Latency
Three separate delays sit between a button going down and a reader acting on it. They have different causes, and only one of them can actually be removed. A fourth, reader gating, is the one buffering exists for.
| Delay | Cause | Removable? |
|---|---|---|
| Polling | Input is sampled, not pushed | No, only shortened |
| The binding step | Update runs after the frame's fixed steps | Yes, with a trade |
| No rollback | A step that ran without the intent is finished | No |
| Reader gating | The reader is not ready to act yet | Tolerated by a window |
Polling delay
You learn about a press at the next input update, so on average you wait half a sampling interval and at worst a whole one. Sampling more often shortens it. Nothing removes it.
The binding step
In the default setup the Input System processes events right before every Update, and Update runs after that frame's fixed steps. A press captured there binds to the next step, which begins on the next frame. At 60 fps that is about 17 ms; at 12 fps it is 83 ms.
This one is removable. The Input System can instead process events right before every FixedUpdate (Update Mode set to Process Events In Fixed Update), and then a press is refreshed and consumed in the very same step with no forward binding at all.
It is a real trade rather than a free win:
- There is no per-frame input, which UI and camera code usually want.
- Two taps falling between two steps coalesce into one, because there is only one input update per step.
Which phase may poll a device
That follows from the setting, and only from it. In the default mode, reading device state from FixedUpdate logs an error in the editor and in development builds, and silently returns the dynamic value in a release player. In fixed mode the error is the other way round.
So poll wherever your project's input system updates, and stamp the tick that phase gives you. Rule 5 holds either way, because it is written in terms of steps rather than phases.
| Input System update mode | Writer polls in | Stamps |
|---|---|---|
| Process Events In Dynamic Update (default) | Update | T + 1, consumed next step |
| Process Events In Fixed Update | FixedUpdate, before its readers | T, consumed this step |
No rollback
None of this is recoverable after the fact. Once a step has run without an intent, that step is finished.
Set will take a tick in the past, but no reader revisits a step it has already run, and the wrapping comparison simply treats a backdated tick as older. Backdating makes an intent expire sooner, never arrive earlier.
Reader gating
A press often arrives while the reader is not yet willing to act, mid-attack or mid-air. A window does not reverse the wait, it tolerates it, which is why widening one feels like removing lag even though the press arrived just as late.