Infrastructure
NVIDIA H100
NVIDIA's 2022 Hopper datacentre GPU — 80GB HBM3, ~3× A100 performance.
Definition
The H100 added FP8 support, the Transformer Engine, and 80GB HBM3 memory, making it the dominant frontier-training GPU of 2023–2024. The SXM5 variant peaks near 4 PFLOPS in FP8 and powers most H100-based clusters.
Common use cases
- LLM training
- High-end inference
- Diffusion training