Gaming PC

Nvidia Gimps H100 Hopper GPU to Sell as H800 to China

Nvidia’s H100 (Hopper) GPU is missing from our list of best graphics cards. However, the H100’s strength lies in artificial intelligence (AI), making it the coveted GPU in the AI ​​industry. And now that everyone is on the AI ​​bandwagon, Nvidia’s H100 is getting even more popular.

Nvidia claims the H100 offers up to 9x faster AI training performance and up to 30x faster inference performance than its predecessor A100 (Ampere). With performance like this, it’s easy to see why everyone would want to own the H100. In addition, Reuters (opens in new tab) reported that Nvidia modified the H100 to comply with export regulations, allowing chip makers to sell the modified H100 to China as the H800.

Last year, US authorities implemented several regulations to prevent Nvidia from selling its A100 and H100 GPUs to Chinese customers. This rule restricted GPU exports with chip-to-chip data transfer rates below 600 GBps. Transfer speed is important in the world of AI, where systems need to move huge amounts of data to train AI models like ChatGPT. Slower data transfer speeds between chips means longer data transfer times, longer training times, and significantly lower performance.

With the A100, Nvidia cut the GPU’s 600 GBps interconnect to 400 GBps and rebranded it as the A800 to commercialize it in the Chinese market. Nvidia has taken the same approach with the H100.

Nvidia has cut the H800’s chip-to-chip data transfer speeds to about half that of the H100, according to Reuters’ Chinese chip industry sources. This leaves the H800 interconnect limited to 300 GBps. That’s a significant drop in performance compared to his A100 and A800, which had a 33% drop in data transfer speed between chips. However, the H100 is significantly faster than his A100, which may be why Nvidia has imposed tighter chip-to-chip data transfer rate limits on the former.

Reuters asked an Nvidia spokesperson about the differences between the H800 and H100. However, an Nvidia representative only said, “Our company’s 800 series products are fully compliant with export control regulations.”

Nvidia already owns three of the most prominent Chinese tech companies using the H800: Alibaba Group Holdings, Baidu Inc and Tencent Holdings. China banned his ChatGPT. So the tech giants are competing with each other to create domestic ChatGPT-like models for the Chinese market and also his H800 with half the chip-to-chip transfer speed his H100 with full fat Definitely slower than , but still not slower. slowWith companies potentially using thousands of Hopper GPUs, they ultimately have to wonder if it means using more H800s and fewer H100s to accomplish the same work. yeah.

Related Articles

Back to top button