FPS Madmanisms

TL;DR - average framerates are bad, so I define the existing terms, show how the current understanding fails us, and propose a new term I call “frame jitter” and measurement process to measure “frame jitter” to better present a solution.

The tyranny of the Second

The biggest challenge since the advent of real-time interactive media, inclusive of video games but also any other, is that performance has been benchmarked and thus averaged across the time domain of a second. While this makes for an acceptable measure of how well the system overall is handling the workload, and is also relatable to non-interactive video media such as movies and TV shows, it unfortunately does not capture the end-user experience well enough.

This is because users - humans - act, process, and react much faster than one second. Thus, any imperfections in the delivery of rendered frames to users can be masked by averaging frame rates over one second, something which is unique to interactive media (and would be rendered outright unacceptable in non-interactive or pre-rendered media).

This is why averaging frame rates over one second - as in Frames Per Second, or FPS, is a bad idea.

Frame Delivery and Frame Pacing

To understand this topic and why there is a need to move beyond FPS, one has to have a basic understanding of frametimes, which are defined as the length of a single frame, the flow of which can be termed frame delivery. Frame pacing then is the regulation of the flow of frames rendered by the GPU and is characterized in terms of how close the outliers of frametimes are to their average, in other words, enforcing smooth frame delivery.

For interactive media, software and hardware solutions must be used in order to better implement proper frame pacing to prevent problems with frame delivery. These solutions need to target problem spots that cause uneven frame delivery, where frametimes can vary wildly from one frame to the next.

Frame Jitter

Frame Jitter, then, I’m defining as “the variance between frametimes measured in milliseconds between the longest frametime and the average frametime”. This can be computed over any length of time, so we can have ‘Frame Jitter per Second’, ‘-per minute’, ‘-per hour’, or ‘-per benchmark run’ as is applicable to the scenario being measured.

Conclusion

The goal is for frame jitter to approach zero, meaning that each frame rendered in the time period measured takes the same length of time.

So, higher frame jitter bad, lower frame jitter good.

(critiques welcome, this is not written by AI, but if it looks like it is, I blame the caffeine)

2 Likes

I understand your viewpoint, but I could say the same thing about engine horsepower… But it isn’t likely to change.

Most benchmarks have a 1% low number, which will give some hints on jitter - but as you say, FPS is imperfect.

Maybe you could introduce the concept of measuring jitter / frame smoothness to MangoHUD, or even code a rough proof-of-concept? Would it be better with a “Average frames per ms” or “Average frames per screen refresh”? Maybe a Box & Whiskers plot would be a better way to indicate FPS averages?

The data is mostly already there, it’s more a question of presenting it and presenting the time it took for each frame to be generated. That’s pretty much no extra overhead.

I do agree with you in general BTW.

This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.