Neat blog! I was intrigued by this bullet point mentioned in passing:
>my four hard drive USB 3.2 ZFS raidz2 array with four 24 TB drives
Can you speak more about this? Which USB array did you choose? How well does it work? I've been slowly planning a transition away from my power-hungry surplus enterprise gear in the 19" rack towards a smaller, quieter, lower power setup ... but storage is the real kicker right now. I have a 12x18TB array in raidz2 built into a 1U NAS case, and I just can't quite figure out a better way to package something like that. I would need three USB arrays if I want to reuse the existing drives, which I think I do given how expensive storage is today.
I strongly considered it, but the one thing that scares me away from wanting to do the spark is you basically have to use nvidia's linux (from what I've read) and it doesn't appear the nvidia is interested in upstreaming their kernel changes.
I'm avoiding where possible buying electronics where support is controlled by the manufacturer and not me.
Many other distros almost work out of the box (as in they boot and run without any modifications). The custom kernel patches you mentioned address mainly non-critical bugs such as a bug where the Realtek r8127 stops working after a reboot (but it works if you turn it off and on again) [1] [2]. I'd consider it in a way better state than trying to run other Linux distros on certain device tree-based devices like, say, Qualcomm Snapdragon machines. The regular NVIDIA drivers with the open source kernel modules work just fine. Talos Linux supports DGX Spark since version 1.12 [3]. I also know of people using Fedora and nixOS successfully.
If I understand your comment correctly, I think he addresses that in the first few paragraphs:
> The DGX Spark runs “DGX OS” but it is in fact just plain old Ubuntu 24.04 with some additions. If you want, you can just install another Linux distribution easily (Fedora works well), although there may be a couple of weird bugs with the Realtek Ethernet driver so the NVIDIA version of the Linux kernel has a couple of patches. Unlike some other ARM devices, the DGX Spark is all ACPI rather than device tree based, so regular Linux builds for arm64 work just fine.
It's interesting how many of these issues don't appear to be specific to the DGX Spark but to the standard "Nvidia GPUs suck on Linux" type of issues that afflict a lot of people.
I'm genuinely disappointed with my Spark. I don't know how anyone can claim it performs decently with LLMs or diffusion models. Back when I worked in VFX in the early 2000s, we had a saying: "Render time is coffee time" and if you try to run this thing with a usable context size, you'll be drinking a lot of coffee. Most of the optimizations it relies on for inference simply aren't available for training, so it crawls like a snail on almost every model. An RTX 6000 Blackwell would have been the better investment for an AI enthusiasts and for general computing there are cheaper offerings.
The DGX Spark hits a sweet spot for me where it can simultaneously work for general computing and run local LLM inference fast enough for some hobbyist dabbling. The RTX 6000 Pro Blackwell is more than twice the price (it has increased quite a bit recently from $8000 to $11600), not to mention the "rest of the PC" needed to get it working, so it's not really a fair comparison. Compared to other 128 GB unified memory devices like the Mac Studio and the Strix Halo, the DGX Spark fairly priced in my opinion.
If you bought it for inference you made a mistake. They aren't good at that. Use it to train models and experiment with ML. It's much better at that.
If you just want local inference buy a Mac.
If you bought early on, like I did, the Spark is probably worth double what you payed now. I think I paid $3,000 retail for mine and the last time I looked they were fetching close to $6k on ebay. I'm not sure if that's still the case, but you can buy a very nice Mac with $6k.
Are you saying that it's slow for training? Sorry, your comment is confusingly worded to me.
I've not had any issues in that regard, but I'm working with LLM's not training diffusion models. Are you following one of the Nvidia provided recipes or inventing something on your own? The last time I looked into it they benchmarked very well, but we both know that doesn't always mean much.
Neat blog! I was intrigued by this bullet point mentioned in passing:
>my four hard drive USB 3.2 ZFS raidz2 array with four 24 TB drives
Can you speak more about this? Which USB array did you choose? How well does it work? I've been slowly planning a transition away from my power-hungry surplus enterprise gear in the 19" rack towards a smaller, quieter, lower power setup ... but storage is the real kicker right now. I have a 12x18TB array in raidz2 built into a 1U NAS case, and I just can't quite figure out a better way to package something like that. I would need three USB arrays if I want to reuse the existing drives, which I think I do given how expensive storage is today.
I strongly considered it, but the one thing that scares me away from wanting to do the spark is you basically have to use nvidia's linux (from what I've read) and it doesn't appear the nvidia is interested in upstreaming their kernel changes.
I'm avoiding where possible buying electronics where support is controlled by the manufacturer and not me.
Many other distros almost work out of the box (as in they boot and run without any modifications). The custom kernel patches you mentioned address mainly non-critical bugs such as a bug where the Realtek r8127 stops working after a reboot (but it works if you turn it off and on again) [1] [2]. I'd consider it in a way better state than trying to run other Linux distros on certain device tree-based devices like, say, Qualcomm Snapdragon machines. The regular NVIDIA drivers with the open source kernel modules work just fine. Talos Linux supports DGX Spark since version 1.12 [3]. I also know of people using Fedora and nixOS successfully.
[1] https://github.com/NVIDIA/NV-Kernels/compare/ea55925ab430f1e...
[2] https://forums.developer.nvidia.com/t/realtek-r8127-ethernet...
[3] https://github.com/siderolabs/talos/issues/12170
If I understand your comment correctly, I think he addresses that in the first few paragraphs:
> The DGX Spark runs “DGX OS” but it is in fact just plain old Ubuntu 24.04 with some additions. If you want, you can just install another Linux distribution easily (Fedora works well), although there may be a couple of weird bugs with the Realtek Ethernet driver so the NVIDIA version of the Linux kernel has a couple of patches. Unlike some other ARM devices, the DGX Spark is all ACPI rather than device tree based, so regular Linux builds for arm64 work just fine.
A good instinct. There are a lot of things a $500 AMD GPU can do in linux that a $5000 DGX cannot.
I know nothing about this topic but your comment piqued my curiosity. What would a $500 AMD GPU do better than DGX?
It's interesting how many of these issues don't appear to be specific to the DGX Spark but to the standard "Nvidia GPUs suck on Linux" type of issues that afflict a lot of people.
it's actually very stable on x86_64 these days, even with optimus.
Super-powerful (if rather pricy) Linux desktop that happens to play games while doing everything else... that man is living my dream :'D
I would avoid the DGX Spark. For that price and its performance on running local models it is a complete scam. This tweet says it all [0]
[0] https://xcancel.com/petergostev/status/1978230978725507108
I'm genuinely disappointed with my Spark. I don't know how anyone can claim it performs decently with LLMs or diffusion models. Back when I worked in VFX in the early 2000s, we had a saying: "Render time is coffee time" and if you try to run this thing with a usable context size, you'll be drinking a lot of coffee. Most of the optimizations it relies on for inference simply aren't available for training, so it crawls like a snail on almost every model. An RTX 6000 Blackwell would have been the better investment for an AI enthusiasts and for general computing there are cheaper offerings.
The DGX Spark hits a sweet spot for me where it can simultaneously work for general computing and run local LLM inference fast enough for some hobbyist dabbling. The RTX 6000 Pro Blackwell is more than twice the price (it has increased quite a bit recently from $8000 to $11600), not to mention the "rest of the PC" needed to get it working, so it's not really a fair comparison. Compared to other 128 GB unified memory devices like the Mac Studio and the Strix Halo, the DGX Spark fairly priced in my opinion.
If you bought it for inference you made a mistake. They aren't good at that. Use it to train models and experiment with ML. It's much better at that.
If you just want local inference buy a Mac.
If you bought early on, like I did, the Spark is probably worth double what you payed now. I think I paid $3,000 retail for mine and the last time I looked they were fetching close to $6k on ebay. I'm not sure if that's still the case, but you can buy a very nice Mac with $6k.
Mhm... maybe read my full comment?
Are you saying that it's slow for training? Sorry, your comment is confusingly worded to me.
I've not had any issues in that regard, but I'm working with LLM's not training diffusion models. Are you following one of the Nvidia provided recipes or inventing something on your own? The last time I looked into it they benchmarked very well, but we both know that doesn't always mean much.
Yes. Waiting for the Windows Version myseflf. RTX Spark Desktop.
I guess copilot needs all the help it can get?