This could be bad or great depending on the use case. - Bad: bottlenecked access to the CPU and its memory - Good: Bypass the CPU on device to device transfer

This could be bad or great depending on the use case. - Bad: bottlenecked access to the CPU and its memory - Good: Bypass the CPU on device to device transfer, in the style of NVidia's GPUDirect (https://developer.nvidia.com/blog/gpudirect-storage/ https://developer.nvidia.com/blog/gpudirect-storage/)