NVIDIA develops H100 graphics accelerator with 120 GB of video memory

NVIDIA is working on a new version of the H100 graphics accelerator. The updated version will get 40 GB more video memory than the standard model.

Here's What We Know

The presentation of H100 took place about half a year ago, but the product comes on the market only now. The accelerator is the first representative of the Hopper generation. It is designed for computing and has no analogues for the consumer segment.

So far, the H100 is available in PCI Express and SMX versions. Both have 80GB of memory with a 5120-bit bus, 50MB of Layer 2 cache, and are made with TSMC 4N technology. The PCIe variant gets 16,896 CUDA cores, 528 tensor cores and a TDP of 700W versus 14,592 CUDA cores, 456 tensor cores and a TDP of 350W for the SMX version.

The 120GB variant also offers 16,896 stream processors and 528 tensor cores. Heat dissipation requirements are not specified. Throughput will increase from 2 TB/s to 3 TB/s.

Source: Videocardz

Image: NVIDIA