It balances the computing and memory bandwidth bottleneck so by a lot, with continuous batching you can easily see a x10, x20 or more.