Yup, people are focusing a bit too much on where the bottleneck is and what Mantle is aimed at. Generally speaking DX isn't brilliant at extracting performance out of gpu's ignoring the CPU. It's using generally higher level concepts, isn't as optimised as it can be(by a fairly long way) for the hardware and currently doesn't send data to the GPU in the best optimised way possible(FAR from it).
Mantle itself will enable dev's to optimise code as it runs on the GPU, use the GPU's most up to date features in the way they deem gets the best performance out of it.
A side benefit is a huge reduction in cpu usage, this is not the intention of Mantle, just what happens when you take the API out of the picture API prevents by a long way the best performance on GPU and CPU.