- Published on
NVIDIA Details BioNeMo Inference Runtime (BioIR): 2.90x Higher Boltz-2 Folding Throughput and 58.5K Residues per GPU-...
- Authors
- Name
- DeepHorus
NVIDIA has detailed BioNeMo Inference Runtime (BioIR), a Python library designed to accelerate biomolecular structure prediction on NVIDIA GPUs. The company reports 2.90x higher Boltz-2 folding throughput and 58.5K residues per GPU-hour on 8xH100 systems. This performance targets the growing bottleneck in proteome-scale worklists, where entire queues of independent targets must move through parsing, featurization, GPU inference, and output writing.
Biomolecular structure prediction has shifted from single-target runs to large-scale production workloads. The key challenge is no longer whether a model can fold a protein, but how quickly a full queue of targets can be processed. BioIR keeps the standard PyTorch workflow while optimizing operations that general-purpose inference stacks do not fully handle, including Pairformer and Evoformer stacks, triangle operations, pairwise attention, diffusion transformers, and atom-level modules.
BioIR has already operated at production scale. It powered the recent expansion of the AlphaFold Database, generating protein-complex structures across 4,777 proteomes, roughly 31 million candidate complexes, with 1.81 million released as high-confidence predictions. The runtime is available now as an open GitHub repository with a wheel containing precompiled CUBINs. Deployment requires Python 3.12+, a compatible NVIDIA GPU and driver, a staged model checkpoint, and per-chain A3M MSAs, but it does not require nvcc, CUDA source, CMake, or the CUDA toolkit.
Developers can use BioIR in two ways. The end-to-end processor moves an InputRequest through parsing, tokenization, feature generation, GPU inference, and PDB or mmCIF writing. Direct PyTorch integration lets developers construct a supported model or reuse selected optimized modules inside custom code. Models remain ordinary torch.nn.Module objects, with no engine build, export step, or separate artifact between checkpoint and forward pass.
The tutorial demonstrates the processor path with Boltz-2. Each protein chain requires an A3M MSA, and paired or unpaired MSAs are accepted for inputs with multiple non-identical chains. Templates must be supplied manually because BioIR does not run HHsearch or HMMsearch. The processor supports ligand structure prediction but not ligand-affinity prediction.
BioIR accelerates at three distinct layers. Kernel selection picks compatible BioIR custom, cuEquivariance, or PyTorch fallback implementations based on model configuration, GPU, data type, and tensor shape. Module optimization enables CUDA Graph capture for compatible modules, cutting launch overhead. Pipeline scaling uses a Ray executor to place one complete model replica on each visible GPU and distribute independent inputs among them, while CPU stages overlap with GPU folding.
Ray does not split a single forward pass across GPUs; replica mode scales worklists, not individual targets. Context-parallel folding is planned but not yet available. BioIR lowers the infrastructure barrier for large-scale structure prediction, and future context-parallel support could extend its reach even further.
