NVIDIA's New AI Accelerator for China Will Ditch Hopper Architecture and HBM Memory
Following the recent export ban HGX H20 accelerators to China, company NVIDIA is preparing a new solution for the Chinese market, but it will not be based on the Hopper architectureThis was announced by the company's CEO Jensen Huang on Taiwanese television, and additional details were revealed by sources of Nikkei Asia.
According to insiders, The new accelerator will switch from HBM to GDDR7 memory, to circumvent American export restrictions. Recall that the US banned the supply of not only the H100 and H200, but also the adapted H800 and H20. The latter was the only flagship GPU NVIDIA, available in China before the new ban in April. The company has already lost $5,5 billion in potential revenue. due to these restrictions.
In response to a question about the successor to the H20, Huang has stated outright that it is "not Hopper" because it is no longer possible to modify the architecture.The US justified the ban by saying that the H20 has too high a memory bandwidth and inter-chip connections, making it suitable for military supercomputers.
The move to GDDR7 represents a radical change in approach, since the GH100 (Hopper core) was originally designed to work only with HBM. A potential alternative could be the Blackwell architecture in the consumer version of GB20X (the RTX 50 series is based on it), where GDDR7 is already in use, but it lacks support for NVLink, a key component for scaling AI systems.
Against the backdrop of this uncertain transition period, Interest in Chinese solutions such as Huawei's Ascend 920 is growing, although they are inferior to NVIDIA in performance. In the coming months, NVIDIA will have to develop a new architecture that meets US requirements in order to maintain its position in China - otherwise, the market may be taken over by local manufacturers.
