NVIDIA BioNeMo Inference Runtime Delivers 2.9x Faster Structure Prediction for Boltz-2
PyTorch · x · 2026-09-18
NVIDIA's new technical blog details BioNeMo Inference Runtime (BioIR), which accelerates biomolecular structure-prediction models while keeping the PyTorch workflow.
Key benchmark (1,000 human dimers, 8xH100):
- BioIR-accelerated Boltz-2: 58.5K successfully folded residues per GPU-hour vs 20.2K for a torch-compiled baseline — a 2.90x throughput gain
- Estimated energy to fold 1M targets drops from 35 MWh to 11 MWh
Three optimization layers: kernel selection, CUDA Graph capture for module optimization, and Ray replicas for pipeline scaling (full model replica per GPU, overlapping CPU stages with GPU folding).
Supports Boltz-2, OpenFold2, and OpenFold3 end-to-end (parsing, tokenization, feature generation, GPU inference, PDB/mmCIF output), and integrates as torch.nn.Module for custom code.
More from Infra
- HyperQwen seeks 4090/5090 owners to push Qwen local inference speeds — iamMess · 2026-09-18
- REBCO superconducting tape: 50-tesla fusion magnets and the rare-earth angle — Anen-o-me · 2026-09-18
- Where does VRAM go during LLM inference? Four buckets explained — blaizedsouza · 2026-09-18
- Planned US data centers hit 342GW, nearly 7x the 50GW installed today — BenBajarin · 2026-09-18
- $42 per billion input tokens with free output: an AI API price that looks like black magic — altryne · 2026-09-18
- openjev-sglang: Qwen3.6-35B-A3B serves Jev-compatible API, 64 tasks in under 1s — andersonbcdefg · 2026-09-18