Also known as: Graphics Processing Unit
In plain English
GPUs were designed for video games, but they turned out to be perfect for AI because they can do thousands of small maths calculations at the same time.
In practice
GPU supply and cost shape the AI market, from cloud prices to which companies can train frontier models. For your own projects, small models can often run on ordinary CPUs; GPUs matter when you need speed or are running larger models.
Under the hood
GPUs execute massively parallel matrix operations, the core workload of neural networks. Other AI accelerators include Google's TPUs and various custom chips. Key specs are memory capacity and bandwidth, which limit model size and speed, plus the interconnect for multi-GPU training.
Example
"We rent cloud GPUs by the hour to test the open model."