NVIDIA announces the Hopper architecture and the H100
The H100 became the part frontier training runs were built around for the next two years, and the unit of account in which compute deals were later measured.
- Compute & infrastructure
- Major
NVIDIA announced the Hopper architecture and its first implementation, the H100 GPU, at its GTC conference. The company said the chip, built on TSMC’s 4N process with 80 billion transistors, delivered an order-of-magnitude performance leap over its predecessor, with roughly 3TB/s of memory bandwidth from HBM3 and support for the newer PCIe Gen5 standard. Availability was announced for the third quarter of 2022.
The specific addition aimed at language models was the “Transformer Engine,” a mixed-precision capability NVIDIA said could speed up transformer-based neural networks by up to six times over the prior Ampere generation without loss of accuracy. Named for the architecture underlying nearly every large language model then in production, it signalled that NVIDIA was designing its flagship hardware around the specific computational pattern that scaling laws had made valuable, rather than treating language models as one workload among many.
The H100 became the accelerator that frontier labs built their next generation of training runs around, and its price and availability shaped how those runs were planned: waiting lists, allocation disputes between labs and cloud providers, and export-control debates over which countries could buy it all followed from decisions made around this one part. Compute deals over the following two years were routinely described in “H100-equivalents,” a unit of account that persisted even after NVIDIA’s next architecture, Blackwell, began shipping. NVIDIA’s stock and revenue, and the broader narrative that AI progress required not just algorithmic ideas but a specific, scarce, and expensive category of hardware, trace back in large part to how central this chip became to the industry’s plans.