[LAD] Not all XHCI (USB-Controllers) are made equal

Hi! This mail is just a heads-up about a finding discovered with the help of the linux-usb mailing list and which I thought some other people might benefit from. If this is nothing new to you, then please do ignore it. Background: The USB audio class 2.0 specification dictates that isochronous transfers (i.e. audio frames to/from an audio interface) happen every "micro-frame". USB micro-frames are 125 microseconds (us) apart. 125 us = 0.125 milliseconds (ms). The majority of USB audio interfaces (at least those that I have) use synchronous audio-streaming, i.e. the sample clock is derived from the bus clock. There are definitely interfaces that use e.g. adaptive or asynchronous modes and this discussion would have to be altered for these. Given the case of synchronous mode isochronous transfers at a sampling rate of 48000 Hz (= 48 kHz) this would correspond to 6 audio frames per USB micro-frame. 48000 frames/second * 0.000125 seconds = 6 frames So in principle a very well behaved audio interface attached to a very well behaved USB controller sitting in a well tuned system _should_ be able to achieve a minimum round-trip latency of 2 * 6 frames = 12 frames or 250 us. This leaves out additional buffering inside the audio interface and additional latency by anti-aliasing and reconstruction filters. The first caveat to above in the context of Linux: The snd_usb_audio driver does not seem to support period sizes of just 6 frames. It _does support 12 frames though, which is nice. And here is the other caveat which lead to the title of this mail: Some controllers are behaving "worse" than others. There's this little thing in the XHCI spec which amounts to the following: The XHCI can specify in a register how many micro-frames have to buffered at all times for outgoing isochronous endpoints. The Intel XHCI in my ASRock N100dc-itx main-board for example request a whole USB frame (which corresponds to 8 micro-frames or 1 ms) to be buffered at all times. Another XHCI that I have (a Renesas controller) only requires one micro-frame. This has direct consequences on the kind of period sizes and number of periods that are usable on these controllers. These consequences don't explain everything but at least you know that you can't ever get better than this limit. For example for a period size of 48 frames I need to use 3 periods on the Intel controller (resulting in 3 ms round-trip latency), but for the Renesas I can use 2 periods at 48 frames (resulting in 2 ms round-trip latency). On the Renesas controller even 2 periods at 24 frames works fine (1 ms round-trip latency). I can lower the latency on the Intel controller by using a smaller period size but more of them as long as the buffering requirement of the controller is satisfied. One stable setting is for example a period size of 24 and 5 periods which results in a round-trip latency of 2.5 ms. So what's the take-away here? In some cases, if you are chasing stable low-latency operation using a USB audio class 2.0 device it might just be worth installing a different XHCI in your computer (the above mentioned Renesas controller is just a PCI-Express card which sits in a slot in my N100DC-itx board) that is better behaved than the one you currently have. Another take-away is that the above limitation only applies to the outgoing direction (playback). If all I was interested in would be the capture direction then 2 periods at 48 frames would work fine even on the Intel controller. Kind regards, FPS