NVIDIA has released the BioNeMo Inference Runtime (BioIR), an open-source tool designed to accelerate protein structure prediction models. The runtime is available on GitHub and supports integration with standard PyTorch workflows without requiring additional compilation steps or engine builds.
The runtime optimizes specific operations, including Evoformer and Pairformer stacks, triangle operations, and diffusion transformers. Developers can use the end-to-end processor for tasks like tokenization and PDB writing or integrate optimized modules directly into custom PyTorch code.
Benchmarking on H100 GPUs shows speedups of 1.55x for OpenFold3, 1.78x for Boltz-2, and 2.56x for OpenFold2 monomer compared to torch.compile baselines. Similar performance gains were observed on H200 hardware across inputs ranging from 29 to 1,734 residues.
Tests on 1,000 human dimer targets using 8xH100 80GB GPUs demonstrated a 2.90x increase in Boltz-2 folding throughput. The system achieves a processing rate of 58.5K residues per GPU-hour, excluding preprocessing and MSA generation stages.
Projections suggest that using BioIR for 1 million targets could reduce energy consumption by 23 to 43 MWh compared to standard open-source implementations. These estimates are based on folding-stage power requirements for IT equipment.
