Nvidia CEO Huang Jen-Hsun said after the keynote speech at the Taipei Computer Show will strive to achieve supply chain diversification, and said the future cooperation with Intel to carry artificial intelligence chips open attitude. So some people suspect that Intel will OEM such top chips as H100 for Nvidia.
Huang Renxun now said that the H100 is produced exclusively by TSMC foundry and will not consider adding a second foundry, the main reason is the entire design foundry process, “do 1 time is very difficult, separate 2 foundries to do 2 times more difficult”.
NVIDIA and TSMC have begun 3/2nm cooperation planning, Huang Jen-Hsun also confirmed for the first time that the next generation of AI chips under the single TSMC, at least in the AI GPU series, for several years will continue to be TSMC “exclusive large single”.
It is worth mentioning that Nvidia A100 / A800 and H100 are ordered TSMC, RTX 40 series has also returned to TSMC 4nm (RTX 30 series is using Samsung 8nm process), the current cooperation between Nvidia and Samsung to memory-based.
NVIDIA’s H100 GPU uses TSMC’s 4nm process, while the A100 is a 7nm process, and NVIDIA’s top-of-the-line H100 Tensor Core GPU uses the Hopper architecture, based on TSMC’s 4nm process, which is equipped with a fourth-generation Tensor Core and Transformer engine, a single card equipped with 80GB of video memory, compared to The A100 GPU is two times more efficient per kilowatt.
As of now, TSMC currently sees no real competition in the sub-7nm process that supports AI chip manufacturing. Intel has returned to the foundry space but has yet to make any meaningful progress.