NVIDIA has layered multiple accelerations around OpenFold3 and related co-folding workflows to make large-scale biomolecular structure prediction practical for drug discovery and protein design. The toolkit includes GPU-based MSA generation (MMseqs2-GPU), the cuEquivariance CUDA-X library for geometric kernel acceleration, an optimized OpenFold3 NIM for inference, and Fold-CP for context-parallel multi-GPU execution. These components are callable from agents via the NVIDIA BioNeMo Agent Toolkit.
Why this matters
Multiple Sequence Alignment (MSA) creation, co-folding inference, serving, and multi-GPU scaling are all required for an end-to-end structure-prediction pipeline. A bottleneck in any stage limits throughput. Speed and memory efficiency are crucial for workflows such as virtual screening (millions to billions of compounds) and prediction of large molecular assemblies spanning thousands of residues.
GPU MSA: removing the MSA bottleneck
Homology search has traditionally been CPU-bound and can dominate wall-clock time. MMseqs2-GPU shifts homology search to NVIDIA GPUs, reducing this bottleneck while scaling with sequence length on both NVIDIA Hopper and Blackwell architectures. The latest GPU-accelerated release includes Hopper- and Blackwell-specific optimizations, support for database searches larger than GPU memory on NVIDIA Grace systems, and additional speedups via Blackwell DPX instructions available in CUDA 13.2.
The MSA Search NIM uses MMseqs2-GPU; according to the cited Nature Methods paper, it achieves up to 177× faster alignments than CPU JackHMMER on a single L40S. NVIDIA benchmarking shows the MSA stage scales smoothly past 10k tokens on H100 and B300 GPUs (the article presents example latencies around ~429 s vs ~463 s at ~10k tokens). The MSA Search NIM can be called directly, self-hosted, or wrapped as an agent tool. (NVIDIA notes that build.nvidia.com hosted endpoints are not intended for high-volume requests; for large workloads the self-hosted NIM endpoint is recommended.)
cuEquivariance and the OpenFold3 NIM: faster folding and longer sequences
cuEquivariance is a CUDA-X library of geometric learning primitives for atomistic modeling, accelerating kernels such as Triangle Attention, Triangle Multiplication, and Attention Pair Bias that dominate co-folding. On B300 it reduces latency by roughly ~1.8–3.1× across several sequence lengths shown by NVIDIA (examples: 1,024 tokens on H100 from 79.4 s to 41.6 s; 1,024 on B300 from 49.1 s to 27.3 s; 1,536 on B300 from 149.0 s to 56.2 s; 2,048 on B300 from 300.1 s to 97.6 s).
cuEquivariance kernels are integrated into open-source models such as OpenFold3 (as an optional dependency), OpenFold2, RosettaFold3, Protenix, and Boltz. Because these accelerations are upstreamed into OSS models, users obtain speedups by running their existing models on NVIDIA GPUs. cuEquivariance also extends maximum sequence length to about ~5.9k tokens, while pure PyTorch runs often run out of memory beyond ~1.5k–2.5k tokens.
On top of cuEquivariance, the OpenFold3 NIM applies additional inference optimizations that compound the gains; NVIDIA reports sequence lengths up to ~6,400 on a single B300 and overall runtime improvements of over 3× on H100 and over 4× on B300 for the fully optimized NIM.
Fold-CP: scaling beyond a single GPU
Single-GPU memory has historically capped co-folding models at a few thousand residues. On NVIDIA B300 (Blackwell Ultra), larger HBM and generation efficiency together with cuEquivariance and inference optimizations raise the single-GPU ceiling without model changes. However, many problems demand more than a single device can hold.
Fold-CP introduces a parallelization technique where per-device memory scales as O(N²/P) (N = token count, P = number of GPUs). NVIDIA demonstrates reaching 32,000 tokens on 64 B300 GPUs using the Boltz-2 model—about a 12× increase over the single-GPU limit. Fold-CP is available in the Boltz-CP codebase and can be invoked from agents or run directly with the provided torchrun examples.
End-to-end acceleration and practical effects
NVIDIA treats structure prediction as a systems problem spanning MSA generation, inference, deployment, and memory limits that determine how large an assembly can be predicted. The combined accelerations are:
- MSA Search NIM: homology search speedups up to 177× versus CPU JackHMMER on a single L40S.
- cuEquivariance + OpenFold3 NIM: inference latency reductions of multiple× (examples above) and single-device sequence length increases up to ~6.4k tokens on B300.
- Fold-CP: context-parallel scaling to 32,000 tokens across 64 B300 GPUs with Boltz-2.
Together, these tools make structure prediction faster, more scalable, and easier to integrate into agentic discovery workflows, helping researchers move from per-structure predictions to modeling larger biomolecular systems.
Practical consequences include:
- Virtual screening: faster co-folding inference enables earlier and larger-scale use of structure-based methods in screening pipelines, improving hit quality and diversity.
- Large assemblies: the combination of extended single-GPU capacity and Fold-CP may make modeling of ribosome- or spliceosome-scale complexes (on the order of 10k residues) tractable across multi-GPU nodes, shifting what structural biology can question computationally rather than merely improving throughput.
Getting started
NVIDIA exposes these accelerations via the BioNeMo Agent Toolkit and NIM endpoints. Key resources mentioned include:
- MSA Search NIM (MMseqs2-GPU)
- cuEquivariance (GitHub)
- OpenFold3 NIM
- Fold-CP (Boltz-CP GitHub repository)
NVIDIA also notes that NIM containers can be downloaded and self-hosted for production-scale workloads.
Acknowledgments
NVIDIA thanks team members involved in benchmark and tool development, including Franco Pellegrini, Lalit Vaidya, Duc Tran, Tien Pham, Maximilian Stadler, Alejandro Chacon, Quan Vu, Simon Chu, Brian Roland, Dejun Lin, Joseph Chang, Hoa La, Jonathan Mitchell, Vishanth Iyer, Timur Rvachov, Christian Dallago, and Christian Hundt, as well as contributors from the OpenFold and OMSF communities.
