Does maximum performance improve CS2 FPS?

The measured ranges overlap for all three metrics, and the paired differences change direction. Small median differences do not establish a winner.

These are benchmark-section measurements on one PC, not a prediction of your match FPS or input latency. TARS is not credited with an FPS gain.

Chronological order: A–B / B–A / B–A / A–B. A = Optimal Power (NVAPI 5); B = Prefer Maximum Performance (NVAPI 1). No outliers removed.

Test setup & definitions

CPU

AMD Ryzen 5 7600X3D

GPU

NVIDIA GeForce RTX 4060

DRIVER

32.0.16.1742

OS

Windows 11 Pro 10.0.26300

CS2BUILD

25966348

DISPLAY

1440 × 1080 · 144 Hz

MSAA

4x

SCENE

FPSHeaven Ancient · workshop 3472126051

CAPTURE

PresentMon 2.6.0; 60 seconds, starting 10 seconds after the observed VProf start. 277.4 ms maximum boundary uncertainty.

CONTROLS

FPS cap removed for tests; other saved graphics settings and map hashes held constant. Only the existing CS2 NVIDIA application profile power setting changed; the global profile stayed unchanged.

Two warmups (A→B), then four measured pairs. Five seconds before each launch; 15 seconds between tests. Original NVIDIA inheritance/override, ten CFG files, file attributes and Steam performance preferences were independently verified after restoration.

Average FPS = 1,000 ÷ mean CPU frame-start interval in ms. 1% low = 1,000 ÷ mean of the slowest ceil(1%) intervals. p99 is the nearest-rank 99th percentile frame interval. These are rendered-frame timing measurements, not displayed-frame or latency measurements. See the PresentMon documentation.

An advantage requires all four paired differences to agree and non-overlapping min–max ranges. Four repetitions are not statistical certainty.

Data quality & downloads

Post-capture analysis correction. The initial export stopped because one complete 60-second capture had 9,907 frames, below an arbitrary 10,000-frame cutoff. That cutoff would reject slower valid results. It was removed and all ten captures were reanalyzed with the same corrected rule. PID, dominant swapchain, full window coverage and timestamp ordering checks remained. The original failed journal was preserved locally.

The lowest result, 25.61 FPS 1% low in the first Optimal measurement, remains included. No test was repeated or discarded to improve the comparison. Earlier incomplete attempts and the separate UI diagnostic are not part of this series.

Game Bar hosts were allowed to start/stop in the background under Windows control. No visible host windows or known recording servers were detected at pass boundaries; that does not prove all overlays were absent. Capture overhead, background variation and the map’s own performance display were not isolated. No ETW loss was reported; zero loss was not independently established.

GPU clocks and temperatures are retained in the derived download. Watt readings were unsupported, so no power-efficiency claim is made. The original 60-second window contains 9,907–10,786 frames per test; all ten metric recalculations independently matched the raw CSV.