Where the milliseconds actually are
Your input crosses four hops. Sensor to OS: modern mice deliver raw events in well under a millisecond. The engine's conversion: Inputize computes curve, smoothing, shaping and the stick position in sub-millisecond tick work. The virtual pad's USB report: set by the poll rate. And the game's own frame — at 60 fps the game samples input once per 16.7 ms, which dwarfs everything upstream combined. The honest conclusion: latency you can influence as a user lives in exactly two places — the poll rate and your machine's scheduling.
Poll rate: 500 is the default for a reason
The stick state is reported to the game at 125–1000 Hz; at the 500 Hz default that is a report every 2 ms. Raising to 1000 halves that — but only if the engine sustains it. The live actual: … Hz indicator next to the setting is the arbiter: green at ≥90 % of the declared value. On a loaded machine a stretched 1000 Hz is measurably worse than a steady 500 or even 250: a report that arrives late is worse than a report that arrives a millisecond less often. Steadier beats faster.
MMCSS: the Games thread and P-core pinning
The MMCSS “Games” thread setting registers the engine's thread with Windows as a game/audio task: the scheduler gives it lower latency and less timer dithering. It is on by default and applies live. On hybrid Intel CPUs, Pin engine thread to P-cores (off by default) stops the thread wandering onto efficiency cores mid-match. Neither is magic — they remove scheduling noise, which is exactly the jitter that makes input feel inconsistent between rounds.
What Inputize doesn't add
No overlays, no screen capture, no injected processes between your mouse and the game. The Input monitor reads the same state the pad reports, so what you see there — including the actual Hz and the live SAB ramp — is the real output, not a simulation of it.
Measuring without fooling yourself
Frame-counting tests at 60 fps mostly measure the game's own 16.7 ms sampling, not your input chain — a 'win' of one frame in such a test is noise. Judge changes the way the engine does: watch the actual-Hz indicator stay green under load, and compare feel across a full session. Consistency is the thing you can actually tune; single-digit milliseconds you cannot control are the thing forums argue about.