Although the NVIDIA H20 accelerator was exported in July, the H20 was almost completely blocked due to the doubts made by the Chinese supervisory authority on the grounds of data safety. Currently, China's AI industry is turning to renovating and using second-hand NVIDIA GPUs to meet the sudden replacement demand.
Market news has been released, as the official has completely banned H20, Chinese companies dismantled and reorganized the A100 and H100 cards into a "low-cost, high-efficiency" customized recommendation system, resulting in a surge in demand for two old cards in the industry.
AI Inference requires less computing power than AI training, and can still operate efficiently even on reconfigured hardware. This is why the A100 launched in 2020 is still valuable.
A100 uses NVIDIA Ampere architecture, equipped with up to 80GB of HBM2e memory and 2 TBps bandwidth. Although it does not achieve the ultimate performance of the Hopper architecture, it is still effective for recommendation tasks due to large-capacity memory and mature CUDA software.
As for the H100 equipped with HBM3 memory, AI training performance is as high as six times higher than that of the A100. In contrast, although China's reduction board H20 is optimized for recommendations, it is over-cut, and its AI performance is only about 1/3 to 1/7 of the H100, which is 30 times slower in FP64 supercomputing. Therefore, the second-hand A100 may all be more attractive than the brand-new H20.
However, NVIDIA is currently in conflict because the government has imposed export restrictions on H20, causing the company to reduce its inventory of US$5.5 billion; on the other hand, NVIDIA GPUs are still the core of promoting China's AI craze, and today's gray market may infringe on profits and slow down the promotion of new architectures.
For China, the gray market is also a hilarious one. Every time the data center uses a second-hand H100, it means that there is another opportunity to use Ascend accelerator, which may delay the progress of local investment. China's AI industry is waiting for the lifting of political and technical obstacles, and it can extend the life of the old NVIDIA hardware as much as possible.
Chinese firms turn to used Nvidia GPUs following Beijing H20 crackdown — industry reports ‘surge’ in A100 and H100 chips being stripped and repurposed Extended reading: Financial Times: China has greatly increased AI chip production and reduced NVIDIA Face Trump? Grofangde: The funds of the Chip Act are not affected, and the framework does not involve stocks