Nvidia: H100 GPU uses TSMC 4nm process

At the NVIDIA GTC conference tonight, NVIDIA will officially release the "Hopper" series of GPUs. Before the press conference, foreign media VideoCardz exposed the specifications of the H100 GPU.

According to reports, it was previously rumored that Nvidia's AI accelerator based on the Hopper architecture's H100 GPU will use TSMC's 5nm process, but the fact is that it uses a 4nm process.

The Nvidia H100 GPU has 3 TB/s of HBM3 memory, 1.5 times that of the older A100. In terms of performance, Nvidia claims that its FP64, TF32, FP16 have 3 times the computing power of the A100, and the FP8 has 6 times the computing power of the A100.

In addition, this accelerator will be divided into two designs, PCIE Gen5 and SXM. The TDP of the latter is 700W, which is 300W higher than the old A100.

In addition, it is reported that Nvidia will also launch two ARM-based Grace CPU platforms today. One uses a Grace CPU combined with a Hopper GPU design with 600GB of memory, while the other uses two Grace CPUs with a total of 144 cores. This series of CPUs will be equipped with LPDDR5X memory.

Post a Comment