TitanCSA

Measurement / Reproducibility

What does an input latency number actually measure?

Read the endpoints, the receiver and the missing evidence before comparing two numbers.

TitanCSA project documentation · Dataset: 17 November 2026 · AI-assisted maintenance

TitanCSA Benchmarks is a publisher-operated Windows input lab. The initial campaign contains 89 validated measured series and retains 31 non-validated series. That total includes separately labeled stability and instrumented trials, not 89 interchangeable latency comparisons.

Start with the two endpoints

The lab's submission-to-observation metric begins immediately before a producer's public SDK call. It ends when a separate executable observes matching content through Windows HID ReadFile, Raw Input, XInput or WinMM. Both processes use QPC in the same guest.

That interval includes SDK work, virtual USB or other Windows stack behavior, queuing, polling and scheduling. It is not a measurement of kernel instructions alone. It is also not physical USB latency: no external analyzer or dedicated physical input device was used.

Compare the same observation contract

ObservationWhat the executed path can showWhat it does not establish
HID ReadFileMatching vendor-HID or DS4 report receptionEvery controller protocol or physical transport
Raw InputExplicit fixture keyboard/mouse delivery, with sequence validationUncoalesced one-to-one free-running mouse reports in every configuration
XInputClosed-loop visibility of validated controller stateA free-running packet-loss rate or hardware polling ceiling
WinMMObserved normalized axes and selected button stateXInput compatibility or all joystick/force-feedback capabilities

HID, Raw Input, XInput and WinMM results should not be pooled into one best-driver ranking. Vdevice/ViGEm controller comparisons and Vdevice/vJoy joystick comparisons have separate profiles, SDK paths and receiver contracts.

Keep the tails and the failed runs

Each run warms up before its measured sequence. Raw timestamps remain available; the reports show per-run quantiles and retain outliers. A median of run P99 values is not a percentile of all pooled samples. A requested send rate is not automatically the observed rate or the physical USB polling frequency.

A failed setup, incomplete sequence, mismatched payload or uncorrelated filter capture must not silently become a performance score. In the initial campaign, extra captured GIP/ViGEm protocol reports prevent some series from passing the strict one-to-one filter rule. Those records remain in the archive.

Separate speed from stability

The campaign reproduced two Vdevice controller-teardown guest crashes at a WDF queue-purge call. The open incident includes a sanitized signature, exact driver hash and the unresolved lifetime-ordering question. Successful timings do not cancel that result.

A five-minute monitored vendor-HID trial received its full sequence, but does not prove multi-hour stability or freedom from leaks. A separate WPR trial recorded module-attributed DPC activity with no lost events reported; it is not total driver CPU or an uninstrumented latency result.

Use the data to ask a narrower question

The guest had fixed resources, but shared-host background load was not frozen. HVCI was off in the guest, while Secure Boot stayed on. The results do not establish production qualification, physical-device compatibility or a universal performance advantage.

The useful next question is specific: does the same change appear with the same profile, receiver, version, fixture and controlled environment? That can lead to a reproducible improvement rather than a headline that outruns the evidence.