AI accelerator
A chip built to run neural networks fast and efficiently, such as Google's TPU, Nvidia's data centre GPUs or Amazon's Trainium.
Also: AI chip, ML accelerator, neural network accelerator
Neural networks spend most of their time multiplying large matrices of numbers, often at low precision. AI accelerators devote most of their silicon to that, using matrix units, wide memory interfaces and fast links between chips, and drop features a general processor needs. They split into training chips, which build models, and inference chips, which run them. Examples include Nvidia's H100 and B200, AMD's Instinct MI300X, Google's TPU, Amazon's Trainium and Inferentia, Intel's Gaudi, and wafer-scale chips from Cerebras. The NPUs in phones and laptops are small accelerators.