AI training cluster
A large group of GPUs joined by a fast network so that they can work together on training one AI model.
Also: training cluster, supercluster
Training a large AI model means running the same calculation on tens of thousands of GPUs at once and constantly sharing results between them. A training cluster links them with a fast, low-latency network such as InfiniBand or specialised Ethernet. xAI's Colossus in Memphis began in 2024 with about 100,000 Nvidia H100 GPUs. Training clusters draw huge, steady power, and their load can swing sharply, which grid operators now study.