https://www.stereotool.com/download/Thi ... 92-060.zip
PiMP can now go down to 2.85 ms, but loading or browsing through the web interface can occasionally cause a small audio glitch. The CPU load is still a bit higher than at higher latency settings, probably due to framework and thread synchronization overhead stuff, and recovering from glitches is a lot harder at these extremely low latencies. (There might still be some settings that we can tweak to improve the behavior).
To run PiMP at 2.85 ms, set the "Latency" setting on the I/O page to 64 samples instead of 128. To get rid of most of the glitches, increase the output buffer size to 0.3 ms. That adds 0.33 ms of latency (3.18 ms in total).
Edit: Good news! It turns out that glitch recovery mainly has a problem with low frequency tones, which are normally not present in microphone input anyway. I have been testing for glitches by playing a 30 Hz tone - after increasing it to 300 Hz I see no unrecovered glitches in a 3-hour measurement, with the latency at 2.85 ms! So... it looks like PiMP can run at 2.85 ms. I do hear some effect in the audio for which I need to search for the cause at that low latency though; not sure yet if it will actually be noticeable in speech.