NVIDIA Details BioNeMo Inference Runtime (BioIR): 2.90x Higher Boltz-2 Folding Throughput and 58.5K Residues per GPU-Hour on 8xH100

Details

Protein structure prediction has moved from single-target experiments to proteome-scale worklists, and the constraint is now throughput, not model capability. NVIDIA's BioNeMo Inference Runtime (BioIR) addresses this directly: it is a Python library that accelerates structure-prediction models on NVIDIA GPUs while keeping models as standard torch.nn.Module objects — no engine build, no export step, no compiler toolchain required. BioIR operates across three acceleration layers. First, kernel selection picks the most compatible implementation — BioIR custom, cuEquivariance, or PyTorch fallback — based on GPU, data type, and tensor shape. Second, CUDA Graph capture via an optimize() call cuts kernel launch overhead for compatible modules. Third, a Ray executor places one complete model replica per visible GPU and distributes independent inputs across them, with CPU stages (parsing, featurization, output writing) overlapping GPU inference. The numbers are concrete: 2.90x higher Boltz-2 throughput versus a baseline stack, 58.5K residues per GPU-hour on an 8xH100 node. The tool is not theoretical — BioIR powered the recent AlphaFold Database expansion across 4,777 proteomes, processing roughly 31 million candidate complexes and releasing 1.81 million high-confidence predictions. Deployment requires Python 3.12+, an NVIDIA GPU with compatible driver, a staged checkpoint, and per-chain A3M MSAs. No nvcc, CMake, or CUDA toolkit is needed. A precompiled wheel is available on GitHub. One scope note for drug discovery teams: BioIR supports ligand structure prediction but not ligand-affinity prediction.