Measuring keyboard-to-photon latency with a light sensor (2023)
thume.ca
thume.ca
I remain puzzled on the goal. Does he optimize his setup for a good coding experience? (this should definitely include the actual keyboard in the measurements!) Is he optimizing editors? Picking a new monitor? As a learning excercise?
I'm all for "because I can" as a valid reason, but that often leaves my projects dangling halfway complete. Are we there yet? Don't know - not sure where I want to be.
It just needs a lot of light (phone flashlight is sufficient at short range), or to emit light by itself like a screen (or microwave segment display, as a random example where you can see it lights up each digit individually and they're never all on at the same time, but we can't perceive that).
Method:
To measure, film the computer screen, positioning the phone in a way that it also captures the keyboard (best head on, since you're interested in when vertical movement stops). Press record, strike a key, and go frame by frame on the recording. The moment the key bottoms out is pretty clear by your finger stopping to move down (if the focus is good, you also see a ripple effect in your hand). I know the trigger point will be before the key's very bottom, but the whole key's travel time is on the order of 2-3 frames = ~2-3 ms, which is good enough for me. Then, count frames until the computer screen updates. I usually wait for all channels before considering it updated (my work laptop emits only red on a separate refresh cycle, sometimes noticeable if you look away with some lucky timing in the right way)
Similarly for the mouse, though I usually pick it up and hold it in one hand before the screen and click the button in the air so that you can hold it in a good (for the camera) position more easily. By lifting your finger off the mouse button until the moment where you're going to press it, the separation is very well visible on camera and the moment of pressing quite exactly determinable
Results:
On the LUKS disk password prompt at boot, latency is about 19 ms on a laptop. It feels extremely fast, a little like it shows the asterisk before I've finished pressing the key. After booting the OS, in gnome-terminal I get 48 ms. Connecting an external screen and keyboard via a cheap USB hub, it grows to 62 ms. Finally, inside the virtual machines we're supposed to use for everything (because security), my input-to-photon latency is 134 ms. Guess I can see why that boot prompt feels like a time traveller compared to what I need to use all day long!
On a different setup (a family member's iirc, running Windows) with an external monitor that has three gaming modes, called standard, fast, and faster, the latencies are 67, 67, and 62 ms. Pretty disappointing for a "low latency gaming" mode!
---
Compared to the submission:
> One source of latency users experience that my tester doesn’t measure is keyboard latency.
The high speed camera method will include this, which is an advantage if you want to know how much latency your setup has, but doesn't help you in picking apart where the latency occurs
I did notice with a daylight lamp, which seemed to instantly go from 0 to fully bright, that my phone would show a left to right progression on some recordings regardless of the lamp's orientation. This must be the sensor being scanned from left to right, but for all I know it completes each scan at the indicated frame rate. If there were interpolation, I'd expect to see things like the microwave digits to appear in a half-on state for each interpolated frame. I can check again but don't think this happened to anything where no half-states are expected
Anyway, when measuring things between a dozen and a gross amount of milliseconds, also the cheap 240 fps (~4 ms) slow motion cameras will be able to give a good estimate. If you think you're capturing 960 but it's actually half, that's lying to the consumer but still a decent measurement
EDIT: well, apparently these are photo resistors, which are much slower and could in fact have substantial transition times (brief lookup indicates 10ms), so could definitely be relevant.