Measure gaming performance before you tweak: frame time, 1% low, latency

Verified 2026-10-272 min read

A repeatable method for Windows 11 26H2 gaming measurements, including passthrough-VM caveats and a ledger for every change.

Tested onPro · Build 26300.9711 · x64

Build: 26300.9711
Edition: Pro
Architecture: x64
Support: Test method
Evidence: E1Official documentation
Verified: 2026-10-27
Page type: Method

The short answer

Average FPS hides the problem. If a game “feels” bad while the counter says 120, look at frame time and 1% low, then change one variable at a time. A tweak without a before/after measurement is a guess.

What to capture

Metric	Why it matters	Tool
Average FPS	Throughput, easy to over-read	CapFrameX / PresentMon
1% low FPS	The stutter you actually feel	CapFrameX
P99 frame time	Worst frames, not the average	CapFrameX
DPC/ISR latency	Driver-level stalls	LatencyMon
Clocks, temps, power	Confirms the run was comparable	HWiNFO

Procedure

Freeze the variablesSettings

Write down the Windows build (winver), GPU driver version, game version, resolution, refresh rate, VRR state, power plan and room temperature. A result without these is not comparable.

Expected: Same build, driver, game version, resolution, quality preset and power plan before and after.

Capture a baselineTerminal (admin)

Use a repeatable scene (built-in benchmark or a saved route). Discard the first cold run. Keep the raw CSV; the page shows the summary, not a screenshot of a number.

Expected: Five runs of the same 60-second scene; median and 1% low recorded.

Change one thingSettings

Snapshot the VM or create a restore point first. If two things change, you cannot attribute the result.

Expected: Exactly one registry, service, driver or in-game setting differs.

Re-capture and compareTerminal (admin)

Run the same five-pass sequence. If the delta is inside the noise band, publish it as “no measurable effect” rather than a win.

Expected: A delta larger than run-to-run noise, with the same scene and settings.

Tweak ledger (schema example)

Rows are filled only after a measured run. “Pending” is deliberate: no data, no percentage.

Change	What it changes	Measured delta	Risk	Rollback
Memory integrity (HVCI)	Disables VBS/HVCI kernel integrity checks; reduces virtualization overhead	Pending (M-V)	R2System-levelStable	Windows Security → Device security → Core isolation → Memory integrity → On
Hardware-accelerated GPU scheduling	Moves GPU scheduling from the CPU to the GPU scheduler	Pending (M-V)	R1ReversibleStable	Settings → System → Display → Graphics → Change default graphics settings → Off
Core parking override	Keeps selected cores unparked; changes scheduler behavior	Pending (M-V)	R2System-levelExtreme Lab	Restore the exported power-plan GUID or `powercfg /restoredefaultschemes`

Passthrough-VM caveats

Label it. A GPU-passthrough VM is M-V. Absolute FPS is not a bare-metal number; use it for same-environment comparisons.
Anti-cheat can refuse. Some titles detect virtual machines. Keep a list; test blocked titles on bare metal or do not publish a result.
Synthetic vs real: a built-in benchmark is repeatable but may not represent a real match. Say which one you used.
No silent edits: if you change the VM CPU model, hugepages or pinning between runs, the comparison is invalid.

What not to claim

“+30% FPS” from a single run.
A registry tweak that “reduces input lag” without a latency measurement.
A VM result presented as if it were a physical PC.
A tweak that permanently disables Defender, updates or a security boundary.
Memory integrity (HVCI): measured cost — check the security/performance trade-off before disabling it.