Fully-engaged nh13/minibwa vs stock lh3/minibwa — same reads, same reference, same compiler:
| workload | arm64 (Graviton4) | x86 (AVX-512) |
|---|---|---|
| short reads (WGS / WES / panel) | up to 1.23× faster | up to 1.08× |
| long reads (HiFi / ONT) | up to 1.16× | up to 1.26× faster |
| Hi-C | up to 1.09× | up to 1.02× |
| ALT-shadowed reads placed confidently | 6.1× more than stock | 6.1× more than stock |
Whole-distribution defaults vs stock lh3/minibwa · hs38DH · -t 16, 2 reps · lh3 has no ALT handling · output byte-identical to upstream except where a feature is engaged (≥99.95% concordance) · measured 2026-09-06.