Benchmark repeatability check
On this page
Archival research. The data below were obtained on 2026-09-04 on version 0.5.2. The tables, CSV and script retain the old Performance / Core Latency / Consistency formula with weights 50/30/20. These are not the current online leaderboard scores and not a test of the new CPU/GPU/Latency model. The CV value of 0,256% cannot be automatically carried over to it. The current formula is described separately.
After this series, the methodology and the execution of GPU loads changed. Recalculating the old numbers does not replace a new hardware series. In this update, the original observations and the published calculations were not changed.
In six runs at 1920×1080, the coefficient of variation of PC Score is 0,256%. Two runs at a lower resolution deviated from the mean of these six by +0,392% and −0,070%. On one PC the spread turned out to be small, but these data cannot be used to estimate noise on other configurations.
Here is the analysis of all eight runs from 2026-09-04. No run was excluded from the tables. The calculations were recomputed on 2026-09-07.
Conditions
Section titled “Conditions”| Parameter | Series data |
|---|---|
| CPU | AMD Ryzen 7 7800X3D, 8 physical cores / 16 logical processors |
| GPU | NVIDIA GeForce RTX 5070 |
| Windows | Build 26200; version string in the results Windows 10.0.26200 |
| Engine | PCBenchmarkX 0.5.2 |
| Profile | Candidate, identical durations and test composition |
| Scoring model | score-model-1.2-candidate-1 |
| Internal resolution | 1920×1080 in all eight runs |
| Output | Full-screen borderless window |
| Environment state | A detected hypervisor is noted in all results |
According to the author, Windows with BoosterX optimizations was used. The full list of settings from the series has not been published. The archive also contains no comparison with non-optimized Windows. The noted hypervisor does not prove that the run took place inside a guest VM.
The summary does not include the graphics driver version, and the archive does not allow determining component temperatures or the full list of settings during the tests.
How the runs are separated
Section titled “How the runs are separated”The first three runs were performed in one Windows boot session, the next three each after a separate reboot. This is visible from the system uptime: the calculated boot moment coincides for No. 1-3 and differs for No. 4, No. 5 and No. 6. The last two resolution tests were performed after No. 6 without a new reboot.
| No. | Windows boot | System uptime before the test, min | Output resolution | PC Score |
|---|---|---|---|---|
| 1 | A | 4,94 | 1920×1080 | 14 198,58 |
| 2 | A | 7,56 | 1920×1080 | 14 229,16 |
| 3 | A | 10,23 | 1920×1080 | 14 186,01 |
| 4 | B | 0,67 | 1920×1080 | 14 239,20 |
| 5 | C | 3,52 | 1920×1080 | 14 224,65 |
| 6 | D | 7,80 | 1920×1080 | 14 141,08 |
| 7 | D | 10,62 | 1280×768 | 14 258,79 |
| 8 | D | 13,48 | 800×600 | 14 193,24 |
Run No. 4 started less than a minute after boot. The waiting time before the tests differs, so it is impossible to isolate a “reboot-only effect” here: the behavior of the software after system startup is superimposed.
Repeatability at 1920×1080
Section titled “Repeatability at 1920×1080”CV here is calculated between complete runs, via the sample standard deviation with divisor n − 1:
mean = sum(x) / ns = sqrt(sum((x − mean)^2) / (n − 1))CV, % = 100 × s / meanrange, % = 100 × (max(x) − min(x)) / mean| PC Score series | n | Mean | CV | Range / mean |
|---|---|---|---|---|
| One boot session, No. 1–3 | 3 | 14 204,58 | 0,156% | 0,304% |
| Different boot sessions, No. 4–6 | 3 | 14 201,65 | 0,373% | 0,691% |
| All 1920×1080, No. 1–6 | 6 | 14 203,12 | 0,256% | 0,691% |
Comparing the first and second triplets gives a difference of −0,021%. The spread between boot sessions is higher than within one boot session, but in this series it is still small. Three runs per group is too few for a strict conclusion about the maximum error.
| Component, No. 1–6 | Mean | CV between runs | Range / mean |
|---|---|---|---|
| Performance | 11 437,74 | 0,121% | 0,330% |
| Core Latency Score | 21 750,66 | 0,392% | 1,129% |
| Consistency | 12 878,65 | 0,832% | 2,156% |
Of the three components, Performance showed the greatest stability. The blocks that use the tail of the latency distribution fluctuated more noticeably. Therefore, final scores that look identical do not yet guarantee an identical picture in terms of frame distributions.
Resolution check
Section titled “Resolution check”| Resolution | PC Score | Δ PC to the mean of No. 1–6 | Δ Performance | Δ Core Latency Score | Δ Consistency |
|---|---|---|---|---|---|
| 1280×768, No. 7 | 14 258,79 | +0,392% | −0,054% | +0,633% | +1,149% |
| 800×600, No. 8 | 14 193,24 | −0,070% | −0,089% | −0,099% | +0,022% |
Lowering the output resolution did not lead to a systematic increase in Performance. This fits the logic of a fixed internal load of 1920×1080.
If we compare only with the nearest 1920×1080 run No. 6 in the same Windows boot session, the PC Score changes are +0,832% and +0,369%. The difference depends on whether we compare with a single run or with the mean of several, so both variants are shown here.
At each low resolution there was only one run. The order was not alternated, and there was no repeat run at 1920×1080 after changing the resolution. The picture matches the hypothesis of a fixed internal load, but does not rule out the influence of resolution on individual components. Verification requires a repeat on several PCs with alternation of order.
Verification of calculations and available data
Section titled “Verification of calculations and available data”All eight runs completed: in the published CSV, all four scores and input aggregates are filled in for each run. Run suitability is assessed by the Run Quality model described in the methodology: it takes into account background CPU load, hypervisor, battery power and remote session. At the same time, external interference could still have affected the measurements: run suitability by this model does not prove its absence.
For all runs, Performance, Core Latency Score, Consistency and PC Score were recalculated from the saved throughput and percentiles. The maximum discrepancy with the publication is less than 0,000001 points. This is a verification of the formulas using aggregates; a full rebuild of percentiles from raw frames was not performed in this analysis.
- Cleaned data of the eight runs, CSV.
- Calculation verification script, Python 3. Only the standard library is required.
Save both files next to each other and run:
python analyze_repeatability.pyThe CSV contains the scores and input aggregates, resolutions, system uptime and boot session labels. The full original archive with identifiers and service data is not published.
Independent results
Section titled “Independent results”To verify repeatability on other computers, send your series to admin@boosterx.org with the subject “PCBenchmarkX: repeatability”.
Send several complete results under identical conditions and specify CPU, GPU, RAM, Windows and driver version, profile, resolution, reboots and changes between tests. If you are checking the influence of resolution, alternate it between runs and do several runs at each value. It is better to send the entire series, including “bad” or rare cases.
Before sending, check the archive for personal data. State whether the results and computer specifications can be published without your personal data. After verification, we will be able to add the series here with the conditions, calculations and limitations. Send series with a large spread as well: they will help clarify under which conditions the results are less stable.
This analysis was prepared by the BoosterX team based on the provided series. It is not an independent certification and does not prove a gain from BoosterX optimization. Independent reproductions will be marked separately.