News Huawei guns for Nvidia market share in China — Ascend 910C GPU customer sampling begins

bit_user

Titan
Ambassador
This article is frustratingly thin on details, especially when tempting us with this morsel:
The Article said:
Interestingly, the 910C may be able to outperform Nvidia’s upcoming Blackwell-based B20 according to a prediction made by SemiAnalysis’s Dylan Patel.

I took a quick look at the home page, but didn't see any recent article specifically concerning Huawei's accelerators. If anyone has more information about them, please share.

I did find this, but note that the 910C specs are just guesses:

nvidia-vs-huawei-gpu-salients.jpg



I'm pretty sure the reason their fp32 TFLOPS overshadows Nvidia's by so much is that they're probably Tensor TFLOPS, whereas Nvidia only supports vector operations on fp32. The closest corresponding spec for Nvidia's H100 would be 989 TF32 TFLOPS (495 without sparsity). TF32 is a specialized number format with the range of fp32, but only the precision of fp16. So, it's not quite equivalent, but should get us into the ballpark, when trying to interpret the above data.
 
Oct 1, 2024
1
0
10
The US government really thought Chinese companies would sit back and do nothing about the sanctions. What China loves most is making money. They will pursue any innovation possible to gain market share in any field. These sanctions are so shortsighted. I guess it's just a four-year window perspective to get some votes. Gotta love democracy, lol.
 

bit_user

Titan
Ambassador
The US government really thought Chinese companies would sit back and do nothing about the sanctions.
You underestimate how far behind China is on lithography, especially if they can't rely on equipment, materials, tools, and software from the west. While they're trying to play catchup, everyone else is continuing to make further progress. It will really take quite a while for China to close the gap, in spite of whatever they try to claim (such as their purported 5 nm capability, which is a stretch they managed to achieve via double-patterning on equipment procured pre-sanctions).

Also, I have a feeling you're going by the headlines and overlooked this caveat:

"Additionally, if the 910C doesn’t make substantial improvements over the 910B, with respect to yields, that would also be a significant problem for Huawei, and at minimum would reduce the new chip’s profitability."

It's one thing just to make a chip. Being able to scale up volume production is what's really needed and that's where there are some serious doubts.