TECH TALK
Nvidia Unveils Hopper H100 Datacenter GPU
Jarred Walton
© NVIDIA
WE’VE KNOWN NVIDIA has been working on next-generation GPUs for some time, but it officially unveiled its Hopper architecture and the H100 datacenter GPU at GTC 2022. Hopper H100 will replace the Ampere A100, which replaced the Volta V100. Designed for supercomputers, with a focus on AI capabilities, H100 provides numerous updates to the current A100.
Built with 80 billion transistors on a custom TSMC 4N process (4nm Nvidia, not to be confused with the generic N4 that TSMC also offers), Hopper ups the ante in many ways. It supports a PCIe 5.0 interface, the first GPU to do so, as well as 4th generation NVLink, which can deliver up to 900GB/s of bandwidth, while PCIe 5.0 tops out at 128GB/s, 1.5x and double the bandwidth of the A100, respectively. H100 will have 80GB of HBM3 memory with 3TB/s of bandwidth, x1.5 faster than the A100’s HBM2E.