I see a few different bwa-mem2 implementations for alignment of Illumina short reads. Do people have thoughts and recommendations for which they prefer (or not)?

Generally, my use case is

• CPU-first by default
• GPU optional
• preference for something stable and easy to support in workflows
• speed matters, but not at the expense of weird RAM requirements—or fragile or unexpected behavior, or behavior that deviates from the original implementation

The options I’ve been looking at are the mainline bwa-mem2 from Intel, BWA-MEME, variants such as bwa-mem2-lisa and bwa-mem2-ert, and possibly Parabricks on the GPU side.

For folks who’ve tried any of these in real workflows, what do you think? Any recs (and why)?