NVIDIA Open-Sources BioNeMo Inference Runtime, Boosting Protein Structure Prediction Throughput up to 2.9x
NVIDIA Health has opened the public beta of BioNeMo Inference Runtime (BioIR), an open-source, PyTorch-native library that exposes the acceleration technology previously behind NVIDIA's NIM microservices, covering protein, nucleic acid, and ligand structure prediction models. Researchers can get near-production-grade speedups on their own GPU clusters without converting models or migrating to NIM. The many posts in this cluster are different retellings of the same announcement and corroborate each other.
Confirmed
- Provides an end-to-end processing pipeline: from FASTA/MSA input parsing and tokenization to structure outputs
- Accelerates inference via dedicated GPU kernels and CUDA Graphs while cutting memory use (multiple posts report roughly one-third the memory) and preserving the familiar PyTorch workflow
- Supports structure prediction models including Boltz-2, OpenFold2, OpenFold3, and Protenix v2, targeting drug discovery and protein design
- NVIDIA reports up to 2.9x Boltz-2 throughput (tested on 8x H100) and over 2x speedup for OpenFold3 on H100
- The team previously saved substantial compute serving the AlphaFold Database with this approach; NVIDIA estimates 1.35 GWh of electricity saved for 31 million protein predictions
- SandboxAQ celebrated the launch and said it will continue collaborating on a pilot stack for "traceable and fast" drug discovery, combining its AQ knowledge capabilities
Why it matters
- Structure prediction is a core computational workload in drug design and biology; the open-source library lowers the barrier to production-grade acceleration without depending on NIM microservices
- The SandboxAQ partnership shows the runtime being applied to real industrial scenarios such as traceable drug discovery
- NVIDIA's energy-saving estimate (1.35 GWh for 31 million predictions) underscores the compute and energy value of the stack at the scale of large structure databases
2026-09-10 ~ 2026-09-12 · 19 related posts
Primary sources
- NVIDIA's BioNeMo Inference Runtime hits public beta, boosting Boltz-2 folding throughput 2.9x — AllThingsApx · 2026-09-10
- [source] NVIDIA open-sources BioNeMo Inference Runtime, the GPU engine behind AlphaFold DB at million-scale — AllThingsApx · 2026-09-10
- NVIDIA open-sources BioNeMo Inference Runtime, saving 1.35 GWh across 31M protein predictions — AllThingsApx · 2026-09-11
- NVIDIA Opens BioNeMo Inference Runtime Beta; SandboxAQ Pilots Traceable Drug Discovery on Top — AllThingsApx · 2026-09-11
- NVIDIA open-sources BioNeMo Inference Runtime to speed up protein structure prediction — AllThingsApx · 2026-09-11
- BioNeMo beta benchmarks show up to ~2.9x structure prediction throughput on 8x H100 — AllThingsApx · 2026-09-11
13 near-duplicate retellings: AllThingsApx · AllThingsApx · AllThingsApx · AllThingsApx · AllThingsApx · AllThingsApx · AllThingsApx · AllThingsApx · AllThingsApx · AllThingsApx · AllThingsApx · AllThingsApx · AllThingsApx