Description of feature

GPU execution of MSA construction using MMSeqs has a 170x speedup compared to AlphaFold2's default JackHMMER execution.

https://developer.nvidia.com/blog/boost-alphafold2-protein-structure-prediction-with-gpu-accelerated-mmseqs2/

I've put GPU formatted MMSeqs DBs on Katana and this bares out on our system. The gpu-server mode takes up about 50GB of VRAM so is suitable for H200 use but not A100. Nevertheless the MMSeqs-GPU speedup is still dramatic even without an index GPU server, provided there's fast local storage.

The MMSeqs team are releasing CUDA12 compatible docker images, for T4 -> HX00 GPUs. Try to apptainer pull from that rather than roll a separate .sif for proteinfold, though some setups may need compilation with CUDA11 compatibility, or particular CPU vector extensions, etc
https://github.com/soedinglab/MMseqs2/pkgs/container/mmseqs2