I think you're right about GPU tending to have worse latency, but I think it's because that's the way the swap chain is set up and the use of blocking present calls. I've been studying this and am pretty sure I know how to make low latency on Windows work, mostly it's following the advice in [1]. If you do it right, then on modern hardware, DirectFlip kicks in and that saves you an additional frame that would ordinarily be eaten by the compositor. I plan to measure this at some point in the not too distant future, or would be happy to mentor somebody interested in taking it on.

[1]: https://devblogs.microsoft.com/directx/dxgi-flip-model/