Hacker News new | ask | show | jobs
by culi 7 days ago
They've achieved self-sufficiency in >14nm chips in remarkable timing. Unfortunately for DeepSeek, it's the <14nm chips that are needed for massive training tasks. I wouldn't be surprised if China backs down and lets them purchase the chips given that they are still years away from being able to make them themselves. Either that or the gov't steps in and forces them to share resources

And even if Huawei's Ascend 910C can compete with NVIDIA's H200, CUDA is still a large moat
3 comments
seewhydee 7 days ago
Bypassing the CUDA moat is, in fact, one of the major tasks Deepseek set for itself. Their efforts in this area are likely one of the main reasons for their slow release cadence, culminating in their v4 inference setup that runs on Huawei Ascend chips.
link
tancop 6 days ago
there are already two independent drop in cuda emulators. zluda is open source with extra features like game physics, scale is production ready and focused on inference workloads. both of them can run pytorch.

and for a company full of brilliant engineers with zero respect for american law its not that hard to make one more by themselves.

link
HarHarVeryFunny 6 days ago
China is already making it's own chips. The one most commonly discussed is the Huawei Ascend 950 (which exists and is in production). NVIDIA say the 950 is about equivalent to an NVIDIA H100, and DeepSeek in this leaked meeting are saying that about 4 950 cards are equivalent to 1 NVIDIA "B" one (Blackwell B200 I think).

Apparently production is constrained though, with DeepSeek saying they were only able to get an allocation of 16,000 950 cards. They have plenty of money, but are limited by the number of cards available to buy, and therefore the size of models they can train. They mentioned having a 20,000 GPU cluster. Other companies like Kimi, with a ~3T param model, clearly have a lot more compute, probably all NVIDIA.

DeepSeek have their own "TileLang" software, which sounds a bit like the Triton kernel compiler, and isolates them from the diffences between NVIDIA and Huawei chips - they are deliberately avoiding any CUDA dependency.

link