I tested 4 #bwa flavours on 2 WGS samples, same #biopype pipeline (markdup, BQSR, QC): Parabricks fq2bam (CUDA), minibwa, bwa-mem3 and good old bwa mem.
fq2bam is ~2x faster than bwa-mem3, which beats minibwa by 20 to 50%. Both still take a swift ~1/3 the time of classic bwa.
The surprise was energy: minibwa and bwa-mem3 use less than the GPU run!
Still diffing the BAMs, but I’m quietly dropping fq2bam from some workflows to see what happens.
Pipeline: https://codeberg.org/bio-pype/test_workflows/src/commit/main/bwas_comparison
#bioinformatics