Key Takeaways
- BioIR folded 58.5K residues per GPU-hour versus 20.2K for torch-compiled open-source Boltz-2 across 1,000 human dimer targets on eight H100 GPUs.
- Acceleration comes from three layers: adaptive kernel selection, CUDA Graph module optimization, and a Ray backend running five-stage pipelines with full model replicas per GPU.
- NVIDIA extrapolates 11 MWh versus 35 MWh for one million targets at 8-GPU TDP, as the AI protein design market is projected to reach USD 12.77 billion by 2035.
Table of Contents
BioNeMo Inference Runtime Clears a 2.90x Throughput Bar
NVIDIA’s developer publication reported on September 10 that BioNeMo Inference Runtime has produced a 2.90x residue-normalized throughput gain over a torch-compiled open-source Boltz-2 implementation.
The matched benchmark loaded 1,000 human dimer targets onto eight H100 GPUs, where BioIR folded 58.5K residues for every allocated GPU-hour against 20.2K from the public implementation.
The worklist represented human dimer targets with combined sequence lengths below 2,800 residues.
The result is configuration-specific: both implementations used the same targets, staged MSAs, inference recipe, and GPU configuration.
NVIDIA cautions that the throughput metric should not be generalized to all BioIR-supported models, datasets, or hardware.
Inside the Three-Layer Acceleration Stack
End-to-End Processor and Direct PyTorch Paths
BioIR keeps the familiar PyTorch workflow while exposing two routes for structure prediction.
The end-to-end processor carries an input request through parsing, tokenization, feature generation, GPU inference, and PDB or mmCIF output.
Direct integration lets developers construct a supported torch.nn.Module or reuse selected modules inside custom code.
The wheel ships with precompiled CUBINs, so runtime use does not require nvcc, CUDA source, CMake, or the CUDA toolkit.
The end-to-end processor supports ligand structure prediction but not ligand-affinity prediction.
Kernel and Module-Level Optimization
The first layer of acceleration sits inside model execution.
Kernel selection chooses BioIR custom kernels, cuEquivariance, or PyTorch fallback based on the model configuration, GPU, data type, and tensor shape.
The second layer applies module optimization with CUDA Graph capture for compatible modules.
This reduces launch overhead and improves synchronization behavior before any multi-GPU scaling is added.
NVIDIA’s model-forward figures show geometric-mean speedups on H100 and H200 relative to OSS torch.compile across 17 inputs ranging from 29 to 1,734 residues.
Both implementations used three recycles, 200 sampling steps, and five diffusion samples per target.
The open-source code also ran out of memory on 29 targets, while BioIR completed all 1,000.
Ray Replicas and the Five-Stage Pipeline
For large worklists, BioIR’s Ray backend places a full model replica on each visible GPU in a single node.
The pipeline then overlaps CPU-bound parsing, tokenization, feature generation, and writing with GPU-bound folding.
Five stages run in dependency order: Parser, Tokenizer, Feature generator, Folding engine, and Writer.
The engine stage reserves one GPU per replica, and the effective capacity rule is engine_stage.compute times engine_stage.num_gpus not exceeding visible GPUs.
Ray does not split one forward pass across GPUs; it multiplies complete replicas to process independent inputs concurrently.
A serial backend runs each stage in sequence to validate setup, while the Ray backend processes independent inputs concurrently.
Energy and Market Forces Behind Proteome-Scale Folding
BioIR’s throughput advantage becomes more visible at proteome scale.
The runtime has already supported the expansion of the AlphaFold Database across 4,777 proteomes, roughly 31 million candidate protein complexes, with 1.81 million released as high-confidence predictions.
From the 1,000-target benchmark, NVIDIA extrapolated folding energy to one million comparable targets.
Using 8-GPU TDP equivalents, BioIR would need an estimated 11 MWh against 35 MWh for the public implementation.
At full-node maximum power, the estimate moves to 21 MWh versus 64 MWh.
These are folding-only IT equipment estimates, not metered energy measurements, and they exclude data center overhead such as PUE.
The market context makes such efficiency gains consequential.
Market estimates put the global AI protein design market at USD 1.45 billion in 2025, rising to a projected USD 12.77 billion by 2035.
The structure prediction and modeling segment held a dominant share in 2025, with North America at 39% and Asia Pacific emerging as the fastest-growing region at approximately 28%.
A January 2026 NVIDIA and Eli Lilly initiative backs this trajectory with up to USD 1 billion over five years for a Bay Area AI co-innovation lab using BioNeMo.
Peer-reviewed work in Briefings in Bioinformatics underscores why accurate protein structures are essential for understanding molecular mechanisms and interpreting sequence variation.
The wider protein engineering market is also projected to grow from USD 3.65 billion in 2025 to USD 14.46 billion by 2034, with rational protein design representing about 41% of method-based revenue.
What Higher Throughput Unlocks for Drug Discovery
The 2.90x result reframes biomolecular structure prediction as a high-volume inference problem, where cost per folded residue and energy draw now define feasibility at proteome scale. For teams turning AI infrastructure advances into authoritative coverage and scalable search visibility, programmatic SEO and AI automation is how Andres SEO Expert approaches high-velocity technical publishing — talk to us.
Frequently Asked Questions
What is BioNeMo Inference Runtime (BioIR)?
BioNeMo Inference Runtime is NVIDIA’s runtime for high-throughput biomolecular structure prediction. It keeps a familiar PyTorch workflow while offering an end-to-end processor and direct PyTorch integration for supported torch.nn.Module components.
How much faster is BioIR than open-source Boltz-2?
BioIR produced a 2.90x residue-normalized throughput gain over a torch-compiled open-source Boltz-2 implementation. In the matched benchmark, BioIR folded 58.5K residues per allocated GPU-hour versus 20.2K for the public implementation.
What benchmark and hardware did NVIDIA use for the 2.90x result?
NVIDIA loaded 1,000 human dimer targets with combined sequence lengths below 2,800 residues onto eight H100 GPUs. Both implementations used the same targets, staged MSAs, inference recipe, and GPU configuration, and NVIDIA cautions the result is configuration-specific.
How does BioIR accelerate model execution?
BioIR uses kernel selection among custom kernels, cuEquivariance, or PyTorch fallback based on model configuration, GPU, data type, and tensor shape. It also applies module optimization with CUDA Graph capture for compatible modules, reducing launch overhead and improving synchronization. The wheel ships with precompiled CUBINs, so runtime use does not require nvcc, CUDA source, CMake, or the CUDA toolkit.
How does BioIR’s Ray backend scale across GPUs?
For large worklists, BioIR’s Ray backend places a full model replica on each visible GPU in a single node. It overlaps CPU-bound parsing, tokenization, feature generation, and writing with GPU-bound folding across five stages: Parser, Tokenizer, Feature generator, Folding engine, and Writer. Ray does not split one forward pass across GPUs; it multiplies complete replicas to process independent inputs concurrently.
What energy savings does BioIR offer at proteome scale?
NVIDIA extrapolated folding energy from the 1,000-target benchmark to one million comparable targets. Using 8-GPU TDP equivalents, BioIR would need an estimated 11 MWh versus 35 MWh for the public implementation. At full-node maximum power, the estimate is 21 MWh versus 64 MWh. These are folding-only IT equipment estimates, not metered energy measurements, and they exclude data center overhead such as PUE.
What market trends are driving proteome-scale protein folding?
Market estimates put the global AI protein design market at USD 1.45 billion in 2025, rising to a projected USD 12.77 billion by 2035. Structure prediction and modeling held a dominant share in 2025, with North America at 39% and Asia Pacific as the fastest-growing region at about 28%. NVIDIA and Eli Lilly announced an initiative of up to USD 1 billion over five years for a Bay Area AI co-innovation lab using BioNeMo.
