|

NVIDIA Details BioNeMo Inference Runtime (BioIR): 2.90x Higher Boltz-2 Folding Throughput and 58.5K Residues per GPU-Hour on 8xH100

ByRicardo

Biomolecular construction prediction has shifted from single-target runs to proteome-scale worklists. The bottleneck is not whether or not a mannequin can fold a protein. It is how briskly a complete queue of unbiased targets strikes by parsing, featurization, GPU inference, and output writing. NVIDIA’s new technical deep dive walks by BioNeMo Inference Runtime (BioIR), a Python library that accelerates supported structure-prediction fashions on NVIDIA GPUs whereas protecting the usual PyTorch workflow. BioIR has already run at manufacturing scale. It powered the recent expansion of the AlphaFold Database, producing protein-complex constructions throughout 4,777 proteomes, about 31 million candidate complexes, with 1.81 million launched as high-confidence predictions.

Is it deployable? Yes. BioIR is accessible now as an open GitHub repository with a wheel containing precompiled CUBINs. Runtime use wants Python 3.12+, a suitable NVIDIA GPU and driver, a staged mannequin checkpoint, and per-chain A3M MSAs. It doesn’t require nvcc, CUDA supply, CMake, or the CUDA toolkit.

What is BioIR

BioIR targets the operations that general-purpose inference stacks don’t absolutely optimize. These embrace Pairformer and Evoformer stacks, triangle operations, pairwise consideration, diffusion transformers, and atom-level modules. Models keep abnormal torch.nn.Module objects. There is not any engine construct, export step, or separate artifact between a checkpoint and a ahead go.

There are 2 methods to make use of it. The end-to-end processor strikes an InputRequest by parsing, tokenization, characteristic technology, GPU inference, and PDB or mmCIF writing. Direct PyTorch integration lets builders assemble a supported mannequin or reuse chosen optimized modules inside customized code. The tutorial demonstrates the processor path with Boltz-2 (model_source="boltz-2"). Each protein chain requires an A3M MSA. Paired or unpaired MSAs are accepted for inputs with a number of non-identical protein chains. Templates might be provided manually as a result of BioIR doesn’t run HHsearch or HMMsearch. The processor helps ligand construction prediction however not ligand-affinity prediction.

Three Layers of Acceleration

BioIR optimizes at 3 distinct layers, every focusing on a unique bottleneck:

  1. Kernel choice: Supported operations choose suitable BioIR customized, cuEquivariance, or PyTorch fallback implementations based mostly on mannequin configuration, GPU, information kind, and tensor form.
  2. Module optimization: A separate optimize() mechanism allows CUDA Graph seize for suitable modules, chopping launch overhead.
  3. Pipeline scaling: A Ray executor locations 1 full mannequin reproduction on every seen GPU in a node and distributes unbiased inputs amongst them. CPU levels (parsing, featurization, writing) overlap with GPU folding.

Note: Ray doesn’t break up a single ahead go throughout GPUs. Replica mode scales worklists, not particular person targets. Per the support matrix, context-parallel folding is deliberate however not but out there. The capability rule is easy: engine_stage.compute x num_gpus should not exceed seen GPUs.

At the model-forward degree, NVIDIA’s early benchmarking reviews geometric-mean speedups over an OSS torch.compile baseline of 1.55x (OpenFold3), 1.78x (Boltz2), and 2.56x (OpenFold2 monomer) on H100. H200 numbers are comparable at 1.54x, 1.75x, and 2.61x. These have been measured throughout 17 inputs spanning 29 to 1,734 residues.

The Benchmark: 1,000 Human Dimers on 8xH100

To quantify end-to-end supply, NVIDIA staff ran a matched benchmark on 1,000 human dimer targets with mixed sequence lengths under 2,800 residues. The comparability pitted BioIR-accelerated Boltz-2 in opposition to a torch-compiled open-source Boltz-2 implementation on 8xH100 80GB GPUs. Both used equivalent targets, staged MSAs, inference recipe (3 recycles, 200 sampling steps, 5 diffusion samples), and GPU configuration.

The outcomes:

  • BioIR accomplished all 1,000 targets and delivered 58.5K efficiently folded residues per allotted GPU-hour.
  • The public implementation delivered 20.2K residues per GPU-hour and ran out of reminiscence on 29 targets.
  • Net outcome: a 2.90x enchancment in residue-normalized throughput.

These numbers are folding-stage measurements particular to this dataset and {hardware}. They exclude MSA technology, preprocessing CPU allocations, storage, information switch, and retries. The weblog explicitly warns in opposition to generalizing them to all BioIR-supported fashions or datasets.

Energy at One Million Targets

Extrapolating the benchmark linearly to 1 million comparable targets, BioIR is estimated to want 11 MWh versus 35 MWh for the general public implementation utilizing 8-GPU TDP equivalents. Using full-node maximum-power equivalents, the estimate is 21 MWh versus 64 MWh. These are rated-power, folding-only estimates for IT tools, not metered measurements, and exclude information heart overhead equivalent to PUE. Still, a 23 to 43 MWh saving per million targets is a cloth quantity for proteome-scale campaigns.

Key Takeaways

  • BioIR accelerates Boltz-2, OpenFold2, and OpenFold3 inference on NVIDIA GPUs whereas staying in plain PyTorch.
  • Matched 8xH100 benchmark: 58.5K vs 20.2K folded residues per GPU-hour, a 2.90x throughput acquire.
  • Ray reproduction mode scales unbiased worklists; it by no means splits 1 ahead go throughout GPUs.
  • Estimated vitality for 1M targets drops from 35 MWh to 11 MWh at 8-GPU TDP equivalents.
  • Already confirmed at scale: 31M candidate complexes generated for the AlphaFold Database growth.

Check out the technical blog, GitHub repo, docs, and the BioNeMo Agent Toolkit for agentic orchestration. Also, be at liberty to observe us on Twitter and don’t neglect to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to companion with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and many others.? Connect with us

The put up NVIDIA Details BioNeMo Inference Runtime (BioIR): 2.90x Higher Boltz-2 Folding Throughput and 58.5K Residues per GPU-Hour on 8xH100 appeared first on MarkTechPost.

Similar Posts