Left: 4 big, powerful CPU cores. Right: a grid of 256 tiny GPU cores. Pick a workload and press run. Watch which chip wins — and notice the only thing that changed is the shape of the work.
Model (simplified): a CPU core is ~20× faster at a single task than a GPU core, but there are only 4 of them. The GPU has 256 slow cores. For 4 heavy tasks, the fast CPU cores win and most GPU cores sit idle. For thousands of tiny tasks, the GPU's sheer width overwhelms the 4 CPU cores — that's throughput. This is why you match the chip to the shape of the work.