[mrb]

Correct, I misread the code. However bumping the pipe size buffer bumps the speed from 28 to 74 GB/s on my Skylake CPU (i5-6500).

Edit: I optimized further and now get 123 GB/s See https://www.reddit.com/r/unix/comments/6gxduc/how_is_gnu_yes...

[gpderetta]

next step is playing tricks with memory mapping so that the buffer is large in virtual memory but only takes a single page in physical memory, fitting in L1 (at least on the read side).