it, as it's just too obviously the main question.
No one seriously cares about this running Windows. We want Steam and CUDA/Ollama, and Windows just gets in the way. nVidia are simply not that oblivious, but I have to admit in their position I'd have considered the Microsoft involvement more trouble than it's worth, which is among the many reasons I'm not a billionaire.
Maybe they think the RAM market is so terrible it will kill the whole initiative regardless.
You misspelled llama.cpp
I’ve read all the stuff about how llama.cpp is much faster and better than ollama, and i believe it - but good god llama.cpp isn’t user friendly.
You’d think in an era where “code is free” there would be an easier story around running local ai than compiling llama.cpp by hand and then spending hours researching flags - only for it to crash from an oom error every ten prompts or so.
You're supposed to use a cheap ChatGPT subscription to run optimization loops over llama.cpp flags with a self-contained reproducible benchmark script and just let it burn for hours/days until it is fully optimized ))))
WSL is the answer in what most folks are concerned.
Has Steam finally started to push for native Linux games instead of translating Windows ones?
Valve did that little more than a decade ago, the original Steam Machines. It didn't take, and despite the success of the Deck and current techy trends, Linux does not have the % to make the ROI worthwhile if it isn't simple for developers. Proton is a wedge in the door that will help Linux get there.
It is simple, Android NDK has all the same APIs for 3D rendering and audio, as do all major middleware engines.
The failure of business, only reinforces Windows as the platform most studios reach for.
Buy Windows, buy Visual Studio, pay game engines licenses, let Valve do the work.
This ignoring that current Valve's management doesn't live forever, so who knows what happens afterwards.
A potential change in Valve's culture/management aside, "let valve do the work" is a feature, not a bug. Studio spends all their budget targeting one platform (which still has ~90+% of the PC gaming market), and get Linux support for free.
Windows' monopoly on game dev isn't just market share either, since game dev isn't just code. You still need Photoshop, Maya, etc. and in smaller studies there's typically a crossover where some devs are doing art as well. Visual Studio's C++ debugger is still one of the best, and the tooling elsewhere hasn't caught up yet (compared to DX + PIX).
Then you also have to solve distribution and handling the fragmented display & audio stack. It's gotten a lot better, but its still a factor.
I'm fine with most of the work going into Wine/Proton. A stable ABI for Linux is a boon, if it happens to be Win32 then so be it.
At this point Valve look more capable of running a platform business than Microsoft do.
Microsoft have spent the whole Nadella era in "oooo cloud" inspired wonder and actively screwed up everything else.
> Valve's management doesn't live forever, so who knows what happens afterwards.
Tens of thousands of Windows games would remain playable with ubiquitous Vulkan-capable hardware and a 500mb Proton runtime?
If it runs faster than the windows ones, who cares?
The game developers that use Windows, with Visual Studio, to develop such games.
This is, admittedly, the great anomaly.
In truth if AMD or nVidia put their mind to having decent profiling tooling on Linux, and the AI wave suggests they will have no option, then this could readily become a thing.
This is strangely absent from the news.
There are two new things being announced here: the GB10 chip being put into laptops, and GB10 running Windows. GB10 running Linux is not news, it's a product that's been shipping since last fall.
It's a collaboration with Microsoft so going to say no, probably not.
I think this is the first time an ARM windows device gets marketed for gaming. Would be interesting to see what kind of performance hit games have on the x86 to ARM translation layer.
Rosetta on Mac was obviously impressive. There was also impressive Arm->Intel translation in the mobile ecosystem at one time.
One reason it works surprisingly well on modern systems is how much is offloaded to the GPU. You aren't going to get great power optimization or anything without it being truly native though.
There are games which are CPU limited though, and it will be interesting how those do. Curiously those also tend to be in engines with Arm support already.
There was a presentation from Valve about their Dex compatibility layer. They did something that seems so obvious in retrospect.
When you lay out the software stack it is essentially OS > Game code > APIs. Both the OS and APIs are native code, it is only that middle point that needs the real work.
This is why x86 to ARM doesn't have such a heavy performance cost. So games can be CPU heavy but if it is heavy at the API end, that isnt a huge issue.
Very cool.
Apple Silicon has a special mode that modified how the ARM chip handles memory transactions to be like x86. Does this nvidia ARM have the same?
What would be interesting to me would be how quickly developers start targeting ARM64 directly.
For Apple use of Rosetta 2 was only temporary as they moved whole lineup to ARM. MS would not abandon x64 anytime soon. So I'm guessing they will try hard to convince developers to release for both architectures.
I'm surprised they released this thing. Brand perception is probably a lot more important to Nvidia than whatever sales they could get from this thing, and if it's basically just DGX Spark, it's likely to underwhelm.
I've heard there's still a large backlog of both software problems, and hardware problems with the platform. The software problems could be fixed with time, but they'll still give a shitty first impression. I'd have thought Nvidia would just bury this and try again with a successor run of silicon with a new design.
This thing seems practically destined to just be a repeat of the Snapdragon laptop debacle.
I cannot think why someone would run those workflows on a Windows laptop, unless someone has way too much money to spend.
> someone has way too much money to spend.
that's what nvidia is hoping for
If the workload is offloaded to the chip, why would the host platform matter?
Lots of machine learning workflows support Linux better than Windows, if they run on Windows at all. (e.g. https://docs.vllm.ai/en/latest/getting_started/quickstart/ )
DGX Spark runs Linux, and nobody is going to install Windows on that machine. This laptop got it backwards.
If someone decides to run Ollama for local inference with this laptop, they fit perfectly into the "has too much money to waste" bracket, which is addressed by a few other comments in the discussion.
WSL
It often works, but you always lose something compared to native Linux.
Believe it or not, Windows (WSL) is the best Linux distro and Nvidia knows that.
vllm-windows works well enough
I am wary of those ARM-based Windows machines because I am unsure how good the ongoing driver support for those SoCs will be. Will they even outlive the Windows version they currently ship with?
Looking at devices like the NVIDIA Shield gives me some hope that NVIDIA will be better than Qualcomm here. I just hope this is not a case where the OEM has to purchase X years of driver support from the chip vendor beforehand, and that NVIDIA will provide support directly itself.
I would love a RTX Spark Shield. ;p
This seems to be an attempt to compete with people running local models on Apple hardware—even though those local Mac Mini setups aren't really powerful.
I expect we'll get there in a few years, so perhaps this is Nvidia taking an early step in that direction.
In that case, this goes against Anthropic and OpenAI's business models. Which is a double whammy after Jensen Huang's recent comment about how agentic coding will only increase demand for software engineers, not reduce it.
So it also feels like a part of a budding shift in the competitive tension between the various parts of the AI supply chain.
Local AI was/is bound to happen, eventually. It'd be smart of Nvidia to get ahead of it.
Non-techy consumers may never do it, but at some point businesses are going to start asking when do they stop paying per token and start running models themselves. Right now the hardware is cost prohibitive, but I doubt that'll always be the case. Eventually the hardware will get cheaper and more available, and Nvidia seems to be betting on that.
They don't care where inference happens, so long as it happens on Nvidia hardware.
IMO it's only a matter of time before "self-hosting local AI" is as complicated as installing an app and clicking a download button.
And when that happens, the pitch to non-techy users is "Free ChatGPT you can use offline with zero privacy risk". Once hardware accessibility and LLM efficiency advance to the point that this becomes feasible, I suspect it'll result in a much bigger hit to the cloud AI market than many expect.
That workflow has been around for awhile now. I'm sure there are others but LM Studio has a model browser in app that effectively simplifies things to hitting download and hitting launch. The complexity tends to be in that there's a lot of models to choose from and also knowing how to set up whatever tool you're using with a local model. None of it's particularly hard, unless you start trying to customize settings.
I think the bigger hang up is that they're still slower and less capable than the frontier models, especially at the hardware specs most home users are likely to have.
LM Studio Link is brilliant, outside their central login/auth requirement. Tailscale is the backbone, I think, so it makes sense but I'm sure a method with wireguard could exist and enable similar performance.
the current dielmma for me is how do I install a model on a remote LM Studio device without bypassing Lm Studio to SSH or remote in?
> lms link [servername] get model ?
> lms get [servername] model ?
> lms get model --link [servername] ?
Maybe I need to read the docs again but I swear the only way is remote or go to that device and download via the GUI, ssh in and use the local cli.
Maybe can copy/paste from one device's downloads dir to the server? Maybe I need to try hosting models on my NAS and see if I can download from device 1 then run on device 2 without install/setup?
Why is it only a matter of time? The AI-as-a-service companies are going to continue to improve their products by improving both the part that could be reproduced in a self-hosted setup, but also the “secret sauce” they put on top of that to make it a better product. There is no incentive for this “secret sauce” to be something that can be reproduced for self-hosting, is there?
What secret sauce? We already have open source tooling for tool use, web browsing, and code execution/computer use. Open weight models will win in the end.
AIaaS might keep an edge with multi-modal agentic workflows, but for 80% of general use cases, no "secret sauce" needed, the open weight models are already there, and tooling is constantly getting better.
The bottleneck is the cost of local hardware right now.
The "secret sauce" is vendor lock-in. A textbook case is the vmware broadcom situation. Vmware was cheap so corporations found little reason to use open source. Broadcom made vmware expensive but now those corporations are finding out that it is a lot of work (aka expensive) to switch infrastructure.
I think a major incentive could be to sell hardware. If Apple is able to get their hands on a local LLM capable of covering a significant % of what people use ChatGPT for, the pitch they can offer is:
"Free, private, offline ChatGPT so long as your laptop has X GB of RAM"
Beyond that, I wouldn't underestimate the incentive of "because I can". The "secret sauce" you refer to is effectively just a DB & a while loop that feeds text to a bunch of tensors. If an indie dev decides they want to release something that dismantles the OpenAI & Anthropic moats, there really isn't all that big of a technical barrier stopping them.
LLM inference decode is heavily dependent on memory speed, not just having lots of memory. You can't say "X amount of ram" because the memory bandwidth on an M1 is 68.3 GB/s versus the 614 GB/s of an M5 Max, or a 4090's 1.01 TB/s over GDDR6X.
This basically creates a bottleneck at the oldest/cheapest Apple Silicon machines, which are already crippled for context prefill.
Thanks for clarifying -- I was oversimplifying.
But honestly, obsoleting a huge number of otherwise great Apple Silicon machines is something Apple would moment consider a major "pro" of building a compelling local AI stack.
With how much speculation around the difficult time Apple has had getting people to upgrade from M1, I'm sure they'd jump at such an opportunity.
this might be a way for Apple to milk product revenue for many years.
- Please buy our new Macbook pro M5 that gives you 20 tokens/s on local 80B LLM
next year - Please buy our new Macbook pro M6 that gives you 25 tokens/s on local 80B LLM
milking product revenue in perpetuity by offering meaningful marginal improvements, while keeping same architecture will be the golden goose for Apple
+plus if it allows to segment market by wallet size into poor/middle/rich classes, thats even better
I'm from the times when you had to purchase a separate chip to perform floating point math. It was called a math co-processor. [1]
After a few generations (and over a decade) that was indistinguishable from the CPU chip itself.
It's a long hyperbole, I know, but I think local inference is inevitable; and the big fishes know it.
Will that be a complex technical setup? An appliance? An additional chip in your motherboard? So transparent it's burned right into the CPU? Those are just implementation details. We're probably just one generational breakthrough away from it.
[1] https://en.wikipedia.org/wiki/X87
Like the math co-processor it might end up just being new instructions for the cpu to handle ai related math.
I think non-techy users will get subsidized hardware with compute workloads running in the background on idle to recover the cost (and lots of ads).
> Non-techy consumers may never do it
They will. As some point in the future, people will want everything, they'll prompt full movies because they're bored and want to watch something.
You’re assuming that owning compute will be possible.
I don't believe Anthropic and OpenAI are any more fearful of local AI than Google or Microsoft are of people hosting their own email.
Local AI capabilities are growing at a rapid pace, but so is hosted AI. While you can do a surprising amount of useful work with a model occupying a few to a few hundred gigs of VRAM, the hosted models are going to be way ahead for a long time.
If it's something like:
- v4.5: 1x cost, 100% quality, 100% speed but maybe sometimes 80% speed because of load - v4.6: 3x cost, 105% quality, 80% speed most of the time depends - v4.7: 9x cost, 115% quality, 90% speed most of the time
Then people will either stick with v4.5 for everything it can do and, if knowledgeable, use v4.7+ for critical or specific tasks.
But if we add the option of:
LocalLLM: one time hardware + electricity cost, good enough quality for 90% of work, good enough speed for 90% of work, no vendor lock in/sudden cost spikes...
Then there is an edge to running it yourself unless you can burn investor cash to get to the next level.
I think the recent headlines on org token spend plus my own experience just today (June 1) with the new Copilot Pro limits is going to push those with the compute to run locally.
As of about 1pm today I did something to hit 47% of my entire June premium requests (copilot Pro, not converted).
As of 2pm I'm using Gemma 4 E4B on a 12gb GPU (with large context window) off my desktop to power VS Code with Copilot on my laptop. I'm going to build an AMD Strix Halo system next week when parts arrive so I can queue up a few models in parallel or work with something I need that much RAM for.
I'm not lifting the earth with my LLM setup. Gemma 4 E4B is solid for accelerating my current projects. and it's costing me pennies more per hour vs blowing half my Copilot Pro plan in a distracted morning.
I'm at a vendor conference this weekend that is showing off their Agent/Agentic workflows. Nobody can tell me how they balance the cost long term. Hopefully whoever the vendor is paying for their cloud LLM token usage doesn't spike cost in a year (or the vendor themselves) after companies convert and are trapped VMware style with these agent processes. You can bring your own (cloud) model subscription. I need to find out if we can point it back to our own local LLM endpoint and try local models for the same processes. Even if it takes 5x longer, it could be cheaper and more secure.
I fear the same thing, but still am unsure why or how :)
Google/Microsoft and hosting your own email is a byproduct of how difficult (socially, not technically) hosting your own email has become - mostly because SMTP protocol is inherently broken by spam and patched by social construct (trusted nodes, abuse@, 3+ DNS entries and counting, etc). Purely technical solutions, such HashCash etc, got discontinued in exchange for social ones. Central providers made (sometimes in exchange for, sometimes as excuse of, spam protection) self-hosting socially hard.
Now, I wonder if, and how, once Anthropic and OpenAI need to demonstrate profitability, could hamstring local AI. Which has been /so far/ very valuable for me in doing things that hosted providers don't want liability for, and align against (even if totally lawful and fair use!).
It's not even anything new, it's basically the mobile version of the DGX Spark. The two chips (N1X/GB10) are pretty similar in terms of architecture and specs. I don't get why this seems to be getting so much attention now.
But I like it. It's a copy of Apple's SoC design philosophy, same as AMD's Strix Halo, which I always thought was really cool both for laptops and home PCs. NVidia's traditional consumer cards pull way too much power and are too noisy to comfortably put them in a living or office environment.
One can only hope.
That said, Apple's vertical integration is a massive competitive advantage here, IMO. Nvidia's reliance on Microsoft & Windows for software support likely makes competing w/ Apple an uphill battle.
If/when Local AI gets good enough to compete with Cloud AI on most inference workloads, Apple starts to look like Nvidia's biggest competitor.
While this is admittedly a dream scenario, the biggest downside would be Apple effectively having a monopoly in "Agent-ready" consumer electronics. Hopefully local AI both becomes the norm, and there is sufficient competition among the consumer platforms.
Side-note: I would love to see an "RTX Spark" Framework 13 mainboard at some point.
I don't understand this stance. Microsoft is reliant on Nvidia, they don't have a good ARM SOC to ship with without them. They will bend over backwards to accommodate these SOCs on Windows, and probably don't have much work to do in the first place.
Apple's vertical integration has led to a Siri overhaul that took half a decade to roll out, and it won't even run locally. They built an NPU coprocessor that's basically dark silicon for expensive inference, and then shipped MLX to stop Tensorflow and Pytorch from replacing Apple's role in the stack entirely. Mac owners are pleading for signed CUDA drivers for the PCIe or Thunderbolt in their $5,000+ Mac Pros. Apple's ecosystem is pure liability for AI, they're not moving any product for datacenter inference and can't even sell the hardware to themselves: https://9to5mac.com/2026/03/02/some-apple-ai-servers-are-rep...
Nvidia's profit margins are safe. Even if the RTX Spark is a completely failed product, Apple is not encroaching on the markets that Nvidia dominates.
Fair points all around. Ultimately it all comes down to execution.
In theory, Apple SHOULD have an advantage given they have everything they need in house and can all pull in a unified direction. In practice, it's not always the case that all the teams in a large corporation are all that much better at pulling in the same direction than multiple different corporations in a partnership. And all this will be moot if Local LLMs never catch up to cloud LLMs in terms of quality.
Regardless, it'll be very interesting to see how Nvidia's partnerships with Microsoft & hardware OEMs play out. If the AI inference compute share shifts appreciably to local consumer hardware, I'll want to see strong competition.
I'd argue that Apple had the upper hand, but they folded super early. They abandoned OpenCL, which was the most promising CUDA competitor with industry-wide buy in from dozens of companies. Then they transitioned to an ecosystem-first mindset prevented Apple from cooperating to take down Nvidia, and their locked-down software stopped the industry's first high-speed ARM servers from reaching their audience. Nvidia capitalized on both opportunities to the tune of trillions in valuation.
Without Khronos involved, I don't think that Apple has the buy-in to create a real industry-scale CUDA alternative. At this point, it might just be most profitable to support CUDA in macOS and give the people what they want.
You can do a lot with existing devices in a medium to decent gaming PC (or probably phone/laptop, I haven't tried.) I think HN tends to skew toward only thinking of LLM as useful for coding, but they are very useful for many non-coding things, and existing local LLMs are quite capable. I imagine it won't be long before apps with LLM-based features will try to run locally first and fall back to cloud LLMs just to save token costs. Actually I'd be surprised if some apps aren't doing this already.
Might be aimed at people who spec out the $5100 Macbook Pros with M5 Maxes and 128GB.
definitely! it has the advantage that it can run CUDA kernels but on the other hand it has lower memory bandwidth and probably loses a token/s fight for many LLMs.
> In that case, this goes against Anthropic and OpenAI's business models. Which is a double whammy after Jensen Huang's recent comment about how agentic coding will only increase demand for software engineers, not reduce it.
The writing is on the wall, neither Anthropic nor OpenAI are anywhere near close to sustainability and if one or, worse, both fail the entire demand bubble for NVDA crashes.
It's smart to set up alternative destination markets while they can do so in peace.
Love seeing AMD forcing Novideo to catch up for once rather than the other way around.
So they have basically reused the same hardware as in the DGX Spark (GB10)... That chip isn't great for LLM inference actually.
https://www.techpowerup.com/gpu-specs/gb10.c4342 https://www.nvidia.com/en-us/products/rtx-spark/
The RTX GPU laptops run very hot. Even though they are pound for pound better, it’s just runs too hot for local llm usage for me at least. Prefer Macs for this. A lot of AMD cards also run cooler. I wonder if undervting would help with smaller models and heat.
I mean the GB10 is pretty efficient for the power it has, but imho is nowhere near the power efficiency of Apple Silicon (it was never intended to be a chip used for mobile devices). I guess this is kind of the movement Apple did with the A12Z and the Mini but... the other way around?
I think its gonna be another failure as we are used to see with the PC market these days.
It is great for inference for single user/single session. it is not replacement for graphical accelerator, that run several concurrent inference sessions in parallel.
Basically the same tradeoff as macmini with unified memory.
>That chip isn't great for LLM inference actually.
Why do I have the feeling it's been intentionally made to be bad in order to get you on to their most pensive datacenter gear.
It's probably more that LLM inference speed comes from having a large amount of fast RAM. And fast RAM is brutally expensive right now.
At this point, your cost-efficient options include used 3090s, "frankenrigs" using recycled data center cards, and a handful of "workstation" class cards, where the originally high margins and the long enterprise purchasing cycles have kept prices from going up too fast.
In contrast, a lot of these "personal" AI systems are basically a GPU-like core wired to larger amounts of slow RAM. Which is still semi-affordable. Generally speaking, they make for OK chatbots but extremely slow coding agents. Whereas you can run a modestly useful coding agent at reasonable speed on a 3090.
So yeah, a lot of these systems are bit scammy. But not because it's a secret conspiracy to protect data center cards. Rather, there simply isn't enough fast RAM in the entire world. So they'll flog you disappointly slow RAM instead.
TL;dr: Might be useful for some use cases, but benchmark very carefully.
Oh, btw, we are only making 10 of these, the rest of our capacity has been sold off to the large AI firms.
For anyone curious to know how this will fare against Macbooks, at least in CPU perf: DGX Spark has the exact same GPU and CPU as the top RTX Spark laptops will, so you can just directly compare from that.
Of course, DGX Spark is a miniPC, so laptops will likely be slower due to power limits/throttling.
DGX Spark is really poor at inference due to the memory bandwidth so hopefully they’ve fixed that before touting this as a way to run local models.
I think DGX Spark has poor memory bandwidth because these laptops were the plan all along. NVIDIA didn't want to commit to the extra costs of a 512-bit memory bus for their first laptop SoC, so they went with the more modest 256-bit bus, same as AMD did for Strix Halo.
UEFI, display panels, wifi, storage controllers, etc would be what I'm worried about. I doubt Microsoft is going to make it easy.
DGX Spark is also $4700
It's been almost 30 years, and a single letter changed. When will we get the Sparkstation, the UltraSpark and the SuperSpark?
SuperSpark and then UltraSpark. And then we can get SparkCube, Sparkii, and SparkiiU.
I'm personally waiting for the OpenSpark.
https://en.wikipedia.org/wiki/OpenSPARC
I'm waiting for the AllSpark
And the top tier QuantumSpark?
With competition from the MegaSpark and SparkGenesis.
you forgot Xerox sPARC with guis, ethernet, laser printers, ...
Will we get enterprise ready open firmware too instead of this "we missed DOS so we invented UEFI" for boot firnware?
The GB10 itself is pretty good and I love using mine for broad Linux development. But it's too expensive for consumer level pricing, and even for the "prosumer" the price is pretty stiff. Even if they dropped the CX-7 and halfed the RAM and shipped a smaller hard drive, would it be below, say, $2500 USD? I guess we'll see, but this variant is coming out pretty late so maybe it's just best to wait for the 2nd generation.
This feels like getting a foot in the door to ensure Apple doesn't entirely eat Nvidia's lunch if AI inference workloads start to shift from cloud to local.
With MLX, Apple is building an answer to CUDA, and if people start switching from ChatGPT & Claude to some app that runs on their M5, suddenly Apple starts to look like Nvidia's biggest competitor.
If Nvidia doesn't have a pathway towards getting hardware into the hands of consumers, it could be a really difficult road ahead for them.
Apple seems to still own the creative space. If those tools are able to run local models for any AI workflows suddenly anthropic/etc could lose a massive segment. Or at least demonstrate to others wanting a slice of the cloud AI profits it can be done.
I'm here for it. Local models can do a lot of what I need at almost no cost, plus the fun of making them work better or building a new system to handle that aspect of my home lab. A Strix Halo system may not be amazingly fast but at 128gb of RAM it can keep up with most open models worth exploring.
Based on June 1 Copilot Pro plan premium token burn and cost, unless you REALLY know how to use cloud AI efficiently and are tooled up to do so a local LLM on hardware you may already own is very appetizing.
I converted a lot of work today to a 6.5gb local LLM on a 12gb GPU and no, it's not as good. But it is 'free' or at least feels that way, especially when I need to redo something and my copilot premium request % doesn't change.
Awesome, won't be buying it all at current prices but once they calm down, I will very much like to get one.
Around 2-3K USD something with a good GPU + CPU + 128GB of integrated RAM is just going to be an awesome experience.
Considering Mac options are north of 5K+ even on a regular day.
DGX Spark is $4700, so I kind of doubt that RTX Spark's top configs will be cheaper than that.
The DGX also contains the 200 GbE networking and linux support.
The ConnectX 7 2x200 Gbps networking card in the DGX Spark alone is worth $700
To be fair the connectx-7 in the spark can't even push 2x200 Gbps since it is connected via 4 pcie lanes.
Technically it's connected via 8 PCIe gen 5 lanes (two 4x connections), allowing ~100Gbps per port.
Thanks for the correction. I should have looked it up; I only remembered it being somewhat odd.
Laptops will also have to contain a much tighter configuration, display, keyboard, camera, etc ;)
there is desktop variant as well
isn't dgx ai first and rtx prosumer first. I think it will be cheaper longer term not atm with component inflation
We'll need to wait for the benchmarks, but this looks great! Windows 11 ARM64 is already amazing, and if these really are an upgrade from the Qualcomm chips we're going to have even better laptops on the market.
With 128GB ram, the price tag would be pretty high. And lots of application does not work Windows on Arm. Even Microsoft provides something like Rosetta 2 for windows, still x86 architecture would be the most popular one for Windows for a looong time.
Saying that I think this is product is kinda dead on arrival.
Is this just dgx spark, but a laptop?
yes, same chip
+ Windows
+ Screen
- ConnectX-7 Smart NIC
+ battery too. I've wondered if a mini pc with battery would make for a good form factor. I often move between places where I have a desk with a screen but still use a laptop because I want to just suspend and resume. If a mini pc had a small battery just to hold its RAM while suspended I could move between places and just plug in a single USB-C cable and have my full workstation up and running. The thermals could be better than in a laptop and having a built-in UPS better than with a desktop. But last time I checked no one packaged things like that.
There's the Khadas Mind series of mini pcs. They have a proprietary docking interface though. Agree that it would be great if this form-factor was more common.
> - ConnectX-7 Smart NIC
Can the link type be toggled between Ethernet and Infiniband? (Don't think I've ever heard of a laptop with IB.)
What about the desktop version? It seemed like it is not a dgx since it has the CPUs cores done by mediatek
The DGX Spark/GB10 has CPU cores from Mediatek (in a pretty odd cluster configuration, too).
They didn't say that Mediatek made the cpu sores. Grace is NVidia's own cpu arm cores. I bet that Mediatek made other parts of SoC necessary for a notebook
MediaTek said MediaTek made the CPU: https://www.mediatek.com/press-room/mediatek-collaborates-wi...
Well, MediaTek actually said they made most of the SoC in fact. But the actual CPU cores themselves are all but certainly off-the-shelf Cortex parts, since MediaTek doesn't have a custom core design at all afaik.
NVIDIA hasn't done custom CPU cores for anything they've yet branded "Grace". The original Grace data center CPU (paired with the Hopper data center GPU) used ARM Neoverse V2 cores. The "GB10" chip shipped in DGX Spark and announced here for RTX Spark uses Cortex X925 and Cortex A725 CPU cores.
Physically, NVIDIA did the GPU chiplet and Mediatek did the other chiplet that has the CPU, DRAM controller, and IO.
desktop is GB300, not GB10 like Spark
GB300 is nominally "available" in desktop form factor workstations priced around $100k. That's a few orders of magnitude away from the ordinary desktop PC market that consumers participate in.
they also announced a GB10/N1X windows desktop mini PC.
I really hope these take off and succeed and they support Linux. Qualcomm is seriously holding back the Linux ARM adoption with their continuous missteps.
"Notify me" -> i.e. when we finally have the DRAM to build this SoC.
I've been daily driving a dgx spark. Once you start there is no going back.
NVIDIA nailed it
Mind sharing more details about your use and experience with DGX? I'm just curious
Will NVIDIA get a monopoly on providing laptops and desktops with a lot of RAM going forward?
No. You can get a PowerBook today with 128 GB ram.
https://www.bhphotovideo.com/c/product/1957120-REG/apple_mbp...
I'm sorry to announce this to you, but the last PowerBook was released 21 years ago
Or get an AMD 395 laptop or mini PC for half the price of an equivalent mac device
https://www.bosgamepc.com/products/bosgame-m5-ai-mini-deskto...
Bosgame M5 AI Mini Desktop Ryzen AI Max+ 395 96GB variant €1.800,95 (sold out)
128GB+2TB variant €2.401,95 (in stock)
I have the latter, it's fantastic
$600 for 32GB ram seems bananas
Unfortunately in the current market 32GB of ddr5 seems to run about $400 as 2x16gb DIMMS, and even more for 1x32GB DIMM (higher density chips are more expensive). So $600 really isn't much over market price, especially considering strix halo uses 8000MHz ram instead of the typical 6000 found in consumer dimms.
https://onexplayerstore.com/products/onexplayer-super-x?vari...
$3649 with 128GB of ram
I didn't see this in the article but elsewhere I've seen the memory bandwidth quoted as 600GB/s [1]. For comparison:
- 5090/6000 Pro: 1792GB/s
- 5080:: 960GB/s
- 5070Ti: 892GB/s
- M3 Ultra: 819GB/s
- DGX Spark: 273GB/s (less than an M5 Pro at 307GB/s)
Memory bandwidth isn't everything but it will cap inference rate pretty heavily. Also, the M3 Ultra is for an almost 2 year old Mac Studio. It's widely expected that it'll be refreshed in Q3 with a likely M5 or M4 Ultra with >1000GB/s. I really hope Apple realizes what a market opportunity Apple has here.
The above shows just how good value the 5090 really is. It basically a RTX 6000 Pro with less RAM (and ~12% fewer CUDA units), which is a ~$10k card, for 20-30% of the price. This also demonstrates how NVidia uses VRAM for market segmentation. As an aside, the true data center cards (eg B100, H100) use HBM memory at ~3.2TB/s.
[1]: https://wccftech.com/nvidia-enters-pc-space-with-rtx-spark/
Spark memory bandwidth is ~300 GB/s. Internal bandwidth is 600 GB/s but that doesn't matter.
128 GB at 600 GB/s for this versus 32 GB at 1800 GB/s for 5090.
This is much better value than 5090, you can run much bigger models.
Here's a pretty detailed breakdown of this [1]:
> tl;dr - For software development, Qwen3.6 27B, 5090 gives you ~3x speed over M5 Max, letting you plow through code, while M5 Max gives you ~4x memory, letting you use higher quantization and bigger context. Which would you choose and why?
I've read a number of things from which the consensus seems to be that yes you can run a larger model and/or have more context with a 128GB+ Mac but the performance gap is still massive and with current hardware we're still talking about inference rates that matter. By this I mean there's a big difference between 10tok/s vs 30. Once we get to t apoint where it's 100 vs 300, it won't be as big of a deal, a bit like FPS in games.
Oh and there are similar concerns with the DGX Spark [2].
[1]: https://www.reddit.com/r/LocalLLaMA/comments/1t5v2gr/need_ad...
[2]: https://www.reddit.com/r/LocalLLaMA/comments/1sqk333/dgx_spa...
He's saying there is a niche where models are too large to practically run on 5090, but are still runnable on DGX Spark.
The larger memory also allows for pre-training / finetuning models, hence why it's aimed at developers.
Yeah and also the quoted 1 PF is only for sparse models (only half that for dense, if that), and the DGX had serious hardware issues: https://x.com/ID_AA_Carmack/status/1982831774850748825
It was wintel (windows + intel) before. This will be what? Windia? Wintek?
Winvidia
Nvideous
Nvidiows
Nvindows
They made their own x86 CPU? Or was that part outsourced? Ok ARM MediaTek.
ARM cpu made by MediaTek.
But probably worth clarifying it's not a typical "MediaTek CPU" some might assume by that. It has Nvidia's customized ARM CPU implementation + their GPU.
This has off-the-shelf Arm cores.
I think that Nvidia made GPU and CPU, and Mediatek made other parts of SoC necessary for a notebook. Grace is Nvidia's own CPU ARM core
I believe Grace is an ARM designed core. Vera is the nVidia designed core.
can it run linux as I am not a windows nor a macos user
I’m getting more and more convinced that we will end up running LLMs in our personal computers. Which makes me wonder where Anthropic/OpenAIs moats will come from.
Convince me
1. in order to run LLMs, especially the best ones, you need complicated devices which are expensive
2. if you buy one for your personal use, you are probably not going to utilize it all the time and it will be idle a lot
It seems to me that it will always be more economical that the LLM-running devices are in a datacenter where it is easier to make sure they are always utilized
If a model is substantially better than most humans at most tasks, the human isn't going to be able to perceive the difference between Claude Opus 7.7 and 8.7. Humans at some point aren't going to be able to perceive the difference on benchmarks either, because they are going to get wildly abstract.
AI vendors are really going to struggle to shift tokens far beyond the frontier of human capabilities. It's reasonable (not guaranteed) to assume that, if the trend of frontier models (doubling capabilities on benchmarks every n months) holds, then the same trend will hold for local models, and those local models will meet and exceed the perception frontier. This would mean a human cannot tell the difference between Mistral-Open-2030 and Claude Opus 2030.
That's a bunch of "ifs", but there's nothing exceptional about those "ifs". They're basically the scenario if nothing changes between now and ~2030 with regards to capabilities trend attainment.
The trend over the past three decades of personal computing has been for devices to become exponentially more powerful regardless of the actual computing needs of users. The excess computing power has famously been requested by projects such as SETI@Home and Folding@Home, and been exploited by bad actors for crypto mining. The most basic laptop today used only for web browsing and word processing would be a powerful workstation 20 years ago, when the most basic laptop was also used only for web browsing and word processing (and arguably for more things, as it was all mostly local software).
There is no ceiling to the power of consumer hardware. If it's cheap enough, it will be bought.
most crypto mining has moved to specialists, even where there were deliberate attempts to make it ASIC-resistant
SETI@Home is a very niche use case
and web browsing still happens by connecting to data centers and server farms, not by connecting to another laptop
I think you missed the point of my message. Web browsing still happens by connecting to data centres, so why are consumer laptops so much more (unnecessarily) powerful today than they were 20 years ago? All the more so given that, at that time, you were running MS Office locally rather than using Office 365 or Google Docs remotely.
This.
Even two or three years people were pointing out "The ChatGPT subscriptions you can buy with $2000 give you much more compute than whatever home setup you come up with" on r/LocalLLM. I did my own elementary school maths and came to the same conclusion.
Yet till this day people still boast how their beefy M4 Pro/Max machine with 32+GB RAM (which is not at all a "normal person's setup" and costs $2000+) runs LLMs smoothly, and "that's the future".
Someone needs to re-learn basic maths and take a walk around Best Buy to understand what "consumer laptop" looks like.
If there end up being useful workflows where you keep stuff running in the background or overnight that's one advantage, compared to a data center that might cut off your access during peak hours or etc.
Think of it like having a graphics card at home versus using a cloud gaming stream? Technically subscribing to GeForce is much cheaper up front than getting a card, but people still do that. So will the audience of people running agents at home be as large as PC gaming? I think that's kind of plausible.
> if there end up being useful workflows where you keep stuff running in the background or overnight that's one advantage
That is not how LLMs are typically used though in my experience
> Think of it like having a graphics card at home versus using a cloud gaming stream?
Latency seems to be much more important in that use case
>2. if you buy one for your personal use, you are probably not going to utilize it all the time and it will be idle a lot
I think consumers are primed for that type of behaviour though. I have an iPhone on my desk. It has something like 2-3tflops CPU+GPU, which is double that of the largest super computer on earth when Jurassic Park came out, and is probably more computing power than existed on earth when I was born in the 80s.
I use this device for around 1hr per day to write text messages.
It's inevitable. What might be a prosumer device today priced at 4000$ will be a regular consumer device in 10 years and models only get better.
Local models today are fine for a lot of mundane tasks and will continue to be so. The use cases where paying for frontier models is worth it, will continue to shrink for folks not doing frontier work.
> models only get better.
Or stall. Acceleration has been slowing significantly and gains seem to be tied to huge memory footprints.
Privacy and offline use would affect the choice as well. How niche are they, I am not sure.
Uploading your IP to the biggest IP thieves in human history seems bad idk.
2. Eventually we'll get to where local models that don't have sycophancy and slot-machine mechanics trained into them will perform better.
3. If your device run on battery, why not using a relatively cheap network call in place of a very power hungry local inference call?
Just like cloud vs private server. It'll be based on use case.
While I agree with that in principle, it is very worrisome that the prices of personal computers, especially of any personal computer that is not a big desktop, have been increasing continuously.
The price of a mini-PC with Intel Panther Lake is at least double in comparison with the price of a mini-PC with Arrow Lake H having similar specifications, and I am talking about barebones, before adding DRAM and SSDs, whose prices have risen even more.
The rise in prices is somewhat obfuscated by the confusing names of CPUs, i.e. some old and new CPUs may seem to be at similar prices and they have similar names, but the new CPU actually corresponds to a lower segment of the market, by having e.g. a smaller GPU and a lower clock frequency, while the CPU model that really corresponds to the old is named such that it seems to belong to the class corresponding to its present price.
As a concrete example of this obfuscation, which may confuse the buyers of laptops or mini-PCs, I have an ASUS 15 Pro with "Core Ultra 5 225H". If I would buy an ASUS 16 Pro now, the corresponding CPU model, the cheapest which is not worse than what I have, would be "Core Ultra X7 358H".
The best open weight LLMs don't run on this computer, or almost any consumer grade computer. Even the memory requirement for Gemma 4 is out of reach for most consumers (by which I mean those who are not on HN). Unless there is some magic that would make high quality LLMs consume no more than 8GB RAM which makes them usable on a 16GB laptop (which is the norm these days), "local LLM for personal computing" is mostly just a myth.
We're hitting the atomic limits of what's possible with minimum feature size in silicon. It's also very hard to remove 1 kW of heat from a laptop, let alone do it quietly or on battery.
My biggest concern with local LLMs is there just isn't enough RAM or HD space to run multiple models, and the generic LLMs are too generic...
I find it hard to see how that would ever be economical. LLMs need very expensive power hungry chips and datacenters have
- bulk discounts - cheaper electricity - high utilisation to spread the costs among many users
I don't see how PCs could ever compete against it. Most users AI demands would probably result in >90% idle time on the GPU.
First we need to actually still be employed, and have them at affordable price.
If we do, it won't be on this chip.
It'll be just another round of the client-side vs server-side processing rounds. We've been through them, we will keep going through them.
i think a lot of that is for government and enterprise use. even for personal computers themselves (i.e.: laptops) they're usually loss leaders, they don't turn profit. You can run a server (and many do) on laptops, but that didn't replace cloud services or server hosting. You can't store enormous amounts of data on your laptop/phone for the llm to use, or access tools the app dev wouldn't want exposed on untrusted devices.
The whole replacing people angle is just the short term use case the more ghoulish executives are thinking about. In practice, lots of lots of new use cases have been made possible by LLMs. A lot of which can be done locally. But whatever capacity you have locally, they can have more of and for cheaper, and they manage the model instead of you doing it yourself. I think you put it nicely though, their moat will be thinned, and I doubt they'll be as profitable as their funding suggests, but at the same time the demand for them won't go away either. I don't know if OpenAI and Anthropic will be viable, but I'm nearly certain Deepseek is.
The tipping point will be power usage, if a local llm can run the same workload for less power that would be a game changer. Nvidia might get decimated, but even Google and others have moved on from GPUs already, they have faster and more power efficient TPUs. Add to that network bandwidth and availability issues, their moat remains. Also consider that even for graphics capabilities, user devices just don't have a consistent spec to make things like widespread 3d graphics and webgl usage viable. Someone's cheap android phone will never run a local llm reliably,same as it won't a 3d game. even if they have a high-end iphone, network providers aren't always performant as they are in western countries, and then there are people that won't want to install your app or local software, and then browser based exposure of the capability to sites which will have similar hardware spec issues, OS instabilities, competing tabs,etc...
No thunderbolt is a big no for me. Its one of the greatest feature of MacbookPro that makes it dockable and expandable as a desktop with a good thunderbolt dock.
Thats also possible with usb-c.
With some caveats, you wouldn't be able to connect two 4k monitors to a dock without TB5.
USB 4 v2 has the same display capabilities as TB5. In fact, TB5 gets its display capabilities from USB 4 v2
Well, it was only a matter of time, since both AMD and now Intel are now switching to APUs. Nvidia could either cede the desktop GPU market to them, going all-in into AI datacenter chips, or it could challenge them.
Maybe the Nth time's the charm and Microsoft+Nvidia will manage to make Windows on ARM a viable platform.
Looks like the MSI one might be a 2-in-1, if it has good stylus support I might have a good candidate for an upgrade, thought my ~3-4 year old Galaxy Book is holding up alright for now.
hope nvidia support driver better than qualcomm. also hope they support linux soon.
Is this finally Macbook Chip Efficiency coming to Windows or will it just be shittier compatibility for slightly better battery life?
I heard leaked geekbench putting it behind the m3, which is couple years old now.
All I care about is if I can get one of these for significantly less than a dgx and get Linux on it for some cuda Blackwell kerneling.
can these do training or only inference? currently working on learning machine learning and I'd love to have a physical machine I could aim to build real workloads on in a few years.
They're Turing complete. What else do you need?
There is a reason why Google has tpu8i and tpu8t
technically in order for something to be turing complete it needs infinite memory
The more (memory) you buy, the more you save!
It’s possible (likely, even) to have a chip fast enough for inference, but not fast enough or with enough memory to do meaningful training runs. Like the current DGX spark.
not for llm full training, but can do some finetuning for sure.
I believe training is way more processor intensive than inference.
Unified RAM means its soldered to the mainboard, right?
I'm not sure if I like this. Sure for a laptop this might be not a big problem but if this ARM ecosystem is a success it will spread to desktop computers and I fear we could lose the existing modularity.
"Unified" means that it's shared between CPU and GPU, I believe.
But yes, it tends to be soldered on.
No, but LPDDR means soldered, there are no LPDDR dimms
There's LPCAMM2, but it's very recent. The Framework Pro laptop supports it, for example, although only on the Intel variant.
I think unified RAM means soldered to the SoC, which is in turn soldered to the mainboard
I have no idea how powerful or power efficient these guys are, but this seems to be the first step in a bigger push towards Windows on ARM (without loosing gaming).
I think more announcements will follow soon from other companies.
My DGX Sparks are the first and only devices I have with 200W USB-C PD. Low power by AI workstation standards, but intolerable in a laptop.
Intolerable? Why?
Battery life
The comment I'm replying to appears to be talking about power DELIVERY, not consumption. Why would extra power-delivery capacity be intolerable?
The DGX Spark doesn't have a battery. If it comes with 200W delivery (actually 240W), it's because it plans on consuming close to that amount.
Although I'm kinda surprised the DGX Spark used USB-C at all for power instead of just like a DC jack or whatever. But whatever.
It's worth noting that Nvidia power management on Linux has been absymal. There also aren't any of the usual power management options to see how much power things are using, which is quite atypical for a modern system.
Nvidia really threw stuff over the wall with the DGX Spark release. They don't seem to really care. I sort of think they'll spend a little more time on Windows, where there's no pesky upstreaming to do and they can just do whatever, but man, it's such typical hubris from Nvidia to build such an expensive box with good chips but make it basically unsupportable and roasty hot all the time.
You also generally have to run an ever more stale two year old Ubuntu derived DGX OS to get anywhere, with bespoke kernel and drivers all. None of it is well supported, none of it just works like a comparable PC or even well behaved arm system would.
As for other ARM, there were rumors AMD Sound Wave is/was going to be a ~10W arm APU, but there hasn't been much said about it lately. Honestly given the ram crunch, it's maybe just not worth trying to build a system with a cheap core, if the rest of your costs are going to stay so stratospheric. https://www.techpowerup.com/341848/amd-sound-wave-arm-powere...
What is this product anyway? Is it a general purpose CPU or is it specifically designed for MS Windows? Nvidia stepping back from the open source?
"Introducing the NVIDIA RTX Spark™ Superchip. The fusion of NVIDIA AI and RTX graphics in a single chip redefines Windows PCs and delivers amazing creating, AI development, and gaming—on the slimmest, most beautiful RTX laptops ever and small, ultra-efficient desktops."
It’s nivdia attempting to compete with Apple’s M-series
Its nvidia attempt to gain additional market share and expected as well. If the whole ecosystem is around nvidia and its the easiest way of running stuff, Nvidia offering more enterprise infrastrcuture allows companies to just buy directly nvidia.
Nvidia is also very very rich and pushes the boundaries of stuff. They stoped waiting for industry standards. You can see this in there network stuff. All nvidia.
Next logical step (at least now, not something i thought about) was there CPU for their GPU racks/clusters/systems.
Now they have everything anyway, RTX Spark is just logical.
I don't think its specificly targeted at Apple at all.
Apple has like 10-15% market share and just because some IT nerds buy themselves a mac mini doesn't mean much.
Plenty of them actually just run openclaw without local models. Something which surprised me quite a lot.
But i have two 4090 at home. They consume a lot of power and i had to research the proper Mainboardmodel and had to mod one 4090 to use water cooling because they run too hot.
There Spark setup was at 3k, way to expensive for normal people. If they can get this down and sell more, great for their ecosystem (strengthening it) and getting more money from people.
It does surprise me though that they have enough capacity for this chip and not just putting everyting in Rubin but perhaps the build out has slowed down a little or they start to diverse already for economic savety
Their target competition is the AMD Strix Halo which is eating the Sparks lunch right now.
Also sounds like they are ditching the discrete GPU altogether.
All the news articles in my feed mentioned Nvidia reinventing personal computing which is laughable given the specs are worse than the m series. I’m guessing they saw how well Apple devices were selling and rushed to get something similar out so they can ride the hype train and have something to fall back on if ai DC spend slows down.
There's a lot of companies trying to support datacenter systems like GH and Rubin that don't have dev hardware remotely resembling it. M-series isn't a good option, speaking from the personal experience of currently using one for this exact purpose.
I wouldn't say it's Nvidia stepping back from open source... if anything this is doubling down on it, as one of the selling points of this is the 128GB of unified memory which will allow for hosting local models (i.e, nvidia's new open model they just released). I guess it's pretty cool, I'm a big supporter of local LLMs/open weight models so seems enticing to me, although I'm not sure this will be super applicable to a lot of regular consumers. Seems like a pretty niche product.
Linux works but MS is just paying them not to mention it.
It’s a step in the right direction, but there’s still a long ways to go in terms of smaller LLMs ability and hardware costs
Great! More pressure on fabs, price of standard GPU will again rise.
Guess I need to postpone my gamer PC renewal to end 2030.
I really like this, but I think the reason Apple Silicon took off was that Apple sort of forced devs to support ARM. Not sure if Microsoft can do the same for Windows…
Developers weren’t really “forced” to support ARM. They simply recognized that all future Macs would be ARM, whereas most new PCs would continue to run on x86. So the incentive to adopt ARM was much weaker on the PC side.
They didn’t though. Rosetta 2.
rosetta is a relatively short term solution. will be supported up to macOS 28
Microsoft can do the same for windows - they need to address the fat bundle solution that Apple came up with, but for Windows, though ..
Anything they can do to avoid producing more 5090 FEs.
"Unified Memory" still means divided address space right? You have to pre-allocate system vs gpu and copy from one to the other?
How would these compare to a MacBook Pro M5 in terms of performance and price?
Yeaaaah . But at what Cost though.
I'm waiting for powerful on device LLM models, since that not worth it
Have you tried Qwen 3.6 or Gemma 4? They're not frontier level but certainly have their uses.
The fact they advertise it as some step forward in PCs is outright bizzare.
It's just worse Strix Halo, as you are landing square in middle of Windows ARM problems
Strix Halo chips have around 210+ GB/sec gpu memory bandwidth and announcements put the new nvidia chip at around 300GB/sec gpu memory bandwidth.
I 'd say that is an improvement if you want to run local llm inference. Still well below with what you can achieve with Apple chips though.
Very exciting! sounds like we're finally leaving x86 behind
After nvidia's many years of neglecting Linux, paired with direct Microsoft's involvement? Are we going to trust them, to allow installing Linux in these easily?
I don't think so.
This most likely be a winmodem situation, again
DGX Spark has the same soc and ships with Ubuntu
Okay, but still it's highly skeptical trusting MS, and NVIDIA.
It ships with DGX OS 7, which includes Ubuntu's 24.04 repos. It is not using mainline Ubuntu, and if you want to run Ubuntu 26.04, you'll have to do some work.
how will this compare to having an rtx pro 6000 for inference? (not training)
Related:
A powerful new chapter for Windows PCs, accelerated by Nvidia RTX Spark
https://news.ycombinator.com/item?id=48352693
Surface Laptop Ultra: Made for World Makers
https://news.ycombinator.com/item?id=48352627
The thing I think is really funny is that if this takes off, frontier model companies and datacenters will end up holding the bag, and as per usual after the last few tech hype cycles, NVIDIA will still be selling.
Eventually a lot of inference will get right-sized into something you affordably run yourself.
First:
> "Our goal is to deliver unmetered intelligence to every home and every desk with Windows," said Satya Nadella, chairman and head of Microsoft.
Then:
> However, Ian Fogg, Research Director at industry analyst firm FDM CCS Insight said the change was "likely to come with a significant price tag" and Nvidia would be targeting "those looking for workstation-class performance".
So... not every desk with Windows.
First, make it possible. Then, expand the market. The early adopters help pay R&D for later efforts. Every desk is a good goal, even if not hit by the first doodad.
It just feels too much like what they said about Apple II and early Windows. A play at nostalgia instead putting real thought into it.
I was an engineer at both MS and Apple, and wholeheartedly agree with you.
My question is, what happens to the people who use RTX cards for gaming? This new solution isn't meant for that. Do they need an "AI accelerator" and a gaming-centric GPU?
I don’t know anyone other than a very small but vocal minority who will give a shit about this.
Even in the analytics side most of the stuff is some shonky ass numpy or excel gank.
I don’t know what the market is. I just can’t see it.
The constant deliberate conflagration of LLMs with general intelligence is so grating.
Question is: "Can it run Doom?"
ARM64+GPU sure seems like the future. I'm still using my M1 and even that can handle models well, has decent graphics, M5 is a beast, and M6 must surely go even bigger on LLM compute. Now Microsoft has a compelling ARM64+GPU future too.
What does AMD or Intel have here?
Don't know about intel, but AMD has Strix Halo with unified memory and really impressive performance.
I think the future will be 50/50 x64 vs arm64 for PCs.
competitor is already on the market and is x86: AMD AI 395+
bechmarks with DGX arnt spectacular for NVIDIAs software and CUDA lead.
wouldnt count on this being a price/compute challenger. especially with overpriced VRAM.
Strix halo's 8060S gpu is very weak, and is roughly equivalent to a 4060 laptop GPU, whereas GB10's gpu is equivalent to a desktop 5070. For LLM throughput, tok/s is similar due to bottleneck by memory bandwidth, but the GB10 has 3x faster prefill. People have also been able to squeeze out much better performance on GB10 using NVFP4 and other improvements in the months after the DGX Spark launch, so don't be misled by early lackluster benchmarks. For the RTX Spark, which also targets gaming and creative applications, the 3x faster GPU is quite nice.
Or like a m4 max? This thing has <300GB/s vs the max with 550GB/s
All those CUDA cores in the sparks but they're starved for memory bandwidth.
I am still waiting for NVidia to release a system that legit beats 3090 maxxing for the home gamer...
I feel like the shape of the market right now for "home lab" inference is:
The sparks are good if your ultimate plan is to spend even more on NVidia hardware in future to run your dev setups at usable speeds. Or, you're developing for a work cluster.
If you mainly want to run local models at acceptable speeds portably, buy a mac with lots of RAM. If you’re happy with non-portable / racked, buy 3090s (dense) or mac studios (MoEs). Buy newer cards if you are restricted on power or slots. If you are rich, buy a6000 blackwells.
The only Question is is it worth suffering hip and x86? I suspect a lot of folks might like a machine that mimics their GB300 But costs less than a dgx.
Also I heard the tensor core instructions on the dgx are gimped and you’re better off with a rtx pro x000. Is that the same with these machines?
Is CUDA really a lead for long? Aren’t all the latest competitive approaches avoiding all the standard software stacks and writing deeply customized software that is very directly tied to whatever hardware they use?
And is it really a way to lock in people? With AI coding tools, isn’t it trivial to write software on top of CUDA and rewrite it to target some other hardware?
yes.
no.
Yeah, there is zero chance I'm ever running Windows ROFL.
However, I'd jump from Mac in a Heartbeat if this supported Linux.
Some other relevant discussions and sources …
NVIDIA and Microsoft Reinvent Windows PCs for the Age of Personal AI
https://news.ycombinator.com/item?id=48352705
NVIDIA DGX Station for Windows Puts a Trillion-Parameter AI Supercomputer on Every Enterprise Desk
https://news.ycombinator.com/item?id=48352691
Introducing Surface Laptop Ultra: Made for world makers
https://news.ycombinator.com/item?id=48352627
Introducing a powerful new chapter for Windows PCs, accelerated by NVIDIA RTX Spark
https://news.ycombinator.com/item?id=48352693
2 comments in total there
It all sounds good on paper. But I have trouble believing Windows can be a good platform for this. Microsoft has lost all trust after inserting ads into windows, slowly removing power user features, and exploiting every dark pattern they can. And for years, the ARM based Windows laptops have been useless due to app compatibility issues. Why would this change now? Is it priced to be a lot cheaper than Apple’s laptops? Or is this a niche product for AI developers basically?
Anecdotally Windows ARM works fine for me, although to be honest most of my work is command line + browser anyway. WSL works like a treat. Steam installs and most lower end games also play fine on my ARM laptop too. Games that require kernel anticheat don't work.
I think they make a great "second device" where you have something meatier to fall back to if something doesn't quite work right. I'm not sure if it's ready to take on the "main device" role just yet. But it's a far far better experience than the Surface RT days.
The "gaming" take is a strange one indeed for an ARM platform. Hopefully they (Microsoft or Nvidia?) put some real effort into the translation layer. They claim modern AAA games, but it is possible they strongarmed the developers to make them an ARM build for a few select titles...
It's clear gaming was not a major concern, it's just "good enough" for someone running AI models and occasionally wants to play some games, not made to primarily play games.
Yep. I noticed the press releases talk about all the partners they have. It seems like a desperate attempt to manufacture a consensus to invest in this new hardware instead of leaving it sort of abandoned like the other Windows ARM stuff. But the problem is that these attempts end up having a few very visible apps working on the architecture and others not actually doing anything substantial.
Sure the graphics capabilities are probably very good. But if you’re a game developer who has traditionally built on Windows on x86 chips, would you want to invest in this new chip or invest in making games for the Apple ecosystem? Aren’t there more new customers to reach in the Apple world than this new Nvidia world?
> But if you’re a game developer who has traditionally built on Windows on x86 chips, would you want to invest in this new chip or invest in making games for the Apple ecosystem?
Windows and the new chip. Higher developer productivity and higher chances of a substantial audience.
Who cares about Windows, the goal is to run local AI models similar to AMD Strix Halo and Apple Silicon machines. The OS is honestly a distant last concern as long as the models work well, as you could put Linux on these too, but not sure how well wake lock works.
Hopefully MSFT would look at this as a do or die system, and go all in on improving the user and ownership experience. Will they? Not so sure.
Microsoft sees windows purely as a platform to sell AI products these days.
That's what they're working on, in theory, with Windows K2.
I would never trust Microsoft. Their next drama is revoking Office 2019 perpetual licenses https://www.youtube.com/watch?v=KRnno9VIZx0. It never ends with them because they know they have you by the balls.
I trust them on a daily basis. No issues thus far..
A lot of the app compatibility issues on current machines are down to Qualcomm's poor drivers - the actual core bits are mostly okay.
So basically Cerebras style?
Not at all. This is a more like what Apple has been doing the past few years. A bunch of decent arm cores paired with a beefy integrated GPU.
No.
Imagine a Beowulf cluster of these. /slashdot
This may finally be the chip family ARM on Windows has always needed. Qualcomm's chips have always been dogs with slow off-the-shelf ARM CPU cores that have pathetic single-threaded performance compared to x86 AMD/Intel or ARM Apple Silicon designs.
For reference, this is just a single benchmark, but as an idea of each vendor's top mobile CPU single-threaded performance:
Geekbench Single Thread Score:
- DGX Spark (same CPU as RTX Spark): 3125
- Snapdragon X1 Elite: 2950
- Snapdragon X2 Elite Extreme: 4050
- AMD Ryzen 9 9955HX: 3225
- Intel Core Ultra 9 290HX Plus: 3175
- Apple M5 Max: 4350
I'm happy to be wrong about Qualcomm's latest X2 chip performance, even if it is shipping in only a single product so far. Their previous best was the lowest in this list.
This will likely have worse single threaded performance than recent Qualcomm CPUs.
These chips also appear to be using off-the-shelf ARM cores.
Qualcomm Snapdragon x1 and upcoming x2 use their Oryon core and have much faster single-thread performance than Intel/Amd and this nvidia soc that uses off-the-shelf arm cores
That wasn't true of the X1, but apparently the X2 (which is only in a single device so far) does appear to finally be fast. The first Windows ARM CPU to be faster than any of its x86 rivals. Competitive with Apple Silicon single-thread performance even.
I was disappointed to see that the RTX Spark has the ARM cores from the DGX Spark. I was hoping it had their new in-house developed cores that Nvidia is starting to use on their latest gen server parts. They look really fast. That said, if RTX Spark has CPU performance like the DGX Spark, it will be almost as fast as the top AMD/Intel parts.
This will crush the M5 Max going by the numbers. I'm curious to see how much they end up costing
It won't, the top tier RTX Spark has the same exact CPU and GPU as DGX Spark, so you can check DGX Spark CPU benchmarks to see how it fares. Spoiler: it's about M3 Max level. And they're only coming this fall.
Nah, still ~300GB/s memory bandwidth. That will be slower than the M5 max, by a wide margin for LLM inference.
M5 max is 3x stronger and 50% more power efficient. nice try though.
... but you'll be rewriting inference for any model that isn't a well-known LLM. Yourself.
AI coding agents can do that pretty nicely already and it will only (slowly) improve over time.
>Lenovo, HP, Dell and Apple accounted for almost 75% of the world's PC market in the first three months of this year, according to research firm Gartner.
https://www.gartner.com/en/newsroom/press-releases/2026-4-10...