The newest Nvidia stuff is offering some of that, though it may be specific to CUDA. The thing is, the bandwidth mismatch between the GPU and its memory and th

The newest Nvidia stuff is offering some of that, though it may be specific to CUDA. The thing is, the bandwidth mismatch between the GPU and its memory and the PC memory to GPU memory is crazy. There seems to be extra latency involved there too. As long as we plug a special-purpose video card into an expansion bus, I suspect the best performance will come from manually-scheduled transfers.