NVLink is the most notable (>600GB/s GPU-to-GPU communication) whereas the best case scenario on 30x0/40x0 GPUs is peer-to-peer PCIe (~30GB/s) and average case is routing PCIe through the host (~15GB/s)