Beside those Youtubers that only know OIlama there is at least one review from people who know what to do with those boxes:
…well they also used Ollama, but they also build SGLang.
When looking at those numbers collected by LMSYS running a preproduction model (I’m aware of that it is not the final version of drivers etc.) why would someone stack two of them together to run a dense 400B model If a 70B model only allows this performance.
But may be FP4 and MoE is the future… 🫢
Looking forward to more tests for different use cases.
1 Like
Came here to post this. Glad to finally get some numbers. Cancelling my pre-order after watching that review. Seriously disappointing for the price tag. Too little, too late…
1 Like
I too am disappointed in the benchmarks.
The review said “The DGX Spark isn’t meant to replace your RTX 5090”
So what will you get instead?
I’m thinking of spending 2.5x more for an RTX Pro 6000 and getting 8x tps.
2 Likes
llama.cpp to the rescue?
So, is everyone as confused as me with this product? I was invited to place my order this morning, and $4k I really don’t see where this fits in anywhere given its performance.
The ONLY scenario I see this as a useful product is for development on the go. And, who is honestly doing this where Cloud resources couldn’t be used instead? At this price, I seriously don’t understand what the point is.
Initially I was thinking this might be an amazing system to inference models for me to use for other development projects without needed subscriptions and the limits those have. The AI in a box concept is what they were initially selling this as. Then it was for development, and only for those on the go, and only to validate some processes to then push to larger DGX systems. Then they massively spiked the price.
Memory is very slow, performance doesn’t seem great at much of anything for this price. Am I missing something? Did this product just miss its mark? THis thing is being pushed and advertised on by stores and youtube channels as if this thing is for general consumers and AI users as well as developers. This doesn’t seem to be the case. I have not canceled the pre-order yet, but I am not sure I can bring myself to drop $4k for this, even if I want it because it looks cool. Even then, how quickly will these dive in value? Now that the race is on, AMD has their version of this at $2k with much better CPU performance and full x86.
Helpy
The SGLang reviewer reported that there is no optimized kernel yet for FP4/sm121a (GB10), but they will support it as soon as it becomes available.
I think just like the other Blackwell GPUs it will take some time to get the necessary optimizations for the full potential of the GB10.
As for the use cases for the box the review of Level1Techs is worth watching. As they don’t stick to inference only.
3 Likes
FWIW, when I cancelled my order this morning, the response said even if it has already shipped you can return for a full refund incl shipping. Don’t remember exactly but I was slightly over #10000 reservation.
You’ll be seeing these on eBay soon.
I use an Mac M1 Max currently. For sure its faster and has more memory.. but to be honest i guess i will wait for the M5 Max . I expected more.
3 Likes
Here is another one I saw: https://www.youtube.com/watch?v=rKOoOmIpK3I
Did you actually get an email to purchase? as I am around 1000 and I did not get a email to purchase, or did you just cancel your reservation. And did you reserve through Nvidia or another supplier?
Yes, I received an email with a link to purchase. I reserved mine back in March through Nvidia. The purchase was easy and the cancellation was even easier. I’m surprised you haven’t received a link yet.
1 Like
That’s why I would always prefer llama.cpp over Ollama.
1 Like
On a positive note, a typical off-site training session is $2500. The DGX Spark does bring all the components together in one small package that will likely be well supported. I worked for Tandy when the first hard drives showed up in retail stores – $2500 for 5mb. That was back in 1982 when I was making $5 an hour. I guess as an alternative I could purchase a…. oh yea…. Nothing else compares….
2 Likes
Any ComfyUI or other image/video generation benchmarks out there?
This review shows image generation. It’s a well done review by a guy who bought his own spark.
2 Likes
They also tested vLLM.
I’ve hacked away for hours with a DGX Spark for the last three days. My review is this:
In 4th grade, I read my first piece of non-fiction, which was the IBM/Microsoft Basic manual. I learned to program reading that book, and with my grandfather’s original IBM PC. Now, decades later, the DGX Spark is the first computer I’ve owned that reminds me of that experience.
For more context, I don’t “need” this machine. I’m fortunate that I have access to thousands of A100/GH200/MI250X GPUs in various clusters which I can use for work. But, a personal computer on my desk that can do what the Spark does, is a better platform on which to learn.
3 Likes
In 1984 Apple released the MacIntosh All-In-One computer for $2,495 (about $7000 today). My dad bought a basic one for the Family. It had a big influence on all of us as it was exciting, challenging, and expensive… it was an inflection point in our lives. I am just saying this DGX moment feels similar to me…
Get your kids/grand-kids a DGX and invest in their potential… Help them embrace AI and teach them not to fear progress… Show them how to use it – explain what it does… Vibe code something together… inspire them to explore their own gifts and potential with this new tool.
If you or your kids are interested then I hope you find a away… the Macintosh was once strange and expensive too.
7 Likes