High-Throughput Structure Prediction with BioNeMo Inference Runtime

NVIDIA Releases BioNeMo Inference Runtime to Accelerate Proteome-Scale Research

NVIDIA has launched the BioNeMo Inference Runtime (BioIR), a new software tool designed to significantly speed up biomolecular structure-prediction models on its GPUs. The runtime addresses the growing need for high-throughput analysis in fields like drug discovery, where researchers now operate at proteome scale. In a benchmark comparing a BioNeMo Inference Runtime-accelerated Boltz-2 model against a torch-compiled open-source implementation, NVIDIA reported a 2.90x improvement in residue-normalized throughput, achieving 58.5K folded residues per allocated GPU-hour versus 20.2K. This performance increase enables researchers to process vast worklists of proteins more efficiently, reducing both time and energy costs for large-scale projects.

The runtime delivers its performance gains through a three-layer optimization strategy while maintaining a familiar PyTorch workflow. For developers, it can be used either as an end-to-end processor that handles the entire pipeline from input parsing to PDB/mmCIF file output, or through direct integration of its optimized modules into custom code. To scale across a single node, the system leverages Ray to place a complete model replica on each GPU, overlapping CPU-bound stages like feature generation with GPU-based folding to maximize hardware utilization. The core optimizations include:

• Kernel Selection: Dynamically chooses optimized custom kernels based on the model, GPU, and data type.
• Module Optimization: Uses CUDA Graph capture where applicable to reduce model-forward latency within each replica.
• Pipeline Scaling: Distributes independent prediction tasks across multiple full-model replicas running on separate GPUs.

The introduction of the BioNeMo Inference Runtime signals a strategic move by NVIDIA to provide domain-specific, full-stack solutions on top of its hardware. By abstracting away complex GPU optimizations and offering a streamlined path to scaling, the company lowers the barrier to entry for computational biology labs to perform massive-scale structure prediction. This has already been demonstrated in practice, with the runtime used to accelerate the generation of 1.81 million high-confidence protein-complex structures for the recent expansion of the AlphaFold Database. This approach not only solidifies the utility of NVIDIA hardware in the scientific research market but also creates an ecosystem that encourages further development and reliance on its platform for critical life sciences workloads.

Strategic Takeaway: NVIDIA is deepening its moat in scientific AI by providing highly-optimized, domain-specific software like the BioNeMo Inference Runtime, which transforms its GPUs from general-purpose accelerators into indispensable, full-stack solutions for critical research workflows like drug discovery.