Hardware advice: RTX 5080 vs 5090 for 50k-100k atom protein systems

User discussions
1

GROMACS version: 2026.3 (planned — conda-forge CUDA build)
GROMACS modification: No

Building two workstations for protein MD at a small research
group (two users, one machine each). Would appreciate input from
people running comparable systems.

Workload

  • Explicit solvent protein systems, ~50k-100k atoms
  • CHARMM36 / AMBER force fields, 2 fs (4 fs with HMR)
  • Multi-day production runs
  • Mostly replicates rather than single long trajectories
  • Will also run ColabFold / Boltz for structure prediction later

Planned build (per machine)

  • GPU: RTX 5080 16GB (alternative: 5090 32GB)
  • CPU: Ryzen 9 9900X, 12 cores
  • RAM: 64GB DDR5
  • Storage: 2TB NVMe (DRAM cache) for runs + 1TB OS
  • PSU: 1000W 80+ Gold
  • OS: Ubuntu LTS

Questions

  1. Any ns/day figures you can share for 5080 or 5090 at this
    system size? VRAM doesn’t look like a bottleneck for MD here,
    but structure prediction may need more — trying to judge
    whether the price gap is justified.

  2. Is 12 CPU cores enough to keep one GPU fed with PME offloaded
    to GPU, or does the CPU become limiting at this ratio?

  3. Recommended starting point for offload flags
    (-nb / -pme / -bonded / -update) on this kind of hardware?

  4. Any known issues with recent consumer GPUs and current
    GROMACS releases on Linux?

  5. For sustained multi-day trajectory writes, anything to watch
    on the storage side beyond picking a DRAM-cache drive?

Happy to post benchmark numbers back once the machines are
running, in case it’s useful to others.

2

For large system the performance of GROMACS is nearly proportional to the flop rate for GPUs of the same generation. Memory size if irrelevant. An RTX 5090 will not reach peak performance on a system of 50k atoms. You would need to run multiple simulations simultaneously to reach maximum throughput.

I see that an 5090 currently is a factor 3-4 more expensive than an 5080. For GROMACS you get more performance with two 5080’s than with one 5090, and it will be cheaper as well.