@HangTime no disagreements.
Not really, PCs and consoles run the same logic, and the compute workload per fps is generally the same. Last gen console games were OK being 30fps so developers had more leeway with game logic, which allowed 60/120fps on PCs with more powerful CPUs. With 60fps being the minimum accepted fps in consoles now generally developers need to limit their games to Zen 2 cores being able to offer 60fps, which is what I'd call adequate gaming. That's 16.6ms per frame, all game AI will be written so that a Zen 2 core can process a frame's worth of compute in 16.6ms, until we get new consoles.
You could accelerate AI on the GPU ever since there were GPUs, just not every type of AI. Those are high-throughput latency-insensitive problems such as parameter optimisation or ML model training, generally matrix operations, these do not happen during gaming. Game AI runs on CPUs in every single game. Now if we're talking chess or go, those are different because nature of the problem completely changes, no longer latency-sensitive, and solutions are matrix operations, so those can run on GPUs even though inference still happens on CPUs in 99.99% of cases.
Generally, games use 6-8 main threads these days doing bulk of the compute, most AI workload is in one or two, depending on the type of game. Nature of game AI is that it's a sequential problem so very ill-suited for GPU processing and the added latency of copying data to GPU memory would just make it unfeasible. In the end game AI is generally not a parallelisable problem so multithreading doesn't help (that's why adding more cores beyond 6-8 doesn't give you more fps).