A CPU, as covered earlier on this trail, is a handful of extremely fast, extremely clever workers. A GPU is the opposite bet: thousands of workers, each individually slow and simple, all working at once. Neither design is "better" — they're built for different shapes of problem, and understanding which shape is which explains the last decade of computing.
Why graphics needed a different kind of worker
Think about what drawing a screen involves. A modern display has millions of pixels, and for a game running at 60 frames per second, every one of them needs its colour recalculated 60 times a second. Here's the key property: each pixel's calculation is independent. Pixel three million doesn't need to wait for pixel one. The work is embarrassingly parallel — a phrase engineers genuinely use.
Handing that to a CPU is like asking four brilliant professors to fill in a million-cell spreadsheet by hand. They're overqualified and outnumbered. A GPU is a lecture hall of a thousand students each doing simple arithmetic simultaneously — vastly more effective for this specific shape of task.
The accidental revolution
Around the mid-2000s, researchers noticed something: lots of important problems have that same shape. Simulating weather, modelling proteins, and — most consequentially — training neural networks are all built on enormous grids of numbers being multiplied together, where each small calculation is independent. The hardware built to draw video game explosions turned out to be almost perfectly shaped for machine learning. That accident is a large part of why AI progress accelerated when it did, and why a graphics-card company became one of the most valuable businesses on Earth.
Rule of thumb for which chip suits which job: if the task is one long chain of decisions where each step depends on the last, that's CPU territory. If it's the same simple operation stamped across millions of independent data points, that's GPU territory.
Where you'll meet GPUs in real IT work
Even outside gaming and AI teams, GPUs show up in ordinary infrastructure: CAD and design workstations, video editing rigs, virtual desktop infrastructure (where one server's GPU is sliced up between many users), and increasingly in cloud bills — GPU-enabled virtual machines cost several times their ordinary siblings, so knowing when a workload actually needs one is a genuinely valuable skill. Plenty of money gets wasted renting lecture halls for jobs that needed one professor.