Tools · reference

GPU memory requirements for biotech tools.

The README minimums are usually wrong. This table shows real production VRAM — what your jobs actually need to finish, not the toy example that fit on a developer’s laptop. Sourced from official docs, practitioner benchmarks, and our own verified runs.

Tool
Min VRAM
Recommended
Recommended GPUs
Notes

Why the README VRAM is usually wrong

Most biotech tool READMEs document the VRAM their toy example needs — the small monomer fold, the single-class 2D classification, the 1×1 docking pose. Production workloads are not toy examples. AlphaFold-Multimer’s official README says it “works on 16 GB” — in practice a 5-chain antibody complex routinely OOMs on a 16 GB T4 and needs at least an A10G (24 GB) to finish.

The numbers above reflect what works for actual production inputs: 3–5 chain complexes, 100K+ particles, 50K+ atom MD systems, real spectral libraries. Where we have first-hand data from a verified Clusterra run, that’s the source. Where we don’t, we cite the most reliable public reference we could find.

Scaling behavior matters as much as the minimum

Two tools with the same “recommended 16 GB” entry can have wildly different scaling. AlphaFold-Multimer scales quadratically with sequence length because of triangle attention — the difference between a 1,000-residue complex and a 2,000-residue complex is 4× the memory, not 2×. ESMFold scales linearly, so the same jump roughly doubles memory. The “Scales with” column is the part to read carefully.

Recommended GPU types are AWS-mapped

We recommend specific GPU types — T4, L4, A10G, L40S, A100, H100 — rather than just a VRAM number because compute capability matters too. AlphaFold runs ~3× faster on an L4 than on a T4 with the same 16 GB. Boltz-2 needs Ampere or newer (sm_75 is not enough). The recommendations bake in compute generation, not just memory.

All GPUs listed are available on AWS in us-east-1. The spot price comparator shows live pricing, typically 50–80% off on-demand.

VRAM tells you what fits; spot pricing tells you what it costs

Once you know VRAM, the next question is “which AWS instance, and what does it cost on spot?” For that, see our spot price comparator or workflow cost calculator. Or just look at the verified benchmarks — each row shows the actual instance, runtime, and dollar figure on a real public dataset.

All tools →