Electronics
GPU — Graphics Processing Unit
CPU's page describes one expert following a recipe, one precise step at a time, extremely fast. A GPU is the opposite shape entirely: not one expert, but thousands of simple workers, each doing the same small task simultaneously on a different piece of data.
- What it is
- A processor built for massive parallelism — thousands of simple cores executing the same operation on different data at once, rather than a few powerful cores doing varied work in sequence.
- Mental model
- A CPU is one expert chef preparing one complex dish, very fast. A GPU is a thousand line cooks, each repeating one simple task on their own ingredient, all at the same time.
- Key numbers
- Thousands of cores (vs. a CPU's handful to dozens); its own dedicated VRAM, measured in GB.
- Reach for it when
- The workload is the same operation, repeated independently across huge amounts of data — pixels, or matrix math.
Massive parallelism
Rendering a frame means computing a color for every one of several million pixels — and critically, each pixel's computation doesn't depend on any other pixel's. That's exactly the shape of problem a GPU is built for: the same small program, run independently across an enormous number of data points at once, rather than one long sequential chain of dependent steps.
CPU vs. GPU, side by side
| CPU | GPU | |
|---|---|---|
| Core count | A handful to a few dozen | Thousands |
| Per-core power | High — complex branching, varied instructions | Low — simple, repetitive operations |
| Best at | Sequential, branch-heavy, varied work | The same operation repeated across huge independent data sets |
| Own dedicated memory | Shares system RAM | Dedicated VRAM, extremely high bandwidth |
Why AI workloads live on GPUs
Neural network computation is, overwhelmingly, matrix multiplication — the exact same operation repeated across enormous, independent grids of numbers. That's precisely the shape of problem a GPU's thousands of simple cores are built for, which is why GPUs (not CPUs) became the default hardware for training and running the AI systems covered in the AI Systems domain — the hardware shape happened to match the math almost perfectly.
VRAM
A GPU keeps its own dedicated high-bandwidth memory, separate from system RAM, because its thousands of cores need to pull huge amounts of data — textures, model weights — at a rate ordinary system RAM isn't built to sustain. VRAM capacity is frequently the real limit on what a GPU can handle at once, sometimes before its raw core count ever becomes the bottleneck.
Real-world applications
- Gaming and graphics rendering — the GPU's original purpose, and still a direct fit for its parallel-pixel shape.
- AI/ML training and inference — the dominant modern use of high-end GPU hardware, driven directly by the matrix-multiplication match above.
- Video editing and encoding — another naturally parallel workload (transforming every pixel of every frame independently).
- Scientific/parallel computing — simulations built from huge independent grids of calculations (weather models, physics simulations) map onto the same core shape.
Gotchas
- More VRAM or more cores doesn't help a workload that isn't actually parallel — matching the problem's real shape to the right processor matters more than either spec sheet number alone.
- A GPU-bound task gets no benefit from a faster CPU, and vice versa — diagnosing which processor is actually the bottleneck is the first real step before "upgrading" anything.
- GPUs draw and dissipate significant power under sustained load — see Cooling and PSU for why a high-end GPU changes both of those requirements for an entire system, not just itself.