Memory bandwidth is the main bottleneck. It got better with M3/M4. ANE is really fast in FLOPS but low in memory bandwidth.