CS2: NVIDIA Low Latency Off vs Ultra

Ultra recorded a 10.16% lower median 1% low and a 7.76% higher p99 frame time. All four pairs favored Off for these metrics, with non-overlapping ranges. Median average FPS was 1.98% lower with Ultra, narrowly below the declared 2% decision threshold. This does not establish which setting has lower input latency.

Input latency: not measured. Input tracking was enabled, but no valid latency samples were captured. Missing values are not zero milliseconds. These frame-timing results cannot establish click-to-photon latency or a universal recommendation.

Two warmups (A → B) precede four measured pairs: A → B, B → A, B → A, A → B. All eight measured tests were accepted on their first attempt; none were 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

CS2 build

25966348

Scene

FPSHeaven Ancient · workshop 3472126051

Display

1440 × 1080 · 144 Hz · fullscreen

Graphics

4× MSAA · NVIDIA Reflex disabled

Windows

High performance power plan · Game Mode on

Capture

PresentMon 2.6.0 · fixed60-offset10-v1

Only the CS2 application profile setting under test changed. The global NVIDIA profile was unchanged. The game relaunched for every pass, with a 5-second prelaunch wait and 15-second cooldown. The other tested NVIDIA setting remained fixed: Texture filtering quality = Quality.

Each PresentMon window begins 10 seconds after the observed VProf start and lasts exactly 60 seconds. Average FPS = 1,000 ÷ mean consecutive CPU frame-start interval (ms). 1% low = 1,000 ÷ mean of the slowest ceil(1%) intervals. p99 is the nearest-rank 99th percentile interval. These are frame-start timings, not physical input latency.

The declared decision rule requires a median difference of at least 2%, all four paired differences in the same direction, and non-overlapping observed ranges. Failing this rule does not prove equivalence or statistical significance.

See the PresentMon measurement documentation.

Data quality & downloads

All 20 PresentMon windows in the combined two-setting session were independently recomputed from frame data; all 20 VProf logs were reparsed. All tests used one dominant presentation chain. Sampled foreground checks had a maximum observation gap of 127.24 ms; benchmark-boundary uncertainty was at most 277.67 ms. Sampling cannot exclude events between observations.

The complete session finished with restoration verified: original CFG hashes and attributes and NVIDIA settings matched their starting state; collectors completed and no cleanup errors remained. Whole-GPU power readings were unavailable, so this report makes no power-saving claim.

• One PC, one CS2 build, one synthetic scene and four measured repeats per setting; results do not predict every match or system.
• No valid input-latency samples were captured. Missing latency is not zero milliseconds.
• Fullscreen is the configured mode; independent-flip presentation does not establish exclusive fullscreen.
• Steam Performance Monitor was off. Non-windowed OS-managed Game Bar hosts were allowed; this is not a zero-background-process claim.
• No outlier removal or automatic retry was used in this completed series.
• The 2% decision threshold is an operational rule, not a statistical confidence interval.

Downloads contain only allowlisted derived measurements, including warmups. Private captures, device paths, game configuration contents and account information remain local.