Transistor count != complexity. A GPU is hundreds of simple ALUs stacked together. Minimal branch handling, no real speculative execution, no reordering buffers

Transistor count != complexity. A GPU is hundreds of simple ALUs stacked together. Minimal branch handling, no real speculative execution, no reordering buffers, etc... Very simple, very slow "cores", and just a whole shitload of them "copy/pasted" together. Some moderately complex management blocks then distribute work to all those ALUs, sure, but still nothing close to the complexity of a modern CPU core. Which is also why GPUs can be built so large. It's much easier for them to handle defects than for CPUs. The loss of a single core cluster on a GPU isn't nearly as significant to the product's overall performance as the loss of a single core on a CPU, and the amount of transistors you need to turn off to handle that defect is much less.