Not sure why that would be true. Slow UMA (like the vast majority of Intel and AMD desktop chips with 128 bit wide memory) hurts GPU performance. However the

Not sure why that would be true. Slow UMA (like the vast majority of Intel and AMD desktop chips with 128 bit wide memory) hurts GPU performance. However the M4 Pro has 256 bits wide, M4 max 512 bits wide, and M2 Ultra has 1024 bits wide. GPU workloads are latency tolerant and embarrassingly parallel, don't see how allowing a CPU to make random accesses is going to hurt the GPU much.