Benchmarking Pearl miners: our test plan
Before we publish a single miner number, here's exactly how we measure, so when the numbers land you can hold us to it.
Mining benchmarks have a credibility problem: peak numbers from the first minute of a run, GPU sensor power instead of wall power, mystery driver versions. Before this section fills up with results, I want the methodology on the record. Partly so you can trust the numbers, partly so you can call me out if a result doesn't follow it.
What we measure
- Sustained hashrate after warm-up. The number a miner prints at startup is marketing.
- Wall power for the whole system, from a meter at the socket. GPU sensor readings miss the CPU, fans, and PSU losses you're actually paying for.
- Efficiency, meaning hashes per joule. This is the number that decides whether a card makes money, and it's the one most write-ups skip.
- Exact versions: miner build, driver, CUDA runtime. If any of those changes, it's a different benchmark.
How a run works
Stock settings first, always. That's the number most people will actually get. Then documented tuning: power limits, clocks, whatever's reproducible from the notes. Runs go long enough for thermals to fully settle, on the same host hardware every time. And the self-reported hashrate gets cross-checked against shares and blocks actually produced, because a miner that says 400 and submits like 350 is a 350 miner.
What’s first on the bench
The RTX 5090, since that's the card I've been porting and tuning the Pearl miner on. The numbers exist; they go up once they survive the methodology above, and not before. If there's a card or a whole rig you want put through this, tell me. Reply to the newsletter or ping us through the About page.
