Why don't we have some programmable, trimmed-down cores acting as DMA engines doing this? This sort of asymmetric multiprocessing would be very useful...