Best GPU for local AlphaFold

[David458-alt]

Hello,

I’m planning a workstation to run AlphaFold locally and need advice on GPU choice.

My use case: Prediction of ~10,000 proteins (750–1500 aa)

I’m considering:

RTX 4090 (24 GB)
RTX 5090 (~32 GB)
RTX 6000 Ada (48 GB)

Questions:

Any experience with these GPUs for AlphaFold or other models like PLACER, RFDiffussion and LigandMPNN?
Better to go with one RTX 6000 or multiple 4090s?

Thanks!

[AravindKurapati]

I can't compare those three cards directly since I haven't run them all, but there is one factor worth checking before you buy: for a run of 10,000 proteins, the GPU probably isn't your main bottleneck.

AlphaFold splits each prediction into two stages, and the GPU only handles the second half:

• The MSA search (the actual bottleneck). Running jackhmmer and HHblits against the databases is bound by CPU and storage speed and never touches the GPU. For most sequences this takes the lion's share of the timeline.
• The model forward pass. The only part that uses the GPU. Since it is the faster half, a top-tier card only shrinks the smaller portion of the wait.

For 10,000 proteins, two specs matter more:

• Fast local NVMe storage. The full database set is about 2.62 TB unpacked (BFD alone is ~1.8 TB). On spinning disks or network storage the search stage crawls. The upstream README recommends an SSD for exactly this reason.
• More CPU cores and RAM. You will spend most of the time waiting on the CPU-bound MSA tools, and extra cores let you run predictions in parallel.

On VRAM: it dictates protein size rather than speed. 24 GB is plenty for standard 750 to 1500 aa monomers. More VRAM buys headroom for very long chains or multimers.

A workflow tip: --use_precomputed_msas lets you run the heavy MSA searches once and reuse them, so a re-fold never repeats the expensive search phase.