---
title: "High-Throughput Structure Prediction with BioNeMo Inference Runtime"
description: "Proteome-scale biomolecular structure prediction requires moving large worklists efficiently, and NVIDIA BioNeMo Inference Runtime (BioIR) accelerates supported models on NVIDIA GPUs while retaining a PyTorch workflow. Its end-to-end processor passes InputRequest objects through parsing, tokenization, feature generation, GPU inference, and PDB or mmCIF writing; a Ray backend places complete model replicas on GPUs for concurrent independent inputs, while direct PyTorch integration supports custom code. In a matched 1,000-target human-dimer benchmark, BioIR completed all targets at 58.5K successfully folded residues per allocated GPU-hour versus 20.2K for the tested public implementation, a 2.90× improvement, which ran out of memory on 29 targets. Rated-power-equivalent estimates for one million comparable targets were 11 versus 35 MWh at 8-GPU TDP and 21 versus 64 MWh at full-node maximum power, excluding data-center overhead and several pipeline costs."
---

# High-Throughput Structure Prediction with BioNeMo Inference Runtime

[NVIDIA](https://yomu.fyi/company/nvidia-developer-blog) · Elizabeth Goodman · Sep 10, 2026

**Type:** Tutorial

## Summary

Proteome-scale biomolecular structure prediction requires moving large worklists efficiently, and NVIDIA BioNeMo Inference Runtime (BioIR) accelerates supported models on NVIDIA GPUs while retaining a PyTorch workflow. Its end-to-end processor passes InputRequest objects through parsing, tokenization, feature generation, GPU inference, and PDB or mmCIF writing; a Ray backend places complete model replicas on GPUs for concurrent independent inputs, while direct PyTorch integration supports custom code. In a matched 1,000-target human-dimer benchmark, BioIR completed all targets at 58.5K successfully folded residues per allocated GPU-hour versus 20.2K for the tested public implementation, a 2.90× improvement, which ran out of memory on 29 targets. Rated-power-equivalent estimates for one million comparable targets were 11 versus 35 MWh at 8-GPU TDP and 21 versus 64 MWh at full-node maximum power, excluding data-center overhead and several pipeline costs.

## Context

Proteome-scale structure prediction involves processing large worklists of independent inputs efficiently. The described benchmark and deployment guidance address throughput, resource efficiency, and operational considerations for running supported biomolecular structure-prediction models at scale.

## Approach / What changed

BioIR provides an end-to-end processor with serial and Ray executor backends, plus direct PyTorch module integration. The Ray workflow uses complete model replicas across visible GPUs and overlaps parser, tokenizer, feature-generation, inference, and writer stages. The tutorial covers input preparation, validation, configuration, measurement, stage balancing, and troubleshooting.

## Takeaways

- Each protein chain requires an A3M MSA; BioIR accepts paired or unpaired MSAs for multiple non-identical protein chains but does not run HHsearch or HMMsearch.
- The Ray replica capacity rule is engine\_stage.compute × engine\_stage.num\_gpus ≤ visible GPUs; each replica loads the full model and reserves its configured GPU resources.
- BioIR’s benchmark metrics separate model-forward time from end-to-end pipeline performance, and the energy figures are rated-power-equivalent estimates rather than metered measurements.

**Tags:** [BioNeMo](https://yomu.fyi/topic/bionemo), [CUDA](https://yomu.fyi/topic/cuda), [Ray](https://yomu.fyi/topic/ray)

- Source: [NVIDIA](https://developer.nvidia.com/blog/high-throughput-structure-prediction-with-bionemo-inference-runtime)
- Source URL: https://developer.nvidia.com/blog/high-throughput-structure-prediction-with-bionemo-inference-runtime
- Ingested by Yomu: 2026-09-10T15:20:44.087Z

[Read original post](https://developer.nvidia.com/blog/high-throughput-structure-prediction-with-bionemo-inference-runtime)