News Asus' mini supercomputer taps Nvidia Grace Blackwell chip for 1,000 AI TOPS

Hasn't released the product page. Another GPT article, or just google translate and no editorial oversight?
I seem to recall having noticed some unusual word usage by this author, previously. I'm pretty sure it's not due to LLM or probably even automated translation, since one notable characteristic of LLMs is a strong preference for more common words, phrases, and writing structures. They actually tend not to sound quirky.

I don't usually comment about it, unless it's bad enough to confuse readers or convey a message different than probably intended. Of all my concerns about the content on this site, I'd have to rank English grammar and style near the bottom. As long as the writing is clear enough and not confusing, it's really the information that I care about.

BTW, the complaint I'd have about this article is that it didn't put "supercomputer" in quotes. Someone at Nvidia seems to be stuck back in the 1990's or 2000's and believes everything which fits that ancient definition of a supercomputer can be labelled as one. Even then, I'm pretty sure its fp64 performance is rubbish, so you'd have to qualify it as an AI "supercomputer".
 
Last edited:
For reference, the 5090 has about 208 TOPs @ fp4.
Sorry, what? Where did you get that? I think you're probably looking at vector int8 dot product + accumulate (i.e. DP4A).

By comparison, its tensor cores can do 419 TOPS @ fp16 (838 with sparsity).

Nvidia claims the RTX 5090 does 1676 tensor TOPS @ fp4 (3352 with sparsity). And we know their GB10 numbers from this article are with sparsity, because Nvidia will always quote the highest number they can possibly justify.
 
I want to Project Digits compared against Strix Halo in everything.
If we're talking native workloads, then I expect Digits would stomp all over it on AI. Digits is made primarily for AI, while Strix Halo sort of stumbled into that niche.

In terms of CPU workloads, the X925 cores are way bigger than their predecessors, the X4 cores (yes, Arm switched naming conventions). They were designed to go up against Apple's P-cores. 10 of those bad boys should pack a punch, even though they're only complemented by 10 A725 cores. It will be enough to pull ahead in a lot of things. I'm not sure who would take the overall lead.
 
If we're talking native workloads, then I expect Digits would stomp all over it on AI. Digits is made primarily for AI, while Strix Halo sort of stumbled into that niche.
I haven't seen anything that isn't Nvidia marketing material, so I'll reserve judgment.

Comparing the two is obviously of interest due to their similar AI target market, pricing, CPU core counts, and (up to) 128 GB of LPDDR5(X).

If Digits/Spark is stomping Halo by 5-10x or something, maybe Strix Halo pricing needs to come down and Medusa Halo needs to use RDNA4 with big TOPS boosts. Meanwhile, Medusa Point may be stuck at RDNA3.5.

10 of those bad boys should pack a punch, even though they're only complemented by 10 A725 cores. It will be enough to pull ahead in a lot of things. I'm not sure who would take the overall lead.
ARM X-cores have lagged behind Apple P-cores. They made big claims about X925 single-threaded uplifts, but Zen 5 cores can easily clock over 30% higher. Strix Halo has 16 to Digits' 10+10 cores. I think it will be similar at best but more software and operating systems will run without issues on x86.
 
Comparing the two is obviously of interest due to their similar AI target market, pricing, CPU core counts, and (up to) 128 GB of LPDDR5(X).

If Digits/Spark is stomping Halo by 5-10x or something, maybe Strix Halo pricing needs to come down
I contend that Strix Halo wasn't primarily designed for AI. Also, it's x86 and a laptop CPU, while GB10 is neither of those things.

So, no. The people putting it in mini-PCs are doing so opportunistically. Maybe Digits shrinks the market for such machines, but we so far have no indication that Ryzen AI Max is under threat in its home territory of premium non-dGPU laptops.

Medusa Halo needs to use RDNA4 with big TOPS boosts.
Yes. Both for AI/FSR4 and also for ray tracing, AMD needs to push for getting RDNA4 in their laptops. Although, you're probably right that it's too late for the mainstream tier of Zen 6 SoCs.

ARM X-cores have lagged behind Apple P-cores. They made big claims about X925 single-threaded uplifts, but Zen 5 cores can easily clock over 30% higher.
If they can come close to Apple on IPC, then X925 cores will definitely bring some heat to Zen 5.

Strix Halo has 16 to Digits' 10+10 cores. I think it will be similar at best but more software and operating systems will run without issues on x86.
When Zen 5 can get 32 threads cranking or exploit its AVX-512, I think it should do well. On scalar and more lightly-threaded workloads, I think it's in trouble.