Hopper (H100)
Nvidia's datacenter GPU generation from 2022. H100 (80 GB HBM3) drove the entire AI boom 2023-2024. TSMC 4N process. 814 mm² die. Transformer Engine: FP8 + sparsity delivers 4x throughput over A100.
H200 (2024): same chip, more HBM (141 GB HBM3e), 1.4x inference bandwidth. NVLink-Switch + NVSwitch for 256 GPU cluster. SXM5 form factor (~$30-40k per card) standard. Allocation problem: 2023-2024 LLM startup CEOs hung on Jensen's sales team to get H100 supply. CSPs (cloud service providers) bought billions worth. Successor: Blackwell (B200, 2024-2025). Hopper name from Grace Hopper.