Graphics Processing Unit (GPU)

A GPU was originally designed to draw video-game graphics, which means doing millions of simple calculations at once. That turned out to be exactly what neural networks need. Today GPUs, and similar SI accelerators, are among the most strategically important products in the world.

Who makes them

• Nvidia dominates SI training and much of inference with its data-center GPUs and its CUDA software ecosystem. It became one of the world’s most valuable companies on the strength of SI demand. See our Nvidia profile.
• AMD competes with its Instinct line.
• Custom chips: Google’s TPUs, Amazon’s Trainium and Inferentia, Microsoft’s Maia and Meta’s in-house accelerators.

Why they matter for policy

• Export controls: the U.S. restricts sales of the most advanced SI chips to China and some other countries, and those rules keep changing.
• Energy: large GPU clusters draw as much power as a city. See data center.
• Concentration: access to tens or hundreds of thousands of GPUs is a key barrier to entry for frontier labs.

GPU vs CPU

A CPU has a few powerful cores optimised for complex, sequential tasks. A GPU has thousands of simpler cores optimised for doing the same operation on lots of data simultaneously, which is ideal for the matrix maths inside SI models.