Model benchmarks: speed, GPU memory and cost per run
These are real measurements from moldesk run on rented NVIDIA GPU hardware, so you can estimate how long a job will take and what it will cost before you start. Every model page shows the same table, and this page collects them all.
How to read these tables
- Cold is the first run after install. It includes one-time work such as caching and kernel compilation.
- Warm is the typical time for later runs. Use this for planning.
- Cost per run is time multiplied by the GPU-hour price. If no price is shown, use
seconds ÷ 3600 × your provider's hourly price. - Peak GPU memory tells you whether a model fits your card. A job that needs more memory than your GPU has will fail, so check this first.
- Times depend on input size. Each table states the exact workload that was measured.
Boltz-2 benchmarks: speed, VRAM and cost per prediction
On an NVIDIA GeForce RTX 3090, Boltz-2 takes about 85 s per prediction once warm (2.3 min on the first run), roughly 42 predictions per GPU-hour, or about $0.012 per prediction.
| Metric | NVIDIA GeForce RTX 3090 |
|---|---|
| Time per run, first (cold) | 2.3 min |
| Time per run, warm | 85 s |
| Predictions per GPU-hour | 42 |
| Cost per run, first (cold) | $0.019 |
| Cost per run, warm | $0.012 |
| Peak GPU memory | 2.6 GiB |
| Peak GPU utilization | 60 % |
| Peak GPU power | 163 W |
| Install time | — |
| Disk per install | 17 GiB |
Workload: Single protein chain, 115 residues, no MSA (examples/boltz/protein.yaml), 1 sample. Median of 3 runs.
Machine: RunPod GPU pod, Linux x64, 125 GiB RAM; AMD EPYC 7H12 (32 vCPU allocated); driver 580.126.20.
Cost: at $0.50/GPU-hour (RunPod Secure Cloud, EU-CZ-1, compute only; storage adds about $0.02/hr); excludes storage and idle time.
Warm time is the mean of runs 2 and 3 (86 s and 84 s). Very first run ever also downloads the CCD data (~198 s total).
ProteinMPNN benchmarks: speed, VRAM and cost per design batch
On an NVIDIA GeForce RTX 3090, ProteinMPNN takes about 8 s per design batch once warm (16 s on the first run), roughly 450 design batchs per GPU-hour, or about <$0.01 per design batch.
| Metric | NVIDIA GeForce RTX 3090 |
|---|---|
| Time per run, first (cold) | 16 s |
| Time per run, warm | 8 s |
| Design batchs per GPU-hour | 450 |
| Cost per run, first (cold) | <$0.01 |
| Cost per run, warm | <$0.01 |
| Peak GPU memory | 0.4 GiB |
| Peak GPU utilization | 9 % |
| Peak GPU power | 108 W |
| Install time | 6.1 min |
| Disk per install | 6 GiB |
Workload: One 20-residue chain (examples/diffdock/protein.pdb), default sampling, 1 sequence. Median of 3 runs.
Machine: RunPod GPU pod, Linux x64, 125 GiB RAM; AMD EPYC 7H12 (32 vCPU allocated); driver 580.126.20.
Cost: at $0.50/GPU-hour (RunPod Secure Cloud, EU-CZ-1, compute only; storage adds about $0.02/hr); excludes storage and idle time.
Warm time is the mean of runs 2 and 3 (7.9 s and 8.1 s). Install time includes downloading PyTorch with an empty package cache. Peak GPU power is near the idle-to-light-load range because the job is tiny.
LigandMPNN benchmarks: speed, VRAM and cost per design batch
On an NVIDIA GeForce RTX 3090, LigandMPNN takes about 10 s per design batch once warm (15 s on the first run), roughly 360 design batchs per GPU-hour, or about <$0.01 per design batch.
| Metric | NVIDIA GeForce RTX 3090 |
|---|---|
| Time per run, first (cold) | 15 s |
| Time per run, warm | 10 s |
| Design batchs per GPU-hour | 360 |
| Cost per run, first (cold) | <$0.01 |
| Cost per run, warm | <$0.01 |
| Peak GPU memory | 0.4 GiB |
| Peak GPU utilization | 10 % |
| Peak GPU power | 108 W |
| Install time | 59 s |
| Disk per install | 7 GiB |
Workload: One 20-residue chain (examples/diffdock/protein.pdb), default sampling, 1 sequence. Median of 3 runs.
Machine: RunPod GPU pod, Linux x64, 125 GiB RAM; AMD EPYC 7H12 (32 vCPU allocated); driver 580.126.20.
Cost: at $0.50/GPU-hour (RunPod Secure Cloud, EU-CZ-1, compute only; storage adds about $0.02/hr); excludes storage and idle time.
Warm time is the mean of runs 2 and 3 (9.7 s and 9.9 s). Install time is with PyTorch already in the package cache; expect several minutes on a first install.
DiffDock-L benchmarks
DiffDock-L is coming soon. Pace and cost numbers will be published here when it ships.
OpenDDE Preview benchmarks: speed, VRAM and cost per prediction
On an NVIDIA GeForce RTX 3090, OpenDDE Preview takes about 57 s per prediction once warm (88 s on the first run), roughly 63 predictions per GPU-hour, or about <$0.01 per prediction.
| Metric | NVIDIA GeForce RTX 3090 |
|---|---|
| Time per run, first (cold) | 88 s |
| Time per run, warm | 57 s |
| Predictions per GPU-hour | 63 |
| Cost per run, first (cold) | $0.012 |
| Cost per run, warm | <$0.01 |
| Peak GPU memory | 5.6 GiB |
| Peak GPU utilization | 28 % |
| Peak GPU power | 124 W |
| Install time | — |
| Disk per install | 12 GiB |
Workload: 9-residue peptide ACDEFGHIK (examples/opendde/tiny.json), default settings: 1 sample, 200 steps, 10 cycles, no MSA. Median of 3 runs.
Machine: RunPod GPU pod, Linux x64, 125 GiB RAM; AMD EPYC 7H12 (32 vCPU allocated); driver 580.126.20.
Cost: at $0.50/GPU-hour (RunPod Secure Cloud, EU-CZ-1, compute only; storage adds about $0.02/hr); excludes storage and idle time.
Warm time is the mean of runs 2 and 3 (57.4 s and 57.7 s). Larger complexes need more time and memory than this smoke-test input.
BindCraft2 benchmarks: speed, VRAM and cost per design campaign
On an NVIDIA GeForce RTX 3090, BindCraft2 takes about 3.8 min per design campaign once warm (2.5 min on the first run), roughly 15 design campaigns per GPU-hour, or about $0.032 per design campaign.
| Metric | NVIDIA GeForce RTX 3090 |
|---|---|
| Time per run, first (cold) | 2.5 min |
| Time per run, warm | 3.8 min |
| Design campaigns per GPU-hour | 15 |
| Cost per run, first (cold) | $0.021 |
| Cost per run, warm | $0.032 |
| Peak GPU memory | 4.5 GiB |
| Peak GPU utilization | 100 % |
| Peak GPU power | 303 W |
| Install time | — |
| Disk per install | 13 GiB |
Workload: Bounded campaign (examples/bindcraft2/campaign.json): 50-residue binder against a short peptide target, 1 trajectory, 1 final design, seed 101. Median of 3 runs.
Machine: RunPod GPU pod, Linux x64, 125 GiB RAM; AMD EPYC 7H12 (32 vCPU allocated); driver 580.126.20.
Cost: at $0.50/GPU-hour (RunPod Secure Cloud, EU-CZ-1, compute only; storage adds about $0.02/hr); excludes storage and idle time.
Runs took 152 s, 232 s and 229 s: design search is stochastic, so time varies run to run and there is no cold-start penalty. The 'warm' figure is the mean of runs 2 and 3. Real campaigns with more trajectories scale roughly linearly with trajectory count.