Repository navigation
Benchmark Results
Open the interactive benchmark dashboard
Compare 13 dense linear-algebra operations across 11 libraries at 64 × 64 and 256 × 256 matrix sizes. The 2026-10-18 run completed 286 measured cases, using two JVM forks per case and ten measured iteration means per case.
- Lower time ratios mean faster execution. NM Dev is the 1× reference.
- A point at 0.5× took half the NM Dev time; a point at 2× took twice the time.
- The horizontal axis is logarithmic. Equal distances indicate equal multiplicative changes.
- Hover, tap or keyboard-focus a point for its library version/backend, operation, matrix size, mean milliseconds, 99.9% confidence interval and ratio to NM Dev. Press Escape to dismiss the tooltip.
- Confidence intervals describe timing uncertainty. The displayed ratio is a ratio of means, without a confidence interval for that ratio.
The dashboard runs on GitHub Pages. The images below are static previews; open the dashboard for interactive tooltips.
NM Dev, EJML, Apache Commons Math, JAMA, Colt, Parallel Colt, jblas, F2J LAPACK, MTJ, ojAlgo and UJMP are included. F2J LAPACK is a replacement distribution for the unavailable original JLAPACK artifact, not the original JLAPACK binary. Exact versions, checksums and backend settings are documented in the report.
All 429 baseline correctness cases passed twice, and 66 targeted pivot/sign checks passed. Each measured fork also validates its selected result before timing. These checks cover the supplied inputs; they do not establish correctness for every possible matrix.
The run used a shared AMD EPYC VM, Java 17, one JMH worker, two forks, three 300 ms warmup iterations and five 500 ms measurement iterations per fork. Native jblas and Java implementations use different backend paths. Native worker counts were not independently audited. Results apply to this environment and these inputs; they are not universal performance rankings.