There seems to have been a bit of activity on my twitter recently regarding SFVAE input lag. I thought I might type out some thoughts on it here, rather than in short little Twitter sentences.
iBananaKiller tested the input lag first, and posted about it on Reddit stating that it was about 4F. This was obviously quite a reduction from what we had seen previously (5.3 by his own tests).
As per iBananaKiller’s own posts, Displaylag also tested, and found no change from previous versions.
Now, iBananaKiller uses a method relying on visualising the button press, which obviously attracted a lot of criticism. The problem as I see it and he made without much success is that even if you take a pesimistic approach to how early the button is pressed, his results would still be noticeably faster than display lags.
So assuming both testing methods are acceptable, how can they both be so different and also correct.
I was overseas when a lot of this was happening, but when I got a chance to test, the results I posted were these.
Now, if you don’t know how to read my report card format, please watch this first.
What is interesting is that most games would look something like this,
You can see that normally the results will be going down one column, and at at threshold they will flick across to the next column (i.e. when the command comes too late in the frame to make it in time, and ends up on the next frame.)
Compare this to SFV’s result. Initially this appears to be happening, with most results occuring in 4F and then some in 5F. Then it goes to six frames. Then it goes down to 3F and back to 4F.
Now, if all you saw were the 4-5F results you would get results similar to displaylag’s.
If all you saw were the 3-4F results, you would get results similar to iBananaKiller’s.
The point being, the input lag for SFV is quite inconsistent. I believe this happens to some extent with T7, and suggests that it may be a UE4 issue. I did not demonstrate this on my latest test for T7, however given displaylag’s input numbers are lower than mine (and should be higher by the nature of the differences in testing process) I would believe it is still there.
This is actually fairly frustrating for input lag testing. What I and presumably most people want is a hard and fast number which after a single test (with ~100 repetitions for my setup) gives a meaningful result. The way that the results have come out for SFV does not provide this.