> inevitable future of computing that is somewhere between those two experiences currently.
Not sure that ever happens. The two are very different in their performance characteristics. CPUs have to stay focused on single-threaded performance.
Single threaded performance = minimizing latency, that’s why in CPUs so many transistors and electricity is spent on cache hierarchy and coherency protocol, instructions reordering, ILP, microops fusion, prefetcher, branch prediction incl. indirect one, and so on.
GPUs don’t care about single-threaded performance, all they care is throughput on massively parallel tasks. They ignore memory latency, switching threads until data arrives, makes caches much simpler. They don’t need high clock frequencies (this alone improves their flops/watt by a large factor), don’t do branch prediction, memory model is less strict, and so on. That’s how they can spend that many transistors on FPUs and ALUs.
P.S. Intel tried many-slow-cores with Larrabee / knights landing / etc., all they got was a very expensive lesson, people didn’t want that stuff. I’ve heard a rumor initially Sony wanted to use another Cell CPU for the GPU in PS3, again, didn’t worked well enough and they negotiated with NVidia instead.