TSMC’s Sub-5nm Process Utilization Rate Dropping Recently
The author is an analyst of NH Investment & Securities. He can be reached at hwdoh@nhqv.com. -- Ed.
The US government announced on Aug. 31 (local time) that it will place restrictions on exports of high-performance GPUs from Nvidia and AMD to China and Russia. This is a negative development for the earnings of memory chipmakers such as Nvidia, AMD, and TSMC. However, it could spell good opportunities for startups related to machine learning computation.
Exports of high-performance GPUs to China to be restricted
Nvidia and AMD are to be subject to recently-announced US government restrictions on exports of high-performance GPUs to China and Russia. Of note, the main products which currently handle computations such as machine learning at data centers are Nvidia’s H100 and A100 and AMD’s M100 and M200. These products are all equipped with Floating Point (FP) 64 calculation function. Given the US restrictions, Nvidia and AMD will now need new licenses from the US government to export these products. Nvidia is saying that the new restrictions could cut its revenue by US$400mn this quarter.
Once the regulations take effect, Chinese Internet companies such as Tencent, Baidu, and Alibaba will be able to use only V100 of 12nm processes for high-performance computing. If a product with this level of performance is used, computational performance will be significantly lowered, and costs at these companies will increase significantly as they will not be able to cover the effects of a recent surge in electricity costs.
TSMC’s sub-5nm node process utilization rate dropping recently
In-house developed machine learning computational processors such as Hanguang, Kunlun, and Xuantie in China are still considered to be technically far inferior in performance to Nvidia and AMD products. As alternatives, Chinese companies may use large amounts of Intel’s high-performance CPUs instead of GPUs or use machine learning processors from startups such as Cerebras.
Meanwhile, it is known that the utilization rate of TSMC’s sub-5nm node process, which produces processors for Nvidia and AMD, has recently started to decline. In response, TSMC has decided to shut down some EUV equipment (which consumes several times more power than DUV) from the end of the year in order to reduce power costs. AMD and Nvidia account for about 10% of TSMC’s sales. Intensifying US sanctions on China are expected to have a further negative impact going forward on TSMC’s utilization rate, which has begun to drop as of late. The new restrictions are to dampen earnings at domestic memory makers as well.