How Intel, NVIDIA and ASML continue to cooperate with China without fear of sanctions. A tricky way that suits everyone
Due to the trade conflict between the United States and China and China, the development of the electronics industry has slowed down. Thus, it is forbidden to import into the country modern lithographic machines, and indeed practically modern technologies, in the development of which American companies took part. Despite all this, the PRC is trying to develop its own electronics development and production industry – Chinese companies create their own lithographic machines, modern chips, etc.
But there is another opportunity for the Celestial Empire to receive a “ban”. And this method was developed just by Western companies, which, apparently, do not benefit from losing such a large market as China. The most interesting thing is under the cut.
Intel will supply “stripped down” Gaudi GPUs
The other day it became known that Intel will start selling Habana Gaudi 2 accelerators to China. They are designed to work with deep learning and inference tasks (ensuring the operation of a pre-trained neural network on a final device). At the same time, the accelerator itself is a system that cannot be delivered to China in its current form, due to restrictions imposed by the United States on this country.
In order to get around the restrictions and make a profit, Intel will produce an export version of Gaudi. If earlier the adjective “export” meant an improved version of the exported product, be it a car, beer or something else. In the case of Intel, the opposite is true – Gaudi will be somewhat cut in functionality. In this form, GPUs will be sold in China without fear of problems with US regulatory authorities.
Unfortunately, what exactly will be cut / reduced, the company does not disclose. Probably, this will become known after the start of GPU sales in China. As for the system itself, it is a competitor to the Nvidia A100 – this is also an accelerator that is used in many industries related to AI, including training large language models, including GPT. The performance of the Intel GPU is about equal to the performance of the competitor’s system, at least when using the ResNet-50 and BERT benchmarks.
And NVIDIA too
This is a 100% American company. Accordingly, it must comply with export regulations set by US regulators. In general, it does just that, but there are some tricks to keep delivering modern equipment to buyers from China. We are talking about the H100 graphics accelerator, which the company is modifying for Chinese consumers. A customized version of the system is sold under a different name – not H100, but H800.
As far as we know, the version of the system that is being shipped to China is artificially “slowed down”. Those. the accelerator has reduced throughput characteristics. So, if the H100 has 300 Gb / s, then the Chinese version has only 150 Gb / s.
It is clear that such powerful accelerators are bought not by individuals or small companies, but by the most powerful corporations from China, including Alibaba Group, Baidu and Tencent. In general, all those who have data centers and a deployed cloud infrastructure.
As for the characteristics of the “original”, i.e. model H100, it includes 80 billion transistors. It is also the world’s first PCIe Gen5 accelerator with HBM3 high-speed RAM. The memory bandwidth is also impressive – about 3 TB / s.
Thanks to the new high-performance NVLink bus, 256 H100 accelerators can be interconnected at once. Of course, a system with such characteristics could not pass by the attention of Chinese corporations. And Nvidia doesn’t want to lose customers, because the Chinese market is really huge, who wants to miss out on hundreds of millions of dollars of revenue?
Dutch ASML also does not want to lose profits
ASML is looking for an opportunity not to refuse money from Chinese customers. According to DigiTimes, the Dutch company is preparing to release a special installation that can be supplied to China without a license. The lithographic machine will allow the production of chips using 28-nm and older technologies.
We are talking about the installation Twinscan NXT:1980Di, which was released in 2016. But it makes it possible to produce chips using the 7nm process technology, which is what the Taiwanese TSMC does with this equipment.
The stripped-down version will not allow the Chinese to produce chips using a process technology that is more modern than 28nm. But this is not bad either, since at least part of the chips the Chinese have problems removed … Taking into account that the lion’s share of SMIC’s income comes from the supply of 28 nm and less modern chips, it is likely that Chinese companies will still be interested in acquiring such tools .
Recent export regulations require U.S. firms and individuals to obtain export licenses for tools and technologies capable of producing chips at 14/16nm and below, 3D NAND with 128 layers or more, and DRAM memory chips at 18nm-pitch or less. The same rules apply to non-US companies that export components from the US, as is the case with ASML and Twinscan NXT:1980Di.
And what are the criteria for determining “prohibition”?
The main requirements for exported products were developed by the US Department of Commerce in October 2022. At that time, the rules for the supply of semiconductors to the PRC, the development of which uses American technologies, were tightened.
The main requirements are restrictions on the performance of semiconductor products supplied to China. Bidirectional transfer rate (BTR) was chosen as the metric. Export to China of semiconductor products whose performance parameters exceed the norm set by US regulators is prohibited. The only way to start exporting is to get a special license.
As for the PRC, the country is currently buying equipment necessary for the development of the electronics industry, injecting big, even very big money into the industry (we are talking about trillions of US dollars), and also strengthening control over the companies that receive these funds.
You may also be interested in these texts:
→ Snort and Suricata – an easy way to use IDPS: from installation on the server to proper configuration
→ “20% of the equipment will not arrive on time, and 10% will never arrive.” Parallel import as seen by the infrastructure provider
→ How to effectively manage the server fleet? Optimizing work with API
