NVIDIA ships BioIR: drop-in runtime folds proteins 2.9x faster
An open GitHub wheel speeds up Boltz-2, OpenFold2 and OpenFold3 - no CUDA build, stays pure PyTorch, and it already folded 31M complexes for AlphaFold.
Copy markdown
2.9x more folds per GPU-hour
On 1,000 human-dimer targets across 8xH100, BioIR folds 58.5K residues per GPU-hour versus 20.2K for the torch-compiled baseline - a 2.90x throughput gain on the same silicon. Single model-forward passes run 1.78x faster for Boltz-2 and 2.56x for OpenFold2 on an H100.
Install is a pip wheel, no CUDA toolkit
It ships as an open GitHub repo with a wheel of precompiled CUBINs: you need Python 3.12+, an NVIDIA GPU and driver, but no nvcc, CMake or CUDA source. Your models stay ordinary torch.nn.Module objects, so there's no export or recompile step to babysit.
It accelerates the folders you already run
BioIR serves Boltz-2, OpenFold2 and OpenFold3 behind one inference recipe - run it end-to-end (MSAs in, PDB/mmCIF out) or drop the optimized modules into a custom pipeline. Ray replica scaling fans independent targets across every GPU you have.
Already proven at 31M complexes
This isn't a benchmark toy: NVIDIA says BioIR already generated 31 million candidate protein complexes for the AlphaFold Database expansion, so the throughput holds up at production volume.
Build this weekend: a proteome-scale folding service
With batch throughput this high on commodity H100s, one builder can stand up a structure-prediction API or screen a full target list in a weekend - point Ray at a GPU pool, queue the sequences, and collect structures that used to need a dedicated infra team.