But before ML was on dedicated hardware it was on GPUs, and programmable GPUs have been around for almost two decades now. They're getting more and more CPU-like, but so far there's no sign of them actually being subsumed by CPUs. If anything, attempts to make CPU-GPU hybrids have historically tended to fail, such as Larrabee or the Cell processor on PlayStation 3. Of course, ML is a somewhat different workload from the usual use of GPUs... but not that different. The future could be different from the past... but as CPU speed increases get slower and slower, why not expect things to shift even further in the direction of dedicated hardware?
But before ML was on dedicated hardware it was on GPUs, and programmable GPUs have been around for almost two decades now. They're getting more and more CPU-like, but so far there's no sign of them actually being subsumed by CPUs. If anything, attempts to make CPU-GPU hybrids have historically tended to fail, such as Larrabee or the Cell processor on PlayStation 3. Of course, ML is a somewhat different workload from the usual use of GPUs... but not that different. The future could be different from the past... but as CPU speed increases get slower and slower, why not expect things to shift even further in the direction of dedicated hardware?