GPU

Overview

The GPU began life rendering graphics, a task that requires the same simple calculation to be applied to millions of pixels at once. That demand shaped a processor with a very different character from a CPU: where a CPU has a handful of powerful cores optimized for sequential logic, a GPU has thousands of simpler cores optimized for doing the same operation across enormous batches of data simultaneously. That architecture, it turned out, is almost perfectly suited to AI.

How it works

A GPU achieves its speed through massive parallelism. Thousands of cores execute together, each working on a small piece of a much larger calculation, fed by high-bandwidth memory designed to keep them all supplied with data. The mathematics of neural networks, dense matrix multiplications repeated at scale, decomposes naturally into exactly this kind of parallel work, which is why a GPU can outpace a CPU on these tasks by a hundredfold or more.

Why it matters

GPUs are the reason modern AI is practical at all. Training and serving models with billions of parameters would be impossibly slow and expensive on CPUs alone, and every advance in model scale has tracked the growing capability of GPU hardware. The specific generation matters too, since memory bandwidth, interconnect speed, and supported precision formats set a hard ceiling on what a given chip can run.