Gpus left x86, sparc, aarch64, riscv in the dust for compute long ago. What is this unicorn compute task that can only execute on a cpu core, and fastest in x86? Or do you mean x86 systems still hit hardest, when running high end gpus? That’s a system space, power and packaging problem, pcs are just built big, with standard busses. You could make aarch64 systems like that, but not cost effective since pcs already exist.
The main problem with x86 is the cisc to rops conversion. In power efficient designs, it is a huge weight to tote around that arm does not require.
Future heavy computing will center on gpus. Already, most people don’t realize how much raw processing power is available in a rtx2080.
The truth is that for most of the things that are still interesting to do on Cpu can be done on a aarch64 core. Heavy compute is handled much more effectively on a gpu.
I suspect the reason we just don’t use repurposed gpus as opposed to current cpus is because most software can’t take advantage of it.
Most software is still written for one to four cores at most. Parallel processing to the tune of hundreds of threads is something the industry just isn’t doing, and isn’t practical for most workloads anyway.