Official Resources
- Homepage: https://github.com/pyscf/gpu4pyscf
- Documentation: https://gpu4pyscf.readthedocs.io/
- Source Repository: https://github.com/pyscf/gpu4pyscf
- Parent Project: https://pyscf.org/
- License: Apache License 2.0
Overview
GPU4PySCF is a GPU-accelerated extension for PySCF, providing CUDA implementations of two-electron repulsion integrals and DFT calculations. It enables significant speedups for Hartree-Fock and DFT calculations while maintaining compatibility with the PySCF ecosystem.
Scientific domain: GPU-accelerated quantum chemistry, HF/DFT
Target user community: PySCF users needing GPU acceleration for larger molecules
Theoretical Methods
- Restricted/Unrestricted Hartree-Fock
- Density Functional Theory
- LDA, GGA, meta-GGA, hybrid functionals
- Density fitting (DF/RI-J/K)
- Direct SCF
- Geometry optimization
- Hessian calculations
Capabilities (CRITICAL)
- GPU-accelerated ERI evaluation
- CUDA kernel implementations
- Density fitting on GPU
- SCF acceleration
- DFT grid calculations
- Gradient calculations
- PySCF API compatibility
- Multi-GPU support
- cuBLAS integration
- Memory-efficient algorithms
Key Strengths
GPU Acceleration:
- CUDA-optimized integrals
- 10-100x speedups
- Modern GPU support
- Multi-GPU capability
- Memory management
PySCF Integration:
- Drop-in replacement
- Same API
- Ecosystem compatibility
- Workflow preservation
- Easy adoption
DFT Performance:
- Fast grid integration
- Hybrid functionals
- Exchange matrices
- Large molecules
Density Fitting:
- RI-J and RI-K
- GPU-accelerated DF
- Auxiliary basis sets
- Efficient memory use
Inputs & Outputs
-
Input formats:
- PySCF mol objects
- Standard coordinates
- Basis set strings
-
Output data types:
- Energies
- Gradients
- Densities
- Orbitals
- Properties
Interfaces & Ecosystem
- PySCF: Full integration
- NumPy/CuPy: Array backends
- CUDA: GPU runtime
- Post-HF: Enable GPU-accelerated reference
Advanced Features
ERI Acceleration:
- Schwarz screening
- Shell pair sorting
- Memory blocking
- Batched evaluation
Density Fitting GPU:
- Three-center integrals
- Coulomb/exchange
- Auxiliary screening
- Efficient contraction
Gradient Calculations:
- Analytical gradients
- Geometry optimization
- Nuclear forces
- GPU-accelerated
Performance Characteristics
- Speed: 10-100x GPU acceleration
- Accuracy: Identical to CPU PySCF
- System size: Significantly larger molecules
- Memory: GPU memory dependent
- Parallelization: Multi-GPU MPI
Computational Cost
- HF: Dramatic GPU speedup
- DFT: Fast grid + integrals
- Hybrid DFT: Excellent acceleration
- DF methods: Very efficient
- Typical: Enables previously infeasible calculations
Limitations & Known Constraints
- GPU requirement: NVIDIA CUDA needed
- Memory: Limited by GPU RAM
- Methods: HF/DFT focus (post-HF separate)
- Basis sets: Standard GTOs
- Platform: Linux primarily
- Installation: CUDA setup required
Comparison with Other Codes
- vs QUICK: Both GPU; GPU4PySCF PySCF ecosystem
- vs TeraChem: GPU4PySCF open-source, TeraChem commercial
- vs Psi4 GPU: Different implementations
- vs CPU PySCF: Same results, much faster
- Unique strength: PySCF integration, open-source, community
Application Areas
Large Molecules:
- Proteins and enzymes
- Organic semiconductors
- Supramolecular systems
- Drug molecules
High-Throughput:
- Database generation
- Virtual screening
- ML training data
- Property prediction
Method Development:
- Reference calculations
- Benchmarking
- Algorithm testing
- Rapid prototyping
Best Practices
GPU Setup:
- Modern NVIDIA GPU
- Sufficient GPU memory
- CUDA toolkit installed
- cuBLAS optimized
Calculation Strategy:
- Use density fitting
- Appropriate cutoffs
- Batch similar molecules
- Monitor GPU memory
Community and Support
- Open-source Apache 2.0
- PySCF community support
- Active GitHub development
- Documentation and examples
- Academic publications
Verification & Sources
Primary sources:
- GitHub: https://github.com/pyscf/gpu4pyscf
- PySCF ecosystem: https://pyscf.org/
- J. Chem. Theory Comput. (2024) GPU4PySCF paper
- Active development
Confidence: VERIFIED
- Source code: OPEN (GitHub, Apache 2.0)
- Documentation: ReadTheDocs
- Active development: Yes
- Part of PySCF project