テスト本体、レポート、およびサイトのその他のページは、現時点では英語です。
診断
どれだけ落ちたかではなく、なぜ落ちたか。 曲線の形そのものが診断であり、レポートはそれを読み解きます。
スコアはマシンが遅くなったことを示すだけです。その裏にある曲線が、何が起きたかを語ります。重要な形は3つ、それぞれ原因が異なります。
A single step down
A power budget expiring. The machine held its boost for as long as it was allowed to, then dropped to the sustained limit in one move.
A steady slide
Cooling falling behind. Every second under load, a little more is lost, and it does not settle until the fan can keep pace.
A sawtooth
A machine cycling between the two: it cools, boosts, hits the limit, drops, and repeats.
4つのステージ、そして起きたことに名前を。 そのマシンは仕様書どおりの性能を維持できているか。できていないなら、なぜか。
Stage 1
Every core, for 60 seconds. The number that matters is the one it settles at.
Parallel Web Workers saturate every logical core and the main thread samples throughput five times a second. Peak arrives in the first second; the rate the machine holds once cooling becomes the limit is what gets reported.
Illustrative run. Your own cores draw their own.
Stage 2
One thread, then six. How much more the memory system gives when asked.
Workers stream 24 MiB arrays to measure bandwidth from one thread and then from all of them, and a pointer chase through a shuffled array measures how long a round trip to memory takes. A machine whose bandwidth barely scales with threads is telling you about its memory channels, not its cores.
One thread
Six threads
Illustrative figures. Not a measurement.
Stage 3
A raster scene tuned to the card. Frame time, pacing, and how far it falls.
A lit scene renders on a slow orbit while a controller ramps the supersampling factor against a 33ms budget and records the ceiling the card can hold. Stutter and slowness look different here, which is the point.
Illustrative frame times. Not a measurement.
Stage 4
Timed per sample, on a fixed clock. So the cost of each sample is the measurement.
A WebGL2 path tracer accumulates one jittered sample per frame into a float texture. Samples per second and delivery consistency are reported; a card that delivers unevenly is distinguishable from one that is merely slow.
Illustrative progress. Not a measurement.
Report
Not how much it slowed. Why. The shape of the curve is the diagnosis.
A plateau that steps down once is a power budget expiring. A steady slide is cooling falling behind. A sawtooth is a machine cycling between the two. The report reads the shape, states its confidence, and names the cause with the number to beat on the retest. The next run on the same machine is shown against this one.
Plateau, then a single step down.
Throughput held for 18 seconds, then fell in one transition and stayed there. One fall accounts for 98% of everything the run gave up.
Illustrative shape. Your run draws its own, and the report reads that one.
限界を、はっきりと
ブラウザに見えないもの。
どのウェブページも温度は読み取れません。それでも冷却に制限されたマシンは、性能の低下という形でそれを示します。低下の測定にセンサーは要りません。
説明用の形です。あるマシンはピークの約96%を維持し、別のマシンは64%で落ち着きます。
ハードウェアを限界まで使う人たちに。 自作派、システム構築者、修理技術者。
“I was about to buy a second-hand ThinkPad for 400 quid. Ran the test in the seller's kitchen, watched the CPU stage lose 38 percent in three minutes, and walked away. Best three minutes I spent that week.”
Marcus R. · Buying a used laptop“I hand out one link to clients now instead of a download. They run it, forward me the report, and I know before I drive over whether the cooling is the problem.”
Priya S. · System integrator“My frames were fine for ten minutes and then fell off a cliff every night. The graphics stage showed the same drop at the same point, and the report said thermal ceiling in plain words. New paste, problem gone.”
Devon L. · Gamer, fixed his FPS drops“Exports kept getting slower the longer I worked. I assumed it was the software. The memory stage showed the laptop was fine on mains and lost a third of its speed on battery. Now I edit plugged in.”
Ana M. · Video editor“It says up front what a browser cannot see, and then measures the rest properly. That honesty is why I trust the number it gives me over a spec sheet.”
Tom K. · Sceptical by trade“Forty refurbished desktops in an afternoon, one link each, same test on every machine. The three that graded D all had dust-packed fans. The report told me before the screwdriver did.”
Jordan P. · IT refurbisher“Watching ops per second sink live while the fans spun up taught me more about my machine than any single score ever did. The retest target told me exactly what to aim for after the fix.”
Sana D. · Curious about her own PC“New build, new cooler, and I wanted proof it was worth it. Before and after in two runs: decay went from 31 percent to 6. That is the number I put in the forum post.”
Leo V. · First-time PC builder“Twelve laptops in the classroom, three complaints a week that they were slow. Ran the test on all of them from the same link. Two were throttling badly, ten were fine. Two repairs instead of twelve.”
Ravi N. · School IT admin“I asked in a forum why my new PC felt slower than the reviews said. Someone sent this. Three minutes later I could see it was hitting a power limit under load, and I knew what to change in the BIOS.”
Emily W. · New PC underperforming“Average FPS looked great and the racing still felt wrong. The consistency figure caught frame pacing my overlay was hiding. Once I could see it, I could fix it.”
Chris B. · Sim racer“Customers say slow, I need a number. I run it in the browser while they watch, the report says what is happening in plain words, and the repair conversation is over in a minute.”
Nadia F. · Repair shop ownerすべてのパーツを、仕様とともに。 そして仕様が語らないことも。
1,671 processors and every graphics card we can source properly, with side-by-side comparisons. 101 carry a full sheet today; the rest say “not verified” rather than guess.
記事
読むほどに深く。 正直に測ることについてのガイドと覚え書き。
- Thermal decay, explainedDecay is the percentage a system's throughput falls from its early peak to its sustained plateau: what it reveals, what it hides, and how to read it.
- What browser benchmarks can and cannot measureThe capability boundary of a web-based benchmark, stated plainly: what the platform exposes, what it refuses to, and where a native tool is the right answer.
- Why peak benchmarks misleadPeak scores measure the best few seconds of a machine's life under ideal conditions. Here is why the industry optimises for them and what it costs you.
- How accurate is a browser benchmark, really?An honest account of what a browser-based benchmark can measure reliably, what it cannot measure at all, and when you should use a native tool instead.