- For those of you wondering if this fits your use case vs the RTX 5090 the short answer is this:
The desktop RTX 5090 has 1792 GB/s of memory bandwidth partially due to the 512 bit bus width, compared to the DGX Spark with a 256 bit bus and 273 GB/s memory bandwidth.
The RTX 5090 has 32G of VRAM vs the 128G of “VRAM” in the DGX Spark which is really unified memory.
Also the RTX 5090 has 21760 cuda cores vs 6144 in the DGX Spark. (3.5 x as many). And with the much higher bandwidth in the 5090 you have a better shot at keeping them fed. So for embarrassingly parallel workloads the 5090 crushes the Spark.
So if you need to fit big models into VRAM and don’t care about speed too much because you are for example, building something on your desktop that’ll run on data center hardware in production, the DGX Spark is your answer.
If you need speed and 32G of VRAM is plenty, and you don’t care about modeling network interconnections in production, then the RTX 5090 is what you want.
by jasoneckert
2 subcomments
- I've got the Dell version of the DGX Spark as well, and was very impressed with the build quality overall. Like Jeff Geerling noted, the fans are super quiet. And since I don't keep it powered on continuously and mainly connect to it remotely, the LED is a nice quick check for power.
But the nicest addition Dell made in my opinion is the retro 90's UNIX workstation-style wallpaper: https://jasoneckert.github.io/myblog/grace-blackwell/
- You can get two Strix Halo PCs with similar specs for that $4000 price.
I just hope that prompt preprocessing speeds will continue to improve, because Strix Halo is still quite slow in that regard.
Then there is the networking. While Strix Halo systems come with two USB4 40Gbit/s ports, it's difficult to
a) connect more than 3 machines with two ports each
b) get more than 23GBit/s or so per connection, if you're lucky. Latency will also be in the 0.2ms range, which leaves room for improvement.
Something like Apple's RDMA via Thunderbolt would be great to have on Strix Halo…
by kristianp
2 subcomments
- I know it's just a quick test, but llama 3.1 is getting a bit old. I would have liked to see a newer model that can fit, such as gpt-oss-120, (gpt-oss-120b-mxfp4.gguf), which is about 60gb of weights (1).
(1) https://github.com/ggml-org/llama.cpp/discussions/15396
- IMHO DGX Spark at $4,000 is a bad deal with only 273 GB/s bandwidth and the compute capacity between a 5070 and a 5070 TI. And with PCIe 5.0 at 64 GB/s it's not such a big difference.
And the 2x 200 GBit/s QSFP... why would you stack a bunch of these? Does anybody actually use them in day-to-day work/research?
I liked the idea until the final specs came out.
by cat_plus_plus
1 subcomments
- I have a slightly cheaper similar box, NVIDIA Thor Dev Kit. The point is exactly to avoid deploying code to servers that cost half a million dollars each. It's quite capable in running or training smart LLMs like Qwen3-Next-80B-A3B-Instruct-NVFP4. So long as you don't tear your hair out first figuring out pecularities and fighting with bleeding edge nightly vLLM builds.
by kachapopopow
1 subcomments
- Dell fixing issues instead of creating new ones? That's a new one for me. Would rather still not deal with their firmware updaters thought.
- Absent disassembly and direct comparison between a DGX Spark and a Dell GB10, I don't think there's sufficient evidence to say what is meaningfully different between these devices (beyond the obvious of the power LED). Anything over 240W is beyond the USB-C EPR spec, and while Dell does have a question ably-compliant USB-C 280W supply, you'd have to compare actual power consumption to see if the Dell supply is actually providing more power. I suspect any other minor differences in experience/performance are more explainable as the consequences on increasing maturity of the DGX software stack than anything unique to the Dell version; particularly any comparisons to very early DGX Spark behavior need to keep in mind that the software and firmware have seen a number of updates.
- Seems you are paying the Dell tax of 15%. The same setup is $4K from NVidia, Lenovo and $3K for 1TB at Asus.
https://www.dell.com/en-us/shop/desktop-computers/dell-pro-m...
- I have NixOS running on my DGX Spark: https://github.com/graham33/nixos-dgx-spark, would be interested to know if the USB image also boots on the Dell Pro Max GB10.
by colordrops
2 subcomments
- I assume they didn't fix the memory bandwidth pain point though.
- A nice little AI review with comparison of the CPU/Power Draw & Networking would be interested in seeing a fine-tuning comparison too. I think pricing was missing also.
by postalrat
1 subcomments
- Spark's biggest paint point is the price. Does it fix that?
by barelysapient
1 subcomments
- Great article but would be nice to see how larger models work.
- Jeff, This is the second time you have been given a prosumer level cluster pretty much built for local LLM inference and on both occasions you have performed benchmarks without batching.
If you still have the hardware (this and the Mac cluster) can you PLEASE get some advice and run some actually useful benchmarks?
Batching on a single consumer GPU often results in 3-4x the throughput. We have literally no idea what that batching looks like on a $10k+ cluster without otherwise dropping the cash to find out.
- It's a product without a purpose.
Hacker News