NVIDIA BioNeMo Runtime Delivers 2.9x Throughput Gain for Protein Structure Prediction
AllThingsApx · x · 2026-09-11
An NVIDIA technical blog introduces the BioNeMo Inference Runtime (BioIR) for accelerating biomolecular structure-prediction models on NVIDIA GPUs while preserving the familiar PyTorch workflow.
- End-to-end processor: covers input parsing, tokenization, feature generation, GPU inference, and PDB/mmCIF output; also supports direct PyTorch module integration for custom code.
- High-throughput design: the Ray executor places a full model replica on each GPU in a node, overlapping CPU stages with GPU folding.
- Benchmark: on 8xH100 with 1,000 human dimer targets, BioIR-accelerated Boltz-2 folded 58.5K residues per allocated GPU-hour vs 20.2K for a torch-compiled open-source implementation — a 2.90x throughput improvement.
- Three optimization layers: kernel selection, module optimization via CUDA Graph capture, and pipeline scaling via Ray replicas.
- Energy: folding one million comparable targets is estimated at 11 MWh vs 35 MWh for the public implementation.
Related event: NVIDIA Open-Sources BioNeMo Inference Runtime in Public Beta(9 posts)→
More from Infra
- 8x RTX 5090 training run hits ~$11/b tokens amid ~50% GPU failure rates — jon_durbin · 2026-09-11
- DeepSeek's New Model: 4x Smaller KV Cache Than DSV4-Flash and More Stable Training — stochasticchasm · 2026-09-11
- Reflect Orbital readies first satellite to sell sunlight, unfolding a volleyball-court-sized mirror in orbit — kyliebytes · 2026-09-11
- LLM inference bottlenecks: weight loading gave way to KV reads as contexts grew — YouJiacheng · 2026-09-11
- After GPUs and memory, AI agents are now driving a CPU shortage — The Pragmatic Engineer · 2026-09-11
- Dev shares training dashboard: ~$11/b tokens cost with 'insane' MFU — jon_durbin · 2026-09-11