The bottleneck is typically CPU <-> GPU transfer. If your UI is processing data on the CPU, that information has to be shuttled over the GPU. On a platform like iOS (or MacOS) the core text rendering frameworks are built on the CPU. Text rendering, for instance, is complicated and despite a lot of experience with both native iOS development and recently with compute shader development, I couldn’t really throw out an estimate of how involved it would be to recreate something like CoreText fully on the GPU. And unless you give up CPUs, you’re still left with the problem of transferring data back and forth.

More and more gets ported to the GPU through parallelized algorithms, and smart GPU memory usage — but CPUs still offer a lot in their ability to do generalized computing.

This is why as I’ve gotten older I’ve come to really respect Computer Science fundamentals; the more you understand how to make the most of resource constraints — whether time, or memory, etc — the more magic you can make.