All three tiers, on one scale.
Saved target curves · 0–3,250 MHz on every frequency plot. The selected anchors are marked; actual loaded clocks are shown below.
The biggest kinks get smaller.
Performance
Slope reveals the cliffs
Clock rise per millivolt · lower spikes mean gentler local steps
Hover or tap either chart to compare a shared voltage. Every frequency plot uses 0–3,250 MHz; voltage axes match within each view. No splines, curve fitting, or averaging.
Shape measurements
Calculated from every adjacent pair on the full 127-point curve.
| Metric | Before | After |
|---|
Operating-point context
Selected curve anchors and recorded workload telemetry are different quantities.
| Profile metadata | Before | After |
|---|
Smaller cliffs. Closer loaded voltage.
All nine curves are monotonic. The latest curves reduce the sharpest slope change and the largest adjacent rise versus both earlier sets. Selected clocks are lower too, so the shape change cannot be attributed to smoothing alone. Rising tails now use 2 bins in every tier; both older sets used 2 / 4 / 4.
Latest saved loaded medians are 855 / 915 / 930 mV: 5 mV above the selected anchors. Performance previously recorded 975 mV at a selected 925 mV in both earlier sets. The saved curve anchor is not a strict voltage lock.
Contributor-reported gameplay: Ernold11 reports the Performance overlay moving from about 990 to 935 mV at a 925 mV anchor, and about 2,860 MHz at the latest 2,880 MHz anchor. This separate observation is not the scan median and was not measured locally for this report. Read the gameplay comment ↗
How the measurements work
- Segment slope: Δclock / Δvoltage, in MHz/mV. Native voltage gaps are 5 or 10 mV, so raw MHz jumps alone are not a fair steepness measure.
- Peak slope change: max |slope(next) − slope(previous)| across 125 internal vertices, in MHz/mV. Smaller means the worst local change in gradient is less abrupt.
- Largest point-to-point rise: the largest positive Δclock between adjacent saved voltage points; the accompanying voltage interval is always shown.
- Downward steps: count of adjacent points with a negative Δclock. No smoothing or interpolation is used for any metric.
- Comparison: all nine files belong to the same GPU and use the same 127 voltages. Their stock references differ by at most 15 MHz and are plotted separately. Different selected operating points still affect shape.
What this comparison cannot establish
These JSON files describe saved curves and aggregate scan measurements. They cannot establish game frametime smoothness, long-term GPU stability, a controlled FPS/W gain, or whether every voltage point was independently stress-tested. The oldest profiles are from August 21 and 30; the intermediate set is September 5 and the latest is September 6, 2026. All nine record final_verified: true. The latest ZIP has no scan log, embedded commit identifier, or explicit final-test duration. Build 1027c01 and uninterrupted completion come from the contributor’s comment.
Source: two contributor attachments
Ernold11 · NVIDIA GeForce RTX 5070 Ti
Driver 610.57.04 (reported)
Hardware limit reported: 300 W default, 250 W minimum, 300 W maximum.
September 6: latest three-tier validation ↗
Original before / September 5 attachment ↗
Reference detail: the September 5 JSON requests 925 mV / 3,000 MHz and records an average loaded clock of 2,950.9 MHz. The new comment calls its previous run “2,950 MHz”. This report uses the JSON’s selected anchor and labels measured clocks separately.
Algorithm walkthrough and local RTX 5080 verification ↗
The HTML contains all chart data and works offline. Downloads are generated locally in your browser.
Files and reproducibility
The report generator and reproduction instructions are in the repository. They authenticate both original ZIPs before generating the report.