Once more time I am ask AMD about improvement of their CPU/GPU/Chipsets.
I propose some cosmetic changes to CPU on-chip memory controller which can boost up any PCIex<->Mem operations.
The main point is dedicated usage of one of two memory channels by PCIex devices without any CPU extra time.
For example CPU work in ganged mode (default for a series of m/b) and it can dynamically switch to un-ganged, and provide icluseve usage of one of two channels to external ( for example PCIex) data transfers using DMA. In this case DMA memory operations will be more efficient, due to minimizing of CPU role in this process. It can improve performance of PCI(ex) devices including GPU.