Comment by adrian_b
1 month ago
Their production capacity is constrained, but not so constrained as to prevent them to demonstrate a supercomputer faster than any US supercomputer, made with custom CPUs designed and fabricated in China.
The limited production capacity means that they will not become exporters of CPUs or GPUs any time soon, but it does not mean that they cannot satisfy their internal necessities in any critical sectors. The technology that they use for CPUs and GPUs is about at the level that TSMC was 5 years ago. That may seem much, but there are a lot of people who are quite satisfied with computers older than that, and do not intend to upgrade them soon.
The main consequence of the older technology is a higher power consumption, because otherwise their designs are quite up-to-date in the attainable throughput. Therefore their datacenters must consume more energy, but that is hardly a problem in China.
The chip export controls have only forced them to design and fabricate their own CPUs and GPUs, instead of buying them from USA, and now they have become able to do this.
Without USA forcing them to do this, they might have remained dependent on imports from USA for decades, but now USA has lost the opportunity to ever play this card again. Thus USA has lost a leverage that could have been useful in a real conflict.
Such export controls are effective only for a short time, so they must be used only when there is a clear immediate goal.
China can produce chips at "7nm", that's more like 8-10 years behind. My M1 macbook pro is 6 years old, and that chip was made using a "5nm" process. The advanced US chips are now on "2nm". Explain to me how China is going to produce a computer that beats US supercomputers using 10 year old technology.
China has already produced a supercomputer that has beaten all US supercomputers, which was made public a month ago, and now it occupies the first place in Top 500.
Not only it has the greatest absolute performance, but it also reaches a much greater percentage of its maximum theoretical performance, due to having a very high memory bandwidth, 8 TB/s per socket (5 times higher than for the fastest US CPUs, faster than for any older US GPUs and equal to the fastest US datacenter GPUs of last year from NVIDIA and AMD), and due to using the Arm Scalable Matrix Extension (SME) instruction set, instead of using GPUs.
The previous AMD datacenter GPUs, which are used in the fastest US supercomputers, had a lower memory bandwidth than the new Chinese CPUs. Only this week AMD has introduced 2 new datacenter GPUs, one for AI and one for HPC, which are considerably faster, but it will take some time until such datacenter GPUs will be deployed in great quantities. NVIDIA has introduced Rubin for datacenters, which also has a greatly improved memory bandwidth (22 TB/s), but it uses extremely expensive memory (with a very high clock frequency) and it is usable only for AI, not for HPC, where AMD remains without competition, except for the new Chinese CPUs.
Recent Chinese chips have been cut and analyzed with electron microscopes and they already have finer metal and gate pitches than anybody has previously reached without EUV lithography, similar to the TSMC "5nm" processes. They are denser than CPUs like the Intel "7 nm" Ice Lake and Tiger Lake and somewhat denser than AMD Zen 3 CPUs.
Their fabrication throughput is much lower than at TSMC, because they must use multiple exposures to reach such a resolution, and they must have a poor fabrication yield, but there is a huge difference between being able to make only slowly something and not being able at all to make it.
Their technology is about at the 2021 level, i.e. 5-year old, not 10-year old, but their design capabilities are much better than their fabrication capabilities. Thus products like their new CPU for HPC offer top performance competitive with anything else available today, not the performance of the CPUs from 5 years ago. Like I have said, the only disadvantage caused by their old fabrication technology is in the worse energy efficiency, i.e. high power consumption, and it does not affect the achievable absolute performance.
While that CPU is intended mainly for HPC, i.e. FP64 computations, it can also perform fast computations with FP16, BF16 and INT8, and it has very fast interconnections between computers (low latency 1.6 Tb/s communication links) so it could do very well AI training or inference.
So there is no doubt that the US policy of export restrictions has been a complete failure, because without it the Chinese would not have had today such CPUs, or such GPUs like the Huawei GPUs.