Performance in OpenFold3 NIM

NIM Accuracy

OpenFold3 is an all-atom biomolecular complex structure prediction model from the OpenFold Consortium and the AlQuraishi Laboratory. OpenFold3 is a PyTorch implementation of the jax-based AlphaFold3 reported in Accurate structure prediction of biomolecular interactions with AlphaFold 3, and like AlphaFold3, OpenFold3 extends protein structure prediction capabilities to model complete biomolecular complexes including proteins, DNA, RNA, and small molecule ligands.

The OpenFold3 NIM’s accuracy should match that of the AlQuraishi Laboratory implementation of OpenFold3, when using equivalent parameters and inputs.

The accuracy of the NIM is measured by structural quality metrics such as lddt (local distance difference test). These scores help assess the reliability of the predicted structures.

Factors Affecting NIM Performance

The performance of the OpenFold3 NIM is determined by several key factors:

Hardware Factors

• GPU type and memory: Different GPU architectures provide different performance levels
• System RAM: Larger proteins and complexes require more system memory
• Storage speed: Fast NVMe SSD storage improves model loading and caching performance

Input Complexity

• Sequence length: Runtime scales approximately linearly with total sequence length
• Number of chains: Multi-chain complexes require more computation than single chains
• MSA size: Larger MSAs can improve accuracy but increase memory usage and computation time
• Ligands, DNA, RNA: Additional molecular components increase computational cost

Model Configuration

• Inference backend: TensorRT + cuEquivariance provides significant speedups over PyTorch
• Diffusion samples: Multiple samples provide diversity but multiply computational cost
• MSA depth: Deeper MSAs improve accuracy with modest runtime increase

Performance Characteristics

Typical Runtimes

For reference, approximate runtimes on high-end hardware (NVIDIA H100 80GB):

Structure Prediction (OpenFold3 v1.0.0 on H100 with TensorRT + cuEquivariance):

• ~200 residues: 11.8-14.5 seconds
• ~300-400 residues: 15.0-20.4 seconds
• ~500-600 residues: 21.5-25.7 seconds
• ~800-900 residues: 32.7 seconds
• ~1300-1500 residues: 46.5-67.9 seconds
• ~1700-1900 residues: 88.7-101.8 seconds
• Memory usage: Varies with sequence length, typically 40-80GB GPU memory

Performance Notes The total runtime for structure prediction depends on:

• Total number of residues in the complex
• Total number of atoms in the complex
• Number of molecules and chains
• Number of sequences in the MSAs
• Number of diffusion samples requested

OpenFold3 v1.0.0 Performance Results on H100 (NVIDIA H100 80GB)

Structure Prediction Performance

The following table shows runtime performance for OpenFold3 NIM v1.0.0 on NVIDIA H100 GPUs, across 3 different inference backends. These backends are described below in Backend Selection below, and in Backend Optimization Options

*All runtimes are in seconds for end-to-end structure prediction with a single diffusion sample

Performance Analysis

Key Observations:

• OpenFold3 NIM Optimization: Provides consistent 1.15x to 1.69x speedup over the open source OF3 baseline
• cuEquivariance Acceleration: PyTorch + cuEquivariance shows speedups for larger proteins, demonstrating the value of cuEquivariance optimization
• Scaling Behavior: Performance scales approximately linearly with sequence length
• Best Performance: Largest speedups (1.6x-1.69x) observed for proteins in the 1400-1500 residue range
• Small Proteins: Still achieve significant speedups (1.2x-1.4x) even for sequences under 400 residues

Backend Comparison:

• Open source OF3 (PyTorch + DeepSpeed): Baseline implementation from the openfold consortium, good for development and debugging
• PyTorch + cuEquivariance: Improved performance, especially for larger proteins, while maintaining PyTorch flexibility
• OpenFold3 NIM (TensorRT + cuEquivariance): Best performance across all sequence lengths, recommended for production deployments

Recommended Configuration:

• For development/testing: Use open source OF3 (PyTorch + DeepSpeed) for easier debugging and flexibility
• For production: Use OpenFold3 NIM (TensorRT + cuEquivariance, default) for optimal performance
• For large proteins (>1500 residues): OpenFold3 NIM (TensorRT + cuEquivariance) provides the best speedups (1.6x-1.7x)
• For sequence lengths outside 4-2048 range: Use PyTorch backend as TensorRT has length limitations

Configuration

The benchmarks above use the following configuration:

Performance Optimization Tips

Backend Selection

• Default (TensorRT + cuEquivariance): Best for most use cases

  # This is the default, no environment variable needed

• PyTorch + cuEquivariance: For flexibility with good performance

  export NIM_OPTIMIZED_BACKEND=torch

• PyTorch + DeepSpeed: For debugging or sequences outside TRT range

  export NIM_OPTIMIZED_BACKEND=torch_baseline

General Optimization Tips

• GPU Selection: Use H100 or H200 GPUs for optimal performance. A100 GPUs are also supported.
• TRT Sequence Length Limits: TensorRT mode (default) supports sequences between 4 and 2048 residues. For sequences outside this range, use PyTorch backend.
• Sequence Length: Performance scales approximately linearly with sequence length.
• Multiple Samples: Setting diffusion_samples > 1 will increase runtime in affine fashion (input featurization time is independent of diffusion_samples).
• MSA Size: While larger MSAs can improve accuracy, they also increase memory usage and computation time. Consider filtering MSAs for very large proteins.
• Batch Processing: For multiple independent predictions, process them sequentially or use multiple NIM instances.
• Memory Management: Ensure adequate GPU memory for your target sequence lengths. Very long sequences (>1800 residues) may require H100/H200 80GB GPUs.

Troubleshooting

General Performance Issues

• Out of memory errors: Reduce MSA depth, decrease sequence length, or upgrade to GPUs with more memory
• Slow performance: Ensure fast storage (NVMe SSD), sufficient CPU cores, and adequate system RAM
• Poor quality predictions: Check input sequence quality, increase MSA depth if available, or adjust diffusion parameters

Backend-Specific Issues

• TensorRT errors with long sequences: Use PyTorch backend for sequences >2048 residues

  export NIM_OPTIMIZED_BACKEND=torch
• TensorRT errors with short sequences: Use PyTorch backend for sequences of fewer than 4 residues

  export NIM_OPTIMIZED_BACKEND=torch
• Inconsistent results between backends: This is expected; TensorRT uses optimizations that may produce slightly different numerical results while maintaining accuracy