High-Throughput Structure Prediction with BioNeMo Inference Runtime
NVIDIA introduced the BioNeMo Inference Runtime (BioIR), a runtime for accelerating supported biomolecular structure-prediction models on NVIDIA GPUs while preserving a PyTorch workflow. The company said BioIR uses optimized kernels and, where applicable, CUDA Graphs to speed model execution, and can place a complete model replica on each GPU of a single node via Ray for large batches of independent inputs.
BioIR supports two usage paths: an end-to-end processor that handles parsing, tokenization, feature generation, GPU inference, and PDB or mmCIF writing, and direct PyTorch module integration for constructing a supported torch.nn.Module or reusing selected modules in custom code. NVIDIA said BioIR optimization operates at three layers: kernel selection, module optimization with CUDA Graph capture, and pipeline scaling through Ray replicas.
NVIDIA reported that BioIR has been used in proteome-scale work, including the expansion of the AlphaFold Database, accelerating protein-complex structure generation across 4,777 proteomes, about 31 million candidate complexes, with 1.81 million released as high-confidence predictions.
In a matched benchmark of 1,000 human dimer targets on 8xH100 GPUs, BioIR-accelerated Boltz-2 delivered 58.5K successfully folded residues per allocated GPU-hour versus 20.2K for a torch-compiled open-source implementation, a 2.90x improvement in residue-normalized throughput. NVIDIA said the open-source implementation ran out of memory on 29 targets. The workflow used three recycles, 200 sampling steps, and five diffusion samples per target.
NVIDIA estimated that folding one million comparable targets would require 11 MWh with BioIR versus 35 MWh for the public implementation using 8-GPU TDP equivalents, and 21 MWh versus 64 MWh using full-node maximum-power equivalents. It described these as folding-only estimates for IT equipment, not metered energy measurements, excluding data center overhead such as PUE.
NVIDIA said throughput metrics depend on the input dataset composition and should not be generalized to all BioIR-supported models, datasets, or hardware. The runtime is available through the BioNeMo Inference Runtime repository.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Biomolecular structure prediction is now often run at proteome scale, where the goal is to move an entire worklist through the pipeline efficiently. NVIDIA...