Is GPU or high-memory + CPU better for mmseqs2 server?

[seanrjohnson]

I'm looking into purchasing a dedicated machine to act as an mmseqs2 server for my organization.

What is the current recommendation for specifications? Is it faster to run in CPU mode with 1 TB memory, or on a recent GPU?

Here are the two options I'm considering:

• 1 TB RAM server with no GPU and run MMSeqs2 on CPU mode.
• 256 GB RAM, plus an NVIDIA RTX PRO 6000 GPU.

Other specs would be similar, i.e. some recent server CPU plus a large recent SSD.

Is either of these options clearly better than the other. Is there some other option I should prefer?

Thanks!

[martin-steinegger]

Below are three scenarios for searching a 200M-sequence database (avg. length 350 aa) with MMseqs2.

1. Fast CPU search (pre-indexed k-mer database; Steinegger & Söding 2017; Mirdita et al. 2019)
   Optimal for many concurrent users (e.g., our ColabFold server).
   Server: ~1 TB RAM to keep the k-mer index in memory
   Memory: 200M × 350 aa × 7 bytes/aa (+ overhead)
   Speed: seconds per average query
   Note: with ~64 cores, similar or higher than ColabFold (~80k requests/day). ColabFold internally performs six searches, so a single 200M-DB search should achieve ≥3x that throughput.

2. Fast GPU search (MMseqs2-GPU; https://www.nature.com/articles/s41592-025-02819-8)
   Optimised for the lowest latency.
   Server: 1× RTX 6000 PRO (96 GB), ~32 cores
   Memory: 200M × 350 aa × 1 byte/aa (GPU VRAM)
   Speed: milliseconds per query
   Note: CPU/RAM less relevant but RAM ≥ GPU VRAM recommended. I would expect a single card could process 50-60K searchers per day.

3. CPU-only ungapped-prefilter search
   Cheapest hardware option; similar to our current Foldseek setup.
   Server: ~128 GB RAM, ~64 cores
   Memory: 200M × 350 aa × 1 byte/aa
   Speed: ~1 minute per average query
   Note: Handles a few thousand requests/day; exact maximum untested

[seanrjohnson]

Thanks for the detailed breakdown. It helps a lot!