Blackwell is an NVIDIA GPU architecture built for training and running very large AI models, including reasoning and mixture-of-experts models.
In plain words
A Blackwell GPU joins two large chip dies with a very fast link so they behave as one GPU, with 208 billion transistors in total. Its second-generation Transformer Engine and new low-precision number formats speed up the maths of language models. At launch, NVIDIA named many large tech companies expected to adopt it.
A quick example
Imagine an AI company planning to serve a large mixture-of-experts model. It chooses Blackwell systems because their Transformer Engine and smaller number formats are designed to speed up exactly that kind of model.
Where you'll meet it
- Large AI data centres, which NVIDIA calls AI factories.
- Training and serving frontier and mixture-of-experts models.
Related
Sources
- DocsNVIDIA Blackwell Architecture, NVIDIA · read 29 Sept 2026
- DocsNVIDIA Blackwell Platform Arrives to Power a New Era of Computing, NVIDIA Newsroom · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.