NVIDIA BioIR boosts Boltz-2 protein folding throughput by 2.9x on H100 GPUs
NVIDIA has detailed BioNeMo Inference Runtime (BioIR), a Python library designed to accelerate biomolecular structure prediction on NVIDIA GPUs. By optimizing kernel selection, module execution via CUDA Graphs, and pipeline scaling with Ray, BioIR maintains standard PyTorch workflows while significantly improving efficiency. The tool is already deployed at scale, having powered the expansion of…
Key points
- BioIR achieves 2.90x higher throughput for Boltz-2 on 8xH100 GPUs compared to standard PyTorch implementations.
- The runtime processed 31 million candidate complexes for the AlphaFold Database expansion, releasing 1.81 million high-confidence predictions.
- NVIDIA estimates BioIR reduces energy usage for one million targets from 35 MWh to 11 MWh, significantly lowering computational costs.
In a matched benchmark using 8x H100 GPUs, BioIR-accelerated Boltz-2 achieved 58,500 folded residues per GPU-hour, compared to 20,200 for a standard torch-compiled implementation. This represents a 2.90x increase in throughput and eliminated out-of-memory errors encountered in the baseline run. NVIDIA estimates that scaling this performance to one million targets could reduce energy consumption from 35 MWh to 11 MWh, offering substantial savings for large-scale proteome campaigns.
BioIR is available as an open-source GitHub repository, requiring Python 3.12+ and compatible NVIDIA drivers. It supports models like Boltz-2, OpenFold2, and OpenFold3, allowing developers to integrate optimized modules directly into custom code or use the end-to-end processor for full pipeline automation.
The story so far
2 episodes →- NVIDIA BioIR boosts Boltz-2 protein folding throughput by 2.9x on H100 GPUsthis story
NVIDIA Details BioNeMo Inference Runtime (BioIR): 2.90x Higher Boltz-2 Folding Throughput and 58.5K Residues per GPU-Hour on 8xH100
MarkTechPost · 10 September 2026Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- Apple delays HomeView EU launch due to Siri’s unavailability · 2 src
- Apple unveils HomeView demo with Siri and facial recognition for up to six users · 2 src
- Microsoft expands Azure Maia system to multi-gigawatt deployments starting in 2027 · 3 src
- Apple adds Siri AI and Wi-Fi 7 to fourth-gen Apple TV 4K for $299 · 1 src
- AI lab replaces scheduler to cut GPU squatting and on-call toil · 1 src
Comments
via GitHub Discussions