The specialised chip that nearly all AI training and running depends on. Originally built for rendering video game graphics, it turned out to be ideally suited to AI because both tasks involve doing enormous numbers of simple calculations at the same time. Access to these chips is the main practical bottleneck in the industry, and the reason for both their cost and their geopolitical significance.
Technical view
GPUs provide massive parallelism suited to the matrix multiplications that dominate neural network computation. Tensor processing units (‘TPUs’) are purpose-built alternatives. In practice, high-bandwidth memory capacity and interconnect speed between chips are often more limiting than raw arithmetic throughput, particularly at inference, which is why memory-reduction techniques such as quantisation translate so directly into speed.
Related
Part of AI Terminology Index, section: Hardware, Compute and Training Mechanics.