NVIDIA Introduces the Hopper GPU Architecture

NVIDIA introduced the Hopper data-center GPU architecture at GTC in March 2022. The flagship H100 was designed primarily for AI and high-performance computing rather than consumer graphics.

H100 added fourth-generation Tensor Cores and a Transformer Engine capable of choosing lower numerical precision such as FP8 for parts of neural-network computation while preserving higher precision where necessary. Lower precision reduces memory traffic and allows more arithmetic operations per second.

Hopper also included faster interconnect support and partitioning features for sharing GPUs securely among workloads. The architecture became one of the central pieces of the generative-AI boom because training and serving large transformer models required enormous matrix-computation throughput and memory bandwidth.