These are going to be so awesome to buy in 5-8 years!
Right now I’m running 2x 3090 and 128 gb ram on 3.8 flash-next180b (the quant is 97gb) and it hums at120t/s. It’s really a surprisingly good model for general tasks, and I’m soon to see how it does with more agentic kit. For agentic work it does slow down quite a bit when the context fills up… but what do you want for running a 180b model on consumer hardware?
Right now I have it set up as an administrator for our asterisk PBX and radius server, and it does higher level judgment on our cctv footage for notifications to security. It’s been really good so far. Agents are the new UI, everything just works over SSH now.
> (...) it hums at120t/s. It’s really a surprisingly good model for general tasks, and I’m soon to see how it does with more agentic kit.
This is exactly why I interpret the aggressive price cutting by OpenAI on their Luna model line as a clear sign they are scared shitless of local models, and how they can trigger a "the king has no clothes" moment for AI companies like OpenAI.
> as a clear sign they are scared shitless of local models
Partially agree and partially disagree. Qwen 3.8 next (120b + 50b ngrams) is at the upper end of a local model, and serving it at good speeds (esp for pp) is not cheap. So no, oAI are not particularly scared that everyone will suddenly start using local models of this size / quality.
I do agree that having these models in the open is great pressure on the price point of "small" API models (like luna, haiku, etc). And that's good news for us consumers and for the field in general - more competition is always better.
And on the cost estimate, the true power of local models is the guarantee of running them on prem w/o any data going out. Price wise it is still cheaper to go API, for small teams.
> as a clear sign they are scared shitless of local models,
The number of people with hardware to run a model like that is very small.
It's going to be many years before hardware to run medium size local models is widely available. Running mid-size local models doesn't even register as a factor in pricing decisions right now.
It's such a small part of the market and will be for many years, until RAM and other hardware prices come down.
> The number of people with hardware to run a model like that is very small.
True, but the number of people who pay the non-subsidized API price is also very small, and it's all in big companies who would have no issue buying or renting hardware for that purpose.
Many years is only long in relation to the AI hype cycle. In terms of a human lifespan it is not that much, maybe 10 years at most and that's pessimistic. We technically live in that future already it is just very unevenly distributed.
> The number of people with hardware to run a model like that is very small.
That's not the problem anyone is facing, and analysis anyone is doing. At all.
The problem is the huge AI-related budget that companies who dove head first onto AI are facing.
Therefore the problem everyone is faced with boils down to whether they keep paying the likes of OpenAI a de-facto business tax, or they replace that tax with a one-off CapEx invoice followed by a utilities bill.
And you don't need a frontier model to do most things. In fact, the biggest challenge in using local models in agentic coding is the models supporting these agent-based tools.
> It's going to be many years before hardware to run medium size local models is widely available.
Nonsense. Speaking for myself, my employer already introduced a hard cap on AI credits while rolling out hardware support for local models, and even went to the extents of rolling out workstations with 64GB of RAM as a baseline. Consequently, we are using the likes of gemma4 for a fair share of everyday tasks. This has a transformative effect on your workflows because suddenly you don't ration tokens, and agent orchestration doesn't become a cash guzzler. So far the only constraints are tool support and token throughput, but that's not a concern.
I'm working off a 128gb Strix system on mostly hydro and it's worth it for me. Qwen 3.8 27B and Flash Next, Gemma 4 and MOE models. It Keeps my thinking and methods local while moving at a speed I can actually read and think about.
It can be annoying when you're running something long and realize there is an issue. But hey, I'm not wasting paid tokens. Reset and try again. I've probably thought of a better way by then too.
hmm...wonder where this will land in market segments.
Seems like neither a consumer product nor server. Maybe if they halve the ram and price it reasonably people with desktops will buy it for MoE like prosumer homelab AI enthusiast type situations?
All these big names in tech will monetize so hard on such products, until this becomes obsolete and available to us all. This should be accessible to everyone at a convenient price, governments should force these companies to mass produce them and sell to people, providing a small subsidy. Everyone deserves to run big models locally.
Given the rumored numbers are all true, this could be a huge contender to M5 Ultra(or M7 Ultra if we consider possible delays in the schedule?). Low-batch single-user serving is usually bottlenecked by `E(sum over (memory bandwidth by tier) * (memory size by tier))`.
That's only partially true - the GDDRX you'd find on say a 3090 is physically very different than what you'd find on a strix halo or mac or vanilla ram. e.g. runs 4 voltage levels not 2 so that it can transmit twice the info per cycle
The large amount of memory means you can serve larger models, or less quantized, but that also means they're going to be slower due to the low bandwidth.
Its bandwidth is around that of 32 DDR5 DIMMS (50ish GB/s). These boards should have built in NICs but they don't because Intel also wants to sell more servers.
The marketing department doesn't have to lift a finger in this market. They could quietly drop it for sale without a single announcement and it would be backordered for a year.
You can tell by the shape of the board that it's meant to go into a server chassis behind a stack of 80mm blower fans. It's not meant for consumers, but you will be able to buy it used for $200 off ebay in 7 years to put in your home lab.
Wonder if there will be consumer video drivers for these GPUs once the AI hardware shortage is over. Hope to buy a gaming rig one day that doesn't cost a fortune.
I remember back in the day of AGP cards around 2002, you could use a 3d card on linux to render the grafix but use a different GPU for the actual output. I used XFree86 on linux with a voodoo banshee (or maybe was a riva tnt2) and a matrox mga200 for its sync-on-green support with my monitor to play games.
I know modern card architecture doesn't support this out of the box, but what is the actuall problem of forwarding a once-rendered frame to another card for output? Sureley DMA transfers PCIe->RAM->PCIe nowaday are fast enough to saturate 60fps at 4K?
Intel ai gpu division is so aggressively anti “small” consumer. I think it has something to do with low volume and a small sales/distribution network. The data center cards I have from amd and nvidia are grey market at best. I don’t think intel moved enough units to establish alternate channels.
What I don't understand is: when you only plan for your product to have 10 customers, why even make a marketing page like this for it? At that point it's essentially the hardware equivalent of consultingware, something you sell to your existing contacts through outbound calls and by shoving ads through the B2B equivalent of a post-consumer sales channel (i.e. proactive outreach by your org's internal champion for that customer, to said customer.)
1.5 TB/s 1.3 PFLOPS with no fast multi-GPU, it's a consumer card that pretends to be a server card.
It would work great on Qwen3.8-Flash-Next or GLM-5.3 as single user decode only hits a few experts, anything with concurrent users and you'd be sitting for a while.
That's the point I believe fast local 2-4 user service for cheap on very large models.
You have to pretend it's for servers to keep investors happy, it would be banks and similar with in house servers that have to run local large models buying these for experiments and cybersecurity. Likely.
Slop and placeholder text on the page, including this one. They should have reviewed their webpage.
"What software frameworks are relevant to Intel GPU D270?
Intel GPU D270 messaging should reference familiar AI software and serving frameworks such as PyTorch, vLLM, SGLang, and OpenAI Triton, where approved. This helps developers and platform teams understand how the product fits into existing AI workflows."
The cost of a 480gb ram video card seems preposterous to be even marketing at this point.
Intel is going to make a card that costs $1 million dollars.. this is like industrial newsletters bleeding into a public forum
> Commercial Scale Pricing: In the industrial data center sector, battery storage systems are priced per kilowatt-hour (kWh) or megawatt-hour (MWh). Industrial battery storage installations typically cost $300,000 to $450,000+ per MWh of capacity depending on software integration, cooling systems, and structural framing.
Looks like Intel is choosing to title its new crescent island cards with the D prefix despite being Crescent Island.
Kind of disappointed that Intel is (probably) not going to release consumer class GPUs. Intel's driver stack was seriously getting mature—both on Linux and Windows. It's always nice to have a third option available.
Agreed. I was really rooting for them to deliver that third iteration to consumers, and retain them as a competitive option in the low to mid range. Pretty sad to see.
"Both" requires more engineering staff on payroll, more wafers to fab/dice/package/test, more boards to manufacture, more inventory to manage, among other things. If their math says it isn't worth it (granted that they don't always get it right), then they don't do it.
> "Both" requires more engineering staff on payroll, more wafers to fab/dice/package/test, more boards to manufacture, more inventory to manage, among other things.
Aren't you describing how a business operates?
I mean, doesn't Intel already hire people for those tasks? Unless you are worried about a specific bottleneck, increasing headcount and production output is what businesses do.
For now. A limited run of Celestials until 2028 when Serpent Lake switches to Nvidia tiles, and a bit longer for the LTS combo (AMD style) IO+GPU die in lower end/embedded products.
Druid is still dead, consumer DGPUs are still, and workstation DGPUs also seems to be dead, and most of the Xe driver team for Windows was laid off, as well as most of the Xe chip design team.
This is a brand new class of GPU that I am excited for, especially because it's rated for air cooling. I hope to see other manufacturers release GPU's with multiple of 100 GB's that can be utilized even in an at-home system's with PCIe 5 support. I probably can't afford this but knowing the engineering is sound at this scale means a similar accelerator maybe attainable.
They originally said it was going to be 120GB but that partners could ship bigger capacities. So much for that!
I do hope this has a price point that makes it somewhat competitive with a Mac! Consumers have had nearly no choices available for mem capacity, and 20 channels of lpddr5 is one way to get that capacity! Go Intel!
Both this GPU and the Mac are constrained by a lack of GPU compute. There's basically no rack-scale GPUs with 480gb on-die that have a sub-400W compute element. It's going to be incredibly prefill-constrained on denser model inference just like the Apple Silicon iGPUs are.
> But I am surprised that they were able to pack so much flops into 300W.
I'm not. When you're doing inference you typically reduce power limits anyway if you're optimizing for efficiency. Even if you don't, it doesn't spike to the max of the card.
My 3090 I power it down to 175W or so for single-digit reductions in performance for like, 60% of the wattage, something like that.
> Efficient token serving
Dense, air-cooled, standard PCIe infrastructure helps maximize tokens per rack and watt, scaling serving capacity and keeping costs in check.
Next the tokens will be asking for tequila and caviar. Bow down to the new token overloads.
These are going to be so awesome to buy in 5-8 years!
Right now I’m running 2x 3090 and 128 gb ram on 3.8 flash-next180b (the quant is 97gb) and it hums at120t/s. It’s really a surprisingly good model for general tasks, and I’m soon to see how it does with more agentic kit. For agentic work it does slow down quite a bit when the context fills up… but what do you want for running a 180b model on consumer hardware?
Right now I have it set up as an administrator for our asterisk PBX and radius server, and it does higher level judgment on our cctv footage for notifications to security. It’s been really good so far. Agents are the new UI, everything just works over SSH now.
How do you do the cctv analysis ?
> (...) it hums at120t/s. It’s really a surprisingly good model for general tasks, and I’m soon to see how it does with more agentic kit.
This is exactly why I interpret the aggressive price cutting by OpenAI on their Luna model line as a clear sign they are scared shitless of local models, and how they can trigger a "the king has no clothes" moment for AI companies like OpenAI.
> as a clear sign they are scared shitless of local models
Partially agree and partially disagree. Qwen 3.8 next (120b + 50b ngrams) is at the upper end of a local model, and serving it at good speeds (esp for pp) is not cheap. So no, oAI are not particularly scared that everyone will suddenly start using local models of this size / quality.
I do agree that having these models in the open is great pressure on the price point of "small" API models (like luna, haiku, etc). And that's good news for us consumers and for the field in general - more competition is always better.
And on the cost estimate, the true power of local models is the guarantee of running them on prem w/o any data going out. Price wise it is still cheaper to go API, for small teams.
> as a clear sign they are scared shitless of local models,
The number of people with hardware to run a model like that is very small.
It's going to be many years before hardware to run medium size local models is widely available. Running mid-size local models doesn't even register as a factor in pricing decisions right now.
It's such a small part of the market and will be for many years, until RAM and other hardware prices come down.
> The number of people with hardware to run a model like that is very small.
Not completely obvious. I have a Ryzen 7900 with 192GB of DDR5 and a 4090. Pedestrian stuff 12 months ago (around 3k€ all-in).
It runs 3.8 Flash-Next at 120tok/s. https://github.com/AnnoyingTechnology/nvidia-4090-flash-next...
[delayed]
12 months ago buying 192 gigs of ram was pedestrian? Really?
> The number of people with hardware to run a model like that is very small.
True, but the number of people who pay the non-subsidized API price is also very small, and it's all in big companies who would have no issue buying or renting hardware for that purpose.
Many years is only long in relation to the AI hype cycle. In terms of a human lifespan it is not that much, maybe 10 years at most and that's pessimistic. We technically live in that future already it is just very unevenly distributed.
A human lifespan doesn't have that many decades in it either...
> The number of people with hardware to run a model like that is very small.
That's not the problem anyone is facing, and analysis anyone is doing. At all.
The problem is the huge AI-related budget that companies who dove head first onto AI are facing.
Therefore the problem everyone is faced with boils down to whether they keep paying the likes of OpenAI a de-facto business tax, or they replace that tax with a one-off CapEx invoice followed by a utilities bill.
And you don't need a frontier model to do most things. In fact, the biggest challenge in using local models in agentic coding is the models supporting these agent-based tools.
> It's going to be many years before hardware to run medium size local models is widely available.
Nonsense. Speaking for myself, my employer already introduced a hard cap on AI credits while rolling out hardware support for local models, and even went to the extents of rolling out workstations with 64GB of RAM as a baseline. Consequently, we are using the likes of gemma4 for a fair share of everyday tasks. This has a transformative effect on your workflows because suddenly you don't ration tokens, and agent orchestration doesn't become a cash guzzler. So far the only constraints are tool support and token throughput, but that's not a concern.
Viewed differently, they make it much cheaper to distill their models.
At cost of power at $.45kwh here and dual 3090 setup easily approaching 1kw per hour at 120 tok/sec, latest Luna is cheaper per million tokens.
Your electrical costs are likely less than mine though
Yeah electric costs are nil on the microgrid but capacity is limited.
I'm working off a 128gb Strix system on mostly hydro and it's worth it for me. Qwen 3.8 27B and Flash Next, Gemma 4 and MOE models. It Keeps my thinking and methods local while moving at a speed I can actually read and think about.
It can be annoying when you're running something long and realize there is an issue. But hey, I'm not wasting paid tokens. Reset and try again. I've probably thought of a better way by then too.
Are you running club-3090 or something else?
Flash-Next runs on Strata, version 0.1.42, an inference engine built specifically for Qwen3.8-flash-next
> These are going to be so awesome to buy in 5-8 years!
Literally my first thought!
hmm...wonder where this will land in market segments.
Seems like neither a consumer product nor server. Maybe if they halve the ram and price it reasonably people with desktops will buy it for MoE like prosumer homelab AI enthusiast type situations?
All these big names in tech will monetize so hard on such products, until this becomes obsolete and available to us all. This should be accessible to everyone at a convenient price, governments should force these companies to mass produce them and sell to people, providing a small subsidy. Everyone deserves to run big models locally.
Given the rumored numbers are all true, this could be a huge contender to M5 Ultra(or M7 Ultra if we consider possible delays in the schedule?). Low-batch single-user serving is usually bottlenecked by `E(sum over (memory bandwidth by tier) * (memory size by tier))`.
edit: typo
With an impressive 1.5TB/s RAM bandwidth you'd think they'd advertise that more.
Marketing department asleep at the wheel.
This is a very low number for a server GPU. My V100 from 2017 is 0.9 TB/s.
A single B300 has almost 8 TB/s.
AMD MI300x from 2023 is 5.3 TB/s
Edit: my bad you are saying that this is an impressive number precisely because it is more or less normal RAM, not VRAM
VRAM isn't a type of memory, it's just RAM used for the GPU. So this LPDDR5 is VRAM, just like it can be with my Strix Halo machine.
That's only partially true - the GDDRX you'd find on say a 3090 is physically very different than what you'd find on a strix halo or mac or vanilla ram. e.g. runs 4 voltage levels not 2 so that it can transmit twice the info per cycle
Very true, though traditionally it did refer to GDDR.
And a 5090 has 1.79 TB/s.
So yeah on its own not very impressive.
The large amount of memory means you can serve larger models, or less quantized, but that also means they're going to be slower due to the low bandwidth.
https://www.techpowerup.com/gpu-specs/geforce-rtx-5090.c4216
Its bandwidth is around that of 32 DDR5 DIMMS (50ish GB/s). These boards should have built in NICs but they don't because Intel also wants to sell more servers.
> Marketing department asleep at the wheel.
The marketing department doesn't have to lift a finger in this market. They could quietly drop it for sale without a single announcement and it would be backordered for a year.
Source? Extremely impressive if true
These things will fly off the shelves with 0 marketing
To be honest these ubiquitous AI hardware news look to me more and more like Maybach commercials.
Not targeting me
Can I actually buy this? Should I prefer it over Blackwell?
You can tell by the shape of the board that it's meant to go into a server chassis behind a stack of 80mm blower fans. It's not meant for consumers, but you will be able to buy it used for $200 off ebay in 7 years to put in your home lab.
Wonder if there will be consumer video drivers for these GPUs once the AI hardware shortage is over. Hope to buy a gaming rig one day that doesn't cost a fortune.
I dont think it has a video out
I remember back in the day of AGP cards around 2002, you could use a 3d card on linux to render the grafix but use a different GPU for the actual output. I used XFree86 on linux with a voodoo banshee (or maybe was a riva tnt2) and a matrox mga200 for its sync-on-green support with my monitor to play games.
I know modern card architecture doesn't support this out of the box, but what is the actuall problem of forwarding a once-rendered frame to another card for output? Sureley DMA transfers PCIe->RAM->PCIe nowaday are fast enough to saturate 60fps at 4K?
You can vibe-code the drivers. Someone just did that with the Nvidia drivers for Intel Macs, bless their soul.
In 2 years once it fails and Intel shuts it down completely you mean :)
Wonder if today's ram will go up price by then...
It is going up one way or another- the price of money will go down - @5% inflation over 8 years the value of today’s dollar will be 67.7c .
> Can I actually buy this?
No. Certainly not now, maybe not ever. That is the least consumer-focused product page I've seen today.
Intel ai gpu division is so aggressively anti “small” consumer. I think it has something to do with low volume and a small sales/distribution network. The data center cards I have from amd and nvidia are grey market at best. I don’t think intel moved enough units to establish alternate channels.
How so? You can walk into Micro Center and buy a B70 right now. Intel is the only one to offer license-free SR-IOV too.
Thats not an AI GPU its a consumer product.
Serious question, would they bother?
You can work with 10 players that will buy all of them with minimal effort.
What I don't understand is: when you only plan for your product to have 10 customers, why even make a marketing page like this for it? At that point it's essentially the hardware equivalent of consultingware, something you sell to your existing contacts through outbound calls and by shoving ads through the B2B equivalent of a post-consumer sales channel (i.e. proactive outreach by your org's internal champion for that customer, to said customer.)
Because they want the hype to make the stock go up.
In fairness, they spent 5 cents generating it with Claude.
1.5 TB/s 1.3 PFLOPS with no fast multi-GPU, it's a consumer card that pretends to be a server card.
It would work great on Qwen3.8-Flash-Next or GLM-5.3 as single user decode only hits a few experts, anything with concurrent users and you'd be sitting for a while.
That's the point I believe fast local 2-4 user service for cheap on very large models.
You have to pretend it's for servers to keep investors happy, it would be banks and similar with in house servers that have to run local large models buying these for experiments and cybersecurity. Likely.
Fun to see the LLMism on the product page: "A, not B"
Productivity!
Slop and placeholder text on the page, including this one. They should have reviewed their webpage.
"What software frameworks are relevant to Intel GPU D270?
Intel GPU D270 messaging should reference familiar AI software and serving frameworks such as PyTorch, vLLM, SGLang, and OpenAI Triton, where approved. This helps developers and platform teams understand how the product fits into existing AI workflows."
Noticed this as well! Was looking for details on the software support and this FAQ entry didn't exactly instill confidence...
I like this. It means a human was involved.
Can intel gift me one?
The cost of a 480gb ram video card seems preposterous to be even marketing at this point.
Intel is going to make a card that costs $1 million dollars.. this is like industrial newsletters bleeding into a public forum
> Commercial Scale Pricing: In the industrial data center sector, battery storage systems are priced per kilowatt-hour (kWh) or megawatt-hour (MWh). Industrial battery storage installations typically cost $300,000 to $450,000+ per MWh of capacity depending on software integration, cooling systems, and structural framing.
Let's all get excited
Looks like Intel is choosing to title its new crescent island cards with the D prefix despite being Crescent Island.
Kind of disappointed that Intel is (probably) not going to release consumer class GPUs. Intel's driver stack was seriously getting mature—both on Linux and Windows. It's always nice to have a third option available.
Agreed. I was really rooting for them to deliver that third iteration to consumers, and retain them as a competitive option in the low to mid range. Pretty sad to see.
Its wild to me that intel isnt trying significantly harder on gpus
Theres so much money to be made
Why make consumer margins on consumer GPUs when they could be making enterprise margins on enterprise GPUs?
One word: Volume.
You can make a nice profit off of the enterprise space, but the big winners are the guys who put the work into servicing the consumer market.
I mean, why not both
"Both" requires more engineering staff on payroll, more wafers to fab/dice/package/test, more boards to manufacture, more inventory to manage, among other things. If their math says it isn't worth it (granted that they don't always get it right), then they don't do it.
"Both" requires allocating limited die capacity to products that have over an order of magnitude less margin.
> "Both" requires more engineering staff on payroll, more wafers to fab/dice/package/test, more boards to manufacture, more inventory to manage, among other things.
Aren't you describing how a business operates?
I mean, doesn't Intel already hire people for those tasks? Unless you are worried about a specific bottleneck, increasing headcount and production output is what businesses do.
At least that driver stack will continue to mature?
On Linux mainly, and almost strictly for compute workloads (not much need to optimize for 3D workloads given the customer base).
They're still using Xe based iGPUs, so 3D still matters.
For now. A limited run of Celestials until 2028 when Serpent Lake switches to Nvidia tiles, and a bit longer for the LTS combo (AMD style) IO+GPU die in lower end/embedded products.
Druid is still dead, consumer DGPUs are still, and workstation DGPUs also seems to be dead, and most of the Xe driver team for Windows was laid off, as well as most of the Xe chip design team.
Seems incredibly little compute for this. Is it meant to be MoE specific?
This is a brand new class of GPU that I am excited for, especially because it's rated for air cooling. I hope to see other manufacturers release GPU's with multiple of 100 GB's that can be utilized even in an at-home system's with PCIe 5 support. I probably can't afford this but knowing the engineering is sound at this scale means a similar accelerator maybe attainable.
They originally said it was going to be 120GB but that partners could ship bigger capacities. So much for that!
I do hope this has a price point that makes it somewhat competitive with a Mac! Consumers have had nearly no choices available for mem capacity, and 20 channels of lpddr5 is one way to get that capacity! Go Intel!
Both this GPU and the Mac are constrained by a lack of GPU compute. There's basically no rack-scale GPUs with 480gb on-die that have a sub-400W compute element. It's going to be incredibly prefill-constrained on denser model inference just like the Apple Silicon iGPUs are.
This GPU has 1.3 PFLOPS of FP8 compute performance and 1.5 TB/s of memory bandwidth: https://videocardz.com/newz/intel-gpu-d270-is-xe3p-crescent-...
This is almost twice better than RTX 5090. This GPU will not be prefill constrained, I feel.
But I am surprised that they were able to pack so much flops into 300W.
> But I am surprised that they were able to pack so much flops into 300W.
I'm not. When you're doing inference you typically reduce power limits anyway if you're optimizing for efficiency. Even if you don't, it doesn't spike to the max of the card.
My 3090 I power it down to 175W or so for single-digit reductions in performance for like, 60% of the wattage, something like that.
It seems like there’s a market out there that is just happy to run the models locally, even if generation is slow.
Am I wrong to assume that large MoE models are likely the point anyway? Seems like this card would love DeepSeek V4.1 Flash.
How much is it expected to cost?
Several hobos worth of organs
How many fps in Cyberpunk 2077?
That brain dance where you end up at the Skav's place in a bathtub
Can It Run Crysis?
What about Windows Vista with Aero enabled?
> Efficient token serving Dense, air-cooled, standard PCIe infrastructure helps maximize tokens per rack and watt, scaling serving capacity and keeping costs in check.
Next the tokens will be asking for tequila and caviar. Bow down to the new token overloads.
Yay more vaporware hardware I get to read about and never see or have a chance to buy.
This stuff you usually buy second hand in 6 years when the data centers get rid of them.
Ai generated no doubt