How we benchmark

SageTweaks publishes performance guidance in optimization guides. That is only useful if you can see how any numbers were measured. This page documents the test rigs, the tools, and the limitations.

If a guide on the site quotes a benchmark, you should be able to map it back to the methodology described here. If you can't, that's a content bug — please email support@sagetweaks.com.

We test on three reference rigs that span the realistic range of reader hardware. Every benchmark in a guide will say which rig the result came from. The rigs are kept on dedicated test SSDs with a clean Windows 11 install image we re-flash before major test runs.

Frame timing is captured with CapFrameX or PresentMon. Both record per-frame present times directly from the DirectX/Vulkan runtime, which is the only honest way to capture stutter.

Each test run is at least 90 seconds in a representative gameplay scenario — not a menu, not a static map view. We report average FPS, 1% lows (the worst 1% of frame times), and the standard deviation of frame time across the run.

Each tweak is benchmarked in isolation against a clean baseline before being combined into a stack. This is the only way to attribute a measured change to a specific tweak rather than to test variance.

Three runs per configuration; the median is reported. If the variance between the three runs exceeds 10% of the average, the test is re-run on a freshly rebooted machine.

For games that support NVIDIA Reflex Analyzer (Valorant, Apex Legends, CS2, Overwatch 2, Fortnite, Warzone), we use the analyzer's end-to-end measurement: button press to photon on screen.

For games without Reflex support, we measure DPC latency with LatencyMon, USB polling rate with the Windows-side reported HID interval, and report only those values rather than implying a full click-to-photon number.

We never quote an input-lag improvement value for a game we did not actually measure on the affected rig.

A clean baseline is captured on a freshly imaged Windows 11 install with default driver versions and no third-party software running.

Each tweak is then applied one at a time on that baseline and measured. The tweak is reverted (we use SageTweaks' own undo where applicable, otherwise a registry export and reapply) before the next tweak.

When tweaks are stacked into a "full optimization run," the stacked result is reported separately so the reader can see the marginal contribution of stacking.

Background system noise is real. Windows Update, scheduled tasks, telemetry, and antivirus scans can each shift FPS by single-digit percentages. We exclude obvious noise events from the recorded results, but we cannot fully eliminate them.

Every measurement is on consumer hardware, not server-validated equipment. We use NVIDIA-grade software measurement tooling, but we are not a hardware lab.

Game patches change performance profiles. A test run from January may not reflect the May build of the same game. The "last verified" date on each guide is the actual date of the most recent benchmark run that produced the numbers in the article.

Hardware variance between identical SKUs is real. Two RTX 4070s from two manufacturers can differ by 2–4% at default boost behaviour. We do not chase that level of precision; we report numbers good to a percent or two on the rigs we tested.