New GPU-based MMseqs2: 20x faster searches on a single L40S (approx. as fast as a RTX 4090) vs. a 128-core CPU. This work enables to set up a very cost-efficient ColabFold MSA GPU server. 🧵🧵 📄 www.biorxiv.org/content/10.1... 💾 mmseqs.com 🗞️ developer.nvidia.com/blog/boost-a... - ThreadSky

martinsteinegger.bsky.social • 225 days ago

New GPU-based MMseqs2: 20x faster searches on a single L40S (approx. as fast as a RTX 4090) vs. a 128-core CPU. This work enables to set up a very cost-efficient ColabFold MSA GPU server. 🧵🧵
📄 https://www.biorxiv.org/content/10.1101/2024.11.13.623350v1
💾 https://mmseqs.com
🗞️ https://developer.nvidia.com/blog/boost-alphafold2-protein-structure-prediction-with-gpu-accelerated-mmseqs2

Comments

martinsteinegger.bsky.social•225 days ago

Unlike most high-speed aligners that focus on multi-query performance, MMseqs2-GPU parallelizes across a single query. Despite this, on an L40S GPU, it is the most cost-effective and energy-efficient option, saving up to 71x in costs and outperforming a 128-core CPU in speed. 2/n

martinsteinegger.bsky.social•225 days ago

ColabFold (MMseqs2-GPU) achieves similar accuracy to AlphaFold2 (JackHMMER+HHblits) but is up to 23x faster! Real-time homology based structure prediction is becoming more of a reality.. 3/n

martinsteinegger.bsky.social•225 days ago

The GPU backend not only boosts MMseqs2 but will also accelerate all tools that rely on it, like Foldseek and FoldMason. It achieves even larger speed gains over traditional CPU setups, further enhancing performance across workflows. Stay tuned! 4/n

martinsteinegger.bsky.social•225 days ago

Great work by Felix Kallenborn, Alejandro Chacon, Christian Hundt, Hassan Sirelkhatim, Kieran Didi, Christian Dallago, @milot.bsky.social, and Bertil Schmidt. 5/n

bioinfo.se•225 days ago

For us it is about 5 times slower on GPU than CPU when we benchmarked it. But that is because we do not load the DB to memory (we do not have enough) and we do not use maximum sensitivity.

martinsteinegger.bsky.social•225 days ago

Yes this search is only faster in a high sensitivity setting (-s 8.5), this is what we use for ColabFold. Regarding memory, you can search through clustered dbs, this requires much less GPU RAM.

maxfus.bsky.social•225 days ago

Ohh this is good to know, please make a note in the docs

milot.bsky.social•225 days ago

What hardware and database is did you use? Was the database on fast local storage or a network share? What CPU setting did you compare against? 5x slower sounds like too much.
You can also cluster the database and do a colabfold style search to reduce memory requirements.

bioinfo.se•224 days ago

Nvidia A100 cluster

bioinfo.se•224 days ago

Details here https://www.nsc.liu.se/systems/berzelius/ CPU took 0.16 seconds per sequence using your script

milot.bsky.social•224 days ago

Thanks a lot! We will investigate

Comments

Posting Rules

Reply