China's AI chip companies live in fear of being dominated by NVIDIA

This article is based on publicly available materials and is for informational purposes only and does not constitute any investment advice.

Recently, I have seen some people in the chip industry express some views on domestic AI chips, which I think is very valuable, and I will briefly refine it here for your reference.

1. Measured data is important

It is of little significance to look at the publicity computing power, and it is necessary to look at the actual measured data.

2. Software ecology is the key

The biggest difficulty of AI chips is that they need to build their own software stacks and establish an ecosystem.

3. Performance improvement of T4/A10/V100/A100

Generally, the 1 and 2 benchmark networks have been tuned to the extreme, which does not mean that all network models have such performance, and the versatility of the chip is not mentioned at all.

4. Why go get ZF's cluster, Xinchuang project?

Because these projects are not demanding, running 1 or 2 models can intersect

5. How to judge the real commercial landing?

See if the Internet manufacturer purchases the board in bulk

Of the previous few, the second is the most difficult place for AI chips.

I quote a passage from Dr. Xia Jingjing, an architect at Huawei Kunpeng, a senior and well-known figure in the chip design industry, as a reference:

"DOCA is the only programming framework for NVIDIA DPUs.

DOCA is compatible with future multi-generation DPU evolutions.

DOCA includes offloading, acceleration, isolation, support for IaaS to supercomputing to disaggregation DC...

DOCA is to DPUs what CUDA is to GPUs...

Have you forgotten the horrors of CUDA? With so many AI chip companies in recent years, who hasn't been scared to cry by CUDA in the middle of the night?

After two years, DOCA matures, and when we look at DoOVS, DoSDI, and DoZIP, we will be powerless to resist.

So when DOCA is not yet mature, YOCA? XOCA？ Where?

It's not as simple as hiring two driver engineers, it's a framework that requires the process of distilling a subtle ISA from HAC, NP, CPU, etc. to an abstract, complete, and stable kernel/API. Compared to the framework in the field of AI, the difficulty is not lower.

This requires recruiting Chen Tianqi, Li Mu, Jia Yangqing in the field of interconnection (compared to AI are networks) ... And there are actually very few talents related to the industry, and they have long been attracted by the fiery AI..."

Of course, Dr. Xia Jingjing mainly talked about the software ecology of DPU, but the truth and the software ecology of AI chips should be the same.

Finally, I have no special research on Cambrian, and if you were to ask me whether Cambrian had investment value, this is not a question I can answer. (Author: Icefighter)