NVIDIA reports faster biomolecular folding with BioIR
NVIDIA has detailed BioNeMo Inference Runtime (BioIR), a Python library designed to speed up supported biomolecular structure-prediction models on NVIDIA GPUs while retaining standard PyTorch workflows. The library targets operations including Pairformer and Evoformer stacks, triangle operations, pairwise attention, diffusion transformers and atom-level modules. It can process complete worklists or provide optimized modules for custom PyTorch applications.
In a matched test involving 1,000 human dimer targets on 8xH100 80GB GPUs, BioIR-accelerated Boltz-2 folded all targets and processed 58.5K successfully folded residues per allocated GPU-hour. A torch-compiled open-source implementation processed 20.2K residues per GPU-hour and ran out of memory on 29 targets, producing a 2.90x residue-normalized throughput advantage for BioIR. The measurements covered folding only and excluded MSA generation, preprocessing, storage, transfers and retries.
BioIR uses GPU-specific kernel selection, CUDA Graph optimization and Ray-based replication across GPUs; it does not split one forward pass between GPUs. NVIDIA also estimated that one million comparable targets would require 11 MWh rather than 35 MWh under an 8-GPU TDP-equivalent calculation. The runtime is available as an open GitHub repository and requires Python 3.12+, a compatible NVIDIA GPU and staged model inputs.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.