NVIDIA Details BioNeMo Inference Runtime (BioIR): 2.90x Higher Boltz-2 Folding Throughput and 58.5K Residues per GPU-Hour on 8xH100
MarkTechPost Asif Razzaq
NVIDIA says its new BioIR runtime speeds protein folding on its GPUs. In one 8xH100 test, it folded 2.9x more residues per GPU-hour than a torch-compiled open-source Boltz-2 build.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
NVIDIA’s BioNeMo Inference Runtime, or BioIR, is aimed at a very specific pain point: getting through huge protein-folding queues faster. The model itself is no longer the hard part. The slog is everything around it — parsing inputs, generating features, running inference, and writing outputs — and BioIR is built to speed that up while keeping the normal PyTorch workflow intact.
The company says BioIR is already beyond the demo stage. It was used in the recent AlphaFold Database expansion, where it helped generate protein-complex structures across 4,777 proteomes. That work produced about 31 million candidate complexes, with 1.81 million released as high-confidence predictions. BioIR is also available now as an open GitHub repository, with a wheel that includes precompiled CUBINs. To run it, users need Python 3.12 or newer, a compatible NVIDIA GPU and driver, a staged checkpoint, and per-chain A3M MSAs. No nvcc. No CUDA source. No CMake. No CUDA toolkit.
BioIR focuses on the parts of inference stacks that tend to get neglected by general-purpose tools: Pairformer and Evoformer stacks, triangle operations, pairwise attention, diffusion transformers, and atom-level modules. Models stay as ordinary torch.nn.Module objects, so there is no export step or separate engine artifact between a checkpoint and a forward pass. There are two paths in: an end-to-end processor that handles parsing, tokenization, feature generation, GPU inference, and PDB or mmCIF output; and a direct integration route for developers who want to use optimized pieces inside their own code.
NVIDIA’s benchmark story leans heavily on worklist throughput, not single-target heroics. In a test of 1,000 human dimer targets on 8xH100 80GB GPUs, BioIR-accelerated Boltz-2 finished every target and delivered 58.5K successfully folded residues per allocated GPU-hour. A torch-compiled open-source Boltz-2 implementation reached 20.2K residues per GPU-hour and ran out of memory on 29 targets. That works out to a 2.90x gain in residue-normalized throughput, though NVIDIA says the numbers only cover the folding stage and should not be generalized to every model or dataset.
The interesting part is that BioIR is not trying to be a grand rewrite of inference. It is a set of targeted optimizations, plus Ray-based replica scaling for independent inputs, wrapped around plain PyTorch. That is the right kind of boring. The AI industry loves dramatic new stacks; the actual money is still in shaving overhead off queues full of work.
My take — AI-written commentary, not fact-checked reporting
BioIR is the sort of unglamorous infrastructure that actually matters, which is why it won’t get nearly the hype of another shiny model launch. NVIDIA is doing the practical thing here: keep PyTorch, cut the waste, and let the GPUs sweat instead of the framework. That is a much better look than another grand “platform” pitch with a thousand moving parts and no throughput to show for them.
Read more about this at: MarkTechPost