Nvidia library triples Boltz-2 protein folding throughput on eight H100s
single source · 1 articles · capability · confidence: high · first seen 2026-09-10 21:57 UTC
NVIDIA has detailed BioNeMo Inference Runtime (BioIR), a Python library that accelerates biomolecular structure prediction on NVIDIA GPUs while keeping models in plain PyTorch. In a benchmark on 1,000 human dimer targets across eight H100 GPUs, BioIR running Boltz-2 folded 58,500 residues per GPU-hour, versus 20,200 for a torch-compiled open-source baseline, a 2.9x increase. The runtime uses custom kernel selection, CUDA Graph capture, and Ray-based scaling that places one full model copy per GPU. BioIR already powered the AlphaFold Database expansion, generating about 31 million candidate protein complexes across 4,777 proteomes.
What this means for you
If you run Boltz-2 or similar protein-folding inference on H100 clusters, evaluate BioIR: the benchmark reports 2.9x more residues per GPU-hour with plain PyTorch. No download, licence or API details are given, so availability remains unclear.
Key facts
- ·BioIR is a Python library that accelerates biomolecular structure-prediction models on NVIDIA GPUs while staying in plain PyTorch. source
- ·In a benchmark on 1,000 human dimer targets across eight H100 GPUs, BioIR-accelerated Boltz-2 delivered 58,500 successfully folded residues per GPU-hour. source
- ·The torch-compiled open-source baseline delivered 20,200 residues per GPU-hour, giving a 2.90x throughput gain. source
- ·BioIR optimises at three layers: custom kernel selection, CUDA Graph capture, and Ray-based replica scaling with one full model copy per GPU. source
- ·BioIR powered the AlphaFold Database expansion, generating about 31 million candidate protein complexes across 4,777 proteomes. source
What the sources say
- MarkTechPost — Details the BioIR benchmark, throughput gain, and its role in AlphaFold Database expansion.
Sources
The original reporting. Follow these — they did the work.