3 ms·

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

• gpderetta 9y ago
  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).