NVIDIA says its BioNeMo Inference Runtime folded biomolecular structures at 2.90 times the throughput of a torch-compiled open-source implementation in a matched benchmark on eight H100 GPUs. The result could make large-scale protein prediction more practical, although it is a vendor benchmark for one model, workload and hardware setup.
NVIDIA Watch analysis
What happened
The runtime supports protein-structure prediction models and moves inputs through parsing, tokenisation, feature generation, GPU inference and PDB or mmCIF output while retaining a PyTorch workflow. For larger worklists, its Ray executor places a complete model replica on each GPU and processes independent inputs concurrently.
In NVIDIA's test, 1,000 human dimer targets were run with Boltz-2 on 8xH100 GPUs. BioNeMo delivered 58.5K successfully folded residues per allocated GPU-hour, compared with 20.2K for the compared implementation. NVIDIA also estimates 11 MWh to fold one million comparable targets with BioIR, versus 35 MWh under its stated rated-power-equivalent assumptions.
Key findings
- 2.90× reported throughput
BioIR reached 58.5K successfully folded residues per allocated GPU-hour, versus 20.2K for the comparison implementation. - Worklist-scale execution
Ray runs complete model replicas across GPUs, allowing independent prediction jobs to run concurrently. - A familiar developer path
The runtime offers an end-to-end processor and direct PyTorch integration, with PDB or mmCIF output. - 11 MWh estimate
NVIDIA estimates lower rated-power-equivalent energy for one million comparable targets, not a full deployment energy audit.
Why it matters
Protein-structure prediction is increasingly a worklist problem. Faster model execution and better pipeline overlap could make proteome-scale processing more accessible to research teams with suitable NVIDIA hardware.
The boundary matters: the benchmark measures the folding stage on one recipe and configuration. It does not establish end-to-end gains for every model, dataset or cluster, and Ray does not split one model forward pass across multiple GPUs.
Our read
This is substantial AI infrastructure for teams already running supported structure-prediction models on NVIDIA hardware. Treat the 2.90× figure as a starting point, then benchmark a representative worklist on your own pipeline.
What to watch
- Reproduction across different proteins and model versions.
- End-to-end throughput including MSAs, feature generation, storage and output writing.
- Which supported models and GPU generations receive the strongest optimisation.
- Whether the energy estimate survives real-world utilisation and cooling overheads.
Discussion spark: For protein-structure workloads, would you prioritise the reported model-forward speedup, Ray worklist scaling or a reproducible end-to-end benchmark?
Sources and evidence
- High-Throughput Structure Prediction with BioNeMo Inference Runtime (10 September 2026, 15:00 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.