AMD s strix halo bude mat trosku mensi vykon, ale vseobecnejsie pouzitie...sa uvidi, o par tyzdnov

Po vydání byly nabízeny krabičky se Strix Halo.
32 GB od $1200
64 GB od $1500
128 GB od $2000

Ale některým to přišlo drahé, protože to nemá CUDA. Oficiálně.

lol.. medium..

'' Here’s the thing I didn’t understand in October: I was evaluating the Spark like a GPU.

It’s not a GPU.

The DGX Spark is a capacity play, not a throughput play. It’s not trying to generate tokens faster than your RTX 5090. It can’t. The memory bandwidth physically won’t allow it. What it can do is hold entire models that would crash your 24GB graphics card. A 120B parameter model in NVFP4 fits in memory. Try that on an RTX 5090.

But I only understood this after the January update forced me to rethink everything.

The Verdict: A Platform, Not a Product
The DGX Spark is the most misunderstood piece of hardware in AI right now.''

to zni jak povinna sebekritika za komousu.. damage control od nVidie jak vysitej.. ((:

a reakce v diskusi tomu odpovidaj:

'' Until the support for vLLM is in a place, where it actually leverages NVFP4 fully without the community needing to hack dependencies together, I would advise people to be cautious. Many blackwell optimisations are also not possible due to GB10 being based on sm121. ''

'' Subject: Your 5,000 words won't hide the painful truth about NVIDIA’s artificial scarcity
To the Author,
You’ve written 5,000 words, but they cannot cover up the painful truth. The "man in the leather jacket" has disappointed us once again. He is simply a master of managing the politics of scarcity. Every RTX card could easily have more VRAM, but they don't because he won't allow a collision with $100k enterprise cards.
Returning to the Spark and your article: I don't know what you need Nvp4 for, but it’s not for anyone serious. The scale of hallucinations statistically blurs every third response. Researchers and scientists need FP16—or, in cases of extreme budget constraints, 8-bit.
This "champagne gold" hardware will never handle that for obvious reasons: a 273 GB/s bandwidth. No amount of marketing talk or "January updates" will turn a lie into the truth.''