a cpu is a few very complicated cores, a gpu is thousands of dumb cores.
its easier to make something doing something low in instructions(gpu) faster than something that has a shit ton of instructions(cpu) due to like you mention, branch prediction.
modern cpu performance gains is focusing more on paralellism and in the case of efficiency cores, scheduling to optimize for performance.
GPU wise, its really something as simple as GPUs are typically memory bottlenecked. memory bandwidth (memory speed x bus width with a few caveats with cache lowering requirements based on hits) its the major indicator on GPU performance. bus width is fixed on a hardware chip design, so the simpilist method to increase general performance is clocks.