The HGX-H20 is a GPU customized by Nvidia to ensure compliance with US Department of Commerce export rules, while still providing powerful artificial intelligence (AI) performance.
This GPU model has a processing speed of 296 INT8 TOPS/FP8 TFLOPS, comes with 96 GB of HBM3 high-bandwidth memory, and a speed of 4.0 TB/s bandwidth. Although the parameters on paper seem less powerful, the HGX-H20 is still able to compete with its direct rival, the Huawei-made AI Ascend 920, thanks to its better memory performance.
Nvidia’s H20 GPUs could face an export ban when Washington reviews semiconductor export policy in October, QZ reported. The ban could take many forms, from banning sales of a specific product to requiring reduced computing power or limiting maximum memory capacity.
Most Chinese AI companies have been developing ecosystems using Nvidia’s CUDA platform, so switching to other products, such as Huawei Ascend, would be costly and time-consuming.
The HGX H20 GPU is fully CUDA compatible, making it an easy choice for many companies despite being significantly slower than the H100 version.
Despite existing export restrictions, Chinese companies still get their hands on cutting-edge Nvidia GPUs through intermediaries, as well as by renting servers from Google and Microsoft.
(According to Tom's Hardware)
Source: https://vietnamnet.vn/nvidia-mat-toi-12-ty-usd-neu-my-cam-ban-chip-ai-sang-trung-quoc-2304389.html
Comment (0)