Domain-specific accelerators have always had their own memory. Look at GPUs. I don't see why a power user who needs speed would be put off by some DMA. Maybe A

Domain-specific accelerators have always had their own memory. Look at GPUs. I don't see why a power user who needs speed would be put off by some DMA. Maybe Apple ends up implementing something like CXL down the line, anyway.