what is a gpu
A GPU, or graphics processing unit, is a chip built around a completely different design philosophy than a CPU: instead of a few powerful cores optimized to handle one complex task after another very quickly, a GPU packs thousands of simpler cores designed to do the same simple operation on massive amounts of data all at once — a structure originally built for rendering graphics that turned out to be exactly what modern AI also needs.
```flowchart
{
"rows": [
{ "type": "nodes", "items": [
{ "title": "A task made of millions of small, similar calculations", "sub": "e.g. shading pixels, or multiplying matrices", "tone": "neutral" }
]},
{ "type": "arrow" },
{ "type": "group", "title": "GPU splits it across thousands of simple cores", "items": [
{ "title": "Each core handles a small piece", "sub": "all running in parallel", "tone": "green" }
]},
{ "type": "arrow" },
{ "type": "nodes", "items": [
{ "title": "Massive parallel throughput", "sub": "far faster than a CPU could do the same task", "tone": "orange" }
]}
]
}
```
**A GPU trades per-core sophistication for sheer numbers.** A CPU core is built to handle complex, varied, unpredictable instructions quickly and correctly, one after another, with a lot of supporting circuitry dedicated to that flexibility; a GPU core is much simpler and less flexible, but a single GPU chip contains thousands of them, all executing the same instruction on different pieces of data simultaneously. For tasks that naturally break into many identical, independent pieces of work, that tradeoff wins decisively over a CPU's smaller number of more capable cores.
```svg
```
| Property | CPU | GPU |
|---|---|---|
| Core design | Few, complex, highly flexible | Thousands, simple, less flexible individually |
| Best suited for | Varied, sequential, decision-heavy tasks | Massively repetitive, parallel tasks |
| Original purpose | General-purpose computing | Rendering pixels for graphics |
| Modern expanded use | Still general computing | Graphics, AI training/inference, scientific simulation |
**A GPU became central to AI almost by accident, because training a neural network is, at its core, an enormous pile of matrix multiplication.** The same parallel architecture built to shade millions of pixels independently and simultaneously turned out to be extremely well suited to performing the same repeated multiply-and-add math across huge neural network layers — which is why GPUs, originally a gaming and graphics technology, ended up powering the vast majority of AI model training happening today.
**A GPU cannot simply replace a CPU, because the two are optimized for opposite kinds of workloads.** A computer still needs a CPU to run its operating system, handle unpredictable branching logic, and manage tasks that don't naturally split into thousands of independent parallel pieces — a GPU excels specifically at the subset of work that does split that way, which is why virtually every GPU-equipped system still relies on a CPU alongside it rather than instead of it.
**Not every workload benefits from a GPU, which is why choosing between CPU and GPU is a real design decision rather than "GPU is always faster."** A task with heavy branching logic, frequent dependencies between steps, or work that doesn't split neatly into many independent pieces often runs poorly on a GPU's simple cores, which are optimized for uniform, parallel work rather than complex decision-making — this is exactly why a well-designed system routes different parts of a workload to whichever chip actually suits it, rather than assuming more parallel cores automatically means better performance.
Read the GPU through a many-simple-workers lens rather than a raw-power lens: its advantage isn't that each of its cores is smarter than a CPU core, it's that thousands of simpler cores working on the same kind of problem simultaneously can outrun a handful of more capable cores working through that same problem one piece at a time.