Patrick Bolen

22 de julio de 2026

Great product and entry level into big models

Great product but beware leads you to buying more and more it's addictive. Core review wish there was faster bandwidth or cheaper price but it is what it is.

B. Jester

25 de abril de 2026

Unified RAM with tradeoff of some speed

MSI's version of the DGX Spark is definitely one of the best. I've seen others complain about overheating and shutdowns, but I've had none of that. Despite that, I decided to put a Wathai 140mm x 25mm fan underneath for extra airflow. Under concurrent inference load, it stays under 80C and performs well.Upon receiving it, I noticed significant disk usage. The manufacturer left disk burn-in files, totaling a few TB, on the device, but once I found them, deleting them resolved that.The DGX platform itself is a mixed bag. Nvidia has been slow to push through better support for the device in libraries like vLLM and llama.cpp. The best performing setup I've used for coding agents so far has been using Intel's qwen3.5-122b-a10-int4-autoround. With that I get generation speeds of 30 t/s, and upwards of 65 t/s under currency=4. Prompt processing speeds are not an issue, usually well into the thousands of t/s.Overall, it's a solid device which handles concurrency really well. Although, to be real, what you're really getting is the 128GB of unified RAM at the cost of slower bandwidth. If you want raw speed, and are willing to sacrifice a little memory and concurrency, go with a Blackwell card. I am hopeful though that libraries, and quants like NVFP4, will be able to maximize performanceVer más

15 personas la encontraron útil

Haruyuki Shini

Not a Frontier-LLM Competitor, but...

This computer is capable of running NVIDIA's Nemotron 3 Super model LLM, which punches way above its weight class. You can easily get a 400k context window while also running Stable Diffusion and generating images that are 1080p or bigger. If you make it a dedicated LLM machine, you can easily get over a 1 million token context window. The only thing I don't like about this machine is the Ubuntu-default stuff like GNOME (ugh!), but Linux is Linux. You can really set this up however you like. I recommend setting it up and then using SystemD to disable gdm for even more RAM headroom (do not do this unless you know *exactly* what that means).It's also pretty quiet. Running both the LLM and Stable Diffusion at the same time pushed it into somewhat uncomfortable thermal territory, but running just one at once tends to stay in the 70s-degree range. The WiFi is good too. Usually I'm a WiFi hater, but honestly the amount of network data throughput this machine needs is rather low. It works great while only being connected to power.Ver más

8 personas la encontraron útil

Rob M.

4 de junio de 2026

More impressive than I initially thought

This is a review of the DGX Spark / GB10 platform in general, not MSI's packaging of it. I am updating my original 1-2 star review and bumping it up to 4. After a month of working with it nearly around the clock, I have found some ways to make it do useful things that my Strix Halo Max+ 395 cannot.Performance of GGUFs through llama.cpp on the GB10 is not much of an improvement over when you'll see on Max+ 395 device, which costs half as much, though it does perform better with the large context prompt processing, just not the inferencing. tensorrt-llm simply is not ready for the GB10 yet, and it's difficult or impossible to get that engine to do anything interesting at all. But vLLM with a properly tuned nvfp4 can be most impressive. A lot of work is required to get a 70B llama 3.3 model into a nvfp4 format optimized for vLLM. Converting from bf16 to nvfp4 cannot be done on the device itself and requires 1-2 hours of B200 or 2xH100 rental time, since the .safetensors will not fit into 128GB. That step will cost about $10. The rest can be done directly on the GB10. This is not the forum for writing instructions on how to go about that, but with a few scripts that Claude or any AI can help you with, you can get appropriate k/v weights into the model, fix some data types, and wind up with a 70B llama 3.3 nvfp4 quant that can handle up to the full supported 128K token context with speeds above 5t/s. I'm getting that speed at about 30K in use. I imagine it will degrade some with additional growth, but this is still better than llama.cpp can do and infinitely better than the Max+ 395 can do. vLLM for GB10 has a special docker container for the 119B nvfp4 Mistral 4 Small MoE model that runs so fast it makes your head spin! 30-40t/s from a 119B on a box you can run in your house is pretty impressive indeed! The Mistral model is not tuned for my use case, but with a little training it might be.With all that said, I STILL think the GB10 is a bit overhyped and not ready for every use case. For this reason, I cannot give it 5 stars. It is immature and the software landscape is in such a state of flux that every document you find was obsolete a week before you read it. But if you like that sort of thing ... have at it! The GB10 does have enormous potential, and it can perform impressively even today.I've been working on this review for a while. My initial one gave it 1 star; I was so furious with my inability to get it to do *anything* Nvidia claimed it could do that I opened a return request. Then I discovered that my Max+ 395 was not able to process the prompt context I had generated on the GB10 without causing a timeout that I could not resolve. So I bumped the review to 2 stars and delayed my return. Then I got vLLM to do these legitimately impressive things, and I had to revise my review altogether, and I've decided to keep the device.I apologize to any who factored my earlier, more scathing, reviews into their decision about whether or not to buy this product. I promise that these are my final words on the topic.

Roman H

23 de mayo de 2026

Great!

Amazing machine!

santa

24 de febrero de 2026

NVIDIA's Pytorch container works on EdgeXpert

Pros: better thermal design. This small AI box seems to work fine during heavy load without thermal throttling.Cons: Official distribution of Pytorch does not support Blackwell GB10 yet. I could not install GB10 enabled Pytorch(I have tried some patches and how-to, but failed.) . If your task will run in a container, you can use GB10 enabled Pytorch container from NVIDIA. This cotainer works fine on EdgeXpert.Ver más

7 personas la encontraron útil

Javier

21 de mayo de 2026

Útil pero requiere trabajo

Cumple con lo esperado y es definitivamente útil para poder trabajar con modelos de LLMs más pesados (fine-tuning, continual pre-training, entre otras técnicas). Sin embargo, al ser el chip bastante nuevo/novedoso no tiene buen soporte con librerías de Python y cuesta un viaje resolver incompatibilidades para que realmente se le saque provecho a este hardware. Para inferencia, es definitivamente más lento que la línea de mayor VRAM RTX.Ver más

2 personas la encontraron útil

Chris JB Gutierrez

28 de abril de 2026