Stability. Every answer, checked.
Every logical core computes answers known in advance, for 5 minutes or a quick 60 seconds. A processor that is stable gets every one right; one wrong answer fails the run.
Or run the 60-second quick check
Illustrative figures, not a measurement. The test reports your own.
Load
Every logical core
Duration
5 min, or 60 s quick
Checks
Integer, float, prime
Pass
No wrong answer
Answers known before the run.
A speed test asks how fast. A stability test asks whether the answers are right. Each worker cycles through three tasks whose answers are fixed by arithmetic, and every answer is checked against the one computed at the start on a different thread.
- Integer
- A fixed-seed xorshift128+ stream of 400,000 numbers, folded into a 64-bit checksum in two 32-bit halves. One flipped bit anywhere changes the checksum.
- Floating point
- A fixed series of 200,000 square roots and divisions summed in a fixed order in double precision. IEEE 754 rounds each of those operations exactly one way, so the sum is the same bits on every correct processor, and it is compared bit for bit.
- Prime
- The Lucas-Lehmer test of Mersenne numbers with big integers, the check Prime95 runs in spirit: 2^127 - 1, 2^521 - 1, 2^607 - 1 and 2^1279 - 1 are prime, 2^67 - 1 and 2^1277 - 1 are not, and the final residue is compared too.
- Checked on another thread
- The answers are computed once on the page's own thread at the start and also stored in the code; if those two disagree the run does not start. Each worker's answers are then checked against them, so a core that gets one wrong shows as a disagreement, with the worker, the task and the seed.
Pass or fail, and what failed.
The verdict is the errors. Throughput is a secondary curve, there to show the load held.
Verdict
No wrong answer passes. Any wrong answer fails, because a correct processor never gets one of these wrong: there is no tolerance to set.
Errors
Counted per worker and per task. The first mismatches are listed with the task, the seed, the expected answer and the one the worker got.
Checks
How many answers were computed and checked, per task and in all.
Per thread
Checks per second per thread, one sample a second, drawn over the run. The median is reported. A secondary figure: a falling curve shows throttling, not an error.
Illustrative figures, not a measurement. The test reports your own.
Hardware
What it needs. Three tiers, no fine print.
Minimum
Any recent browser
Chrome, Edge, Firefox or Safari with Web Workers and BigInt.
Recommended
Plugged in, other apps closed
The test checks correctness, but on battery the load is lighter and an instability may not show.
Optimal
The 5-minute run, warm
Heat-related errors take time to appear. The quick 60-second run is a smoke test, and no run here is shorter than 30 seconds.
Questions
CPU stability, answered
- What does a CPU stability test check?
- Whether a processor computes the right answer under sustained load on every core. This one runs an integer checksum, a floating-point sum and a Lucas-Lehmer prime test whose answers are known, and checks every result.
- What does a failed result mean?
- That at least one answer was wrong. A correct, stable processor never gets one wrong, so a mismatch points to an instability. Possible causes include an overclock or undervolt set too aggressively, heat, or failing hardware such as the processor, its power delivery or memory. The test cannot say which.
- Is a pass here the same as passing Prime95 or y-cruncher?
- No. A browser cannot reach AVX-512 or the heaviest vector paths a native torture test uses, so it draws less power and makes less heat. A pass here is not a pass of Prime95's small FFTs. For a full stress test, use Prime95 or y-cruncher natively.
- Why does my browser say which worker failed, but not which core?
- A browser runs workers as threads and does not say which core ran each one. The worker number tells you that one thread disagreed; it does not name a physical core.
- Is this part of the main SystemCheck run?
- No. It runs on this page only, on its own, and it does not change the score or the report.
- Why 5 minutes, and why a 60-second option?
- Errors from heat or marginal voltage tend to appear as the chip warms, so the default runs 5 minutes. The quick option is a smoke test. Neither length is a standard; they are our line, and no test here runs under 30 seconds.