Two kinds of chip
A core is the part of a chip that runs instructions. A CPU has a few big, clever cores. Each one runs its own program fast and copes with any job. The Intel Core i9-13900K has 24 cores.
A GPU (graphics processing unit) has thousands of small, simple cores. Each one does only simple arithmetic, but there are very many. The NVIDIA GeForce RTX 4090 has 16,384 of them.
Same job, two ways
A picture is a grid of pixels, tiny dots of colour. Here each pixel is a grey number from 0 (black) to 255 (white). A full HD screen has 2,073,600 pixels.
Job: make every pixel brighter by adding 40. The CPU's 4 cores take 4 pixels each round, so 32 pixels need 8 rounds. For two million pixels the queue gets very long.
The GPU sends one instruction, px + 40, to every lane at once. Each lane holds its own pixel, so all 32 change together. This is SIMD: single instruction, multiple data.
Inside the GPU
Each pixel's work is a thread. NVIDIA groups 32 threads into a warp; AMD calls its group a wavefront, of 32 or 64. One scheduler hands the whole warp the same instruction, and all lanes do it in step.
Thousands of cores need data quickly. The RTX 4090's memory moves 1,008 GB per second; the Core i9-13900K's memory link manages 89.6 GB per second. GB means gigabyte: a billion bytes.
GPUs are poor at branching. With if and else, light pixels take one path and dark ones the other. A warp runs one instruction at a time, so half the lanes wait, then the other half.
Where each wins
Machine learning is mostly matrix maths: grids of numbers multiplied together. Each answer number is its own sum of products, independent of the others, so each lane can work out one. That is why AI runs on GPUs.
Use the CPU for step-by-step work full of decisions: the operating system, game rules, most apps. Use the GPU for the same small sum on huge piles of data: pixels and AI. Your computer uses both.
The numbers
;; step 1/10 · cpu ;; about: CPU Intel Core i9-13900K (2022, Intel ARK) 24 cores = 8 performance + 16 efficient 32 threads The figure draws 4 big cores. Each core: its own instruction stream, branch prediction, big caches.
Core counts: Intel Core i9-13900K, 24 cores (8 performance + 16 efficient), max memory bandwidth 89.6 GB/s (Intel ARK); NVIDIA GeForce RTX 4090, 128 SMs x 128 = 16,384 CUDA cores, 1,008 GB/s (NVIDIA Ada whitepaper). A "CUDA core" is not a CPU core: it is one lane that does arithmetic. Instructions are fetched and scheduled per warp, not per lane, which is why a CUDA core is so much smaller.