How scores work
hwbench measures hardware and publishes the numbers. There are no editorial adjustments and no penalties or bonuses for any brand. Everything on this page comes from the scoring configuration in the open-source agent repository, so anyone can check it.
Reference machine
Intel Core i7-12700K, DDR4-3200 dual channel, Linux. 100 points = reference single-thread rate.
CPU
Five workloads (hash, compress, sort, float, physics) each run for 1.5 seconds on 1 thread, 4 threads, and every available thread. Each sub-score is the geometric mean of throughput relative to the reference CPU's single-thread throughput, times 100.
The overall CPU score weights single-core 40%, quad-core 30% and multi-core 30%. All three sub-scores appear next to it, so you can apply your own weighting.
Memory
Read, write and copy bandwidth over a buffer larger than any CPU cache, plus random-access latency. The score is the geometric mean of the four ratios against the reference.
Component scores
A component's score is the median of its verified results, never the mean, so a handful of broken or fake results can't move it much. The full distribution is shown next to it.
Keeping results honest
- Every CPU result includes proof-of-work digests for a one-time server challenge. The server re-runs a random sample of chunks and rejects any mismatch.
- Timing and consistency checks catch impossible results, for example multi-core scaling above the core count or timings longer than the run actually took.
- Statistical outliers are held for review and left out of averages until then.
- Runs where the OS limited available threads (VMs, containers) are excluded from multi-core averages.
Read the full write-up in METHODOLOGY.md.