Chronologie
Jalons de recherche, générations matérielles, sorties de modèles, déploiements industriels et décisions politiques — avec des liens vers les fiches concernées.
1990–2011 · The long ramp
2012–2016 · Deep learning works
A 15-month crash ASIC project puts custom inference silicon in production — revealed to the world in 2016.
2017–2019 · Transformers
NVIDIA builds matrix math into silicon — the first AI-specific GPU feature.
2020–2022 · Scaling
Ampere becomes the workhorse of the first scaling wave.
Hopper's Transformer Engine and FP8 are built for the models to come.
2023–present · The build-out
Two-die B200 and the GB200 NVL72 make the rack the unit of computing — and liquid cooling mandatory.
1.742 exaFLOPS on AMD MI300A — the public sector's biggest disclosed machine.
AWS commits its own silicon to Anthropic at hundred-thousand-chip scale.
Google's inference flagship: 9,216-chip pods for the 'age of inference'.
AMD's CDNA4 reaches Blackwell class; OpenAI appears on stage.