Ten Standard runs on one Intel Core i5-1245U laptop on mains power, five minutes apart, all ten completed.
- Steady: CPU sustained throughput varied by 2.8% (coefficient of variation), the overall score by 3.4% and memory latency by 5.8%.
- Spread: one-thread memory bandwidth varied by 26.5%, and one run read far off the others on GPU frame rate and memory bandwidth for a reason not yet found.
- One machine, one browser, one power state. It says how this method behaves, not how every machine will.
The short answer
On the one laptop measured, the headline figures repeated within a few percent between runs, and some secondary figures did not. That is why SystemCheck states an interval on each figure, names the conditions of a run, and asks for a repeat before a small change is read as real.
How the runs were made
On 5 October 2026 a script drove ten Standard runs through the real site flow against a local production build: about three and a half minutes of run, then 90 seconds idle, five minutes apart. The machine was an Intel Core i5-1245U laptop with Intel Iris Xe graphics, in Google Chrome on Windows, at measurement revision 2026.09.14. The power source was read before each run, and all ten were on mains.
What repeated
Coefficient of variation is the standard deviation divided by the mean:
- CPU sustained throughput, all threads: 40.5 to 44.0 million operations a second, 2.8%.
- Overall score: 61 to 67, 3.4%.
- Memory latency: 158 to 182 nanoseconds, 5.8%.
- GPU sustained frame rate with one outlying run set aside: 141 to 162 frames a second, 3.9%.
What spread
One-thread memory bandwidth ran from 10.1 to 24.0 GB/s, 26.5%. Path tracing ran from 42.3 to 73.0 samples a second, 13.2%. Peak and lowest CPU levels spread more than the sustained level did.
One run read 265 frames a second on the GPU against 141 to 162 in the other nine, and low on memory bandwidth at the same time. Nothing in its quality notes marked it, and the reason has not been established. It is reported rather than dropped.
The CPU curve label also varied across these ten runs. That finding led to revision 2026.09.15, which separated the opening boost from decay; on the same ten recorded runs the new rule gave the same label in eight.
What to take from it
Treat one run as a reading, not a verdict. Compare runs made under the same conditions and the same revision, and repeat before reading anything into a difference of a few percent. These results come from one machine; a second machine, battery power and other browsers have not been measured this way yet.
Terms used here
Questions
- Is a browser benchmark accurate enough to compare two runs?
- On the laptop measured, the CPU sustained figure and the overall score repeated within a few percent. Compare runs under matching conditions, repeat each, and treat small differences as noise until a repeat confirms them.
- Why was the outlying run kept?
- Because nothing measured explained it. Dropping a run because it looks wrong would hide exactly the kind of spread a reader needs to know about.