I expect GPU hardware to specialize like Google’s TPU. The TPU feels like ARM in these AI workloads where when you start to run these at scale, you’ll care about the cost perf tradeoff for most usecases.

> CPU/GPU share the same RAM AFAIK.

This depends on the GPU I believe Apple has integrated memory, but most GPUs from my limited experience writing kernels have their own memory. CUDA pretty heavily has a device memory vs host memory abstraction.