GPU acceleration of rank-reduced coupled-cluster singles and doubles

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• High-Performance Computing

Background:

• Coupled-cluster singles and doubles (CCSD) is a standard for accurate electronic structure calculations.
• Canonical CCSD is computationally expensive, limiting its application to smaller systems.
• Developing scalable and efficient quantum chemistry methods is crucial for modern computational science.

Purpose of the Study:

• To present a GPU-accelerated implementation of the rank-reduced coupled-cluster singles and doubles (RR-CCSD) method.
• To demonstrate the scalability and efficiency of RR-CCSD for large electronic systems.
• To assess the accuracy of RR-CCSD compared to canonical CCSD.

Main Methods:

• Implemented rank-reduced coupled-cluster singles and doubles (RR-CCSD) utilizing low-rank approximations for doubles amplitudes and electron repulsion integrals (via Cholesky decomposition).
• Leveraged graphical processing units (GPUs) for accelerated computation.
• Constructed only a single fourth-order tensor as an intermediate during amplitude equation solution.

Main Results:

• Achieved excellent parallel efficiency (95% on eight GPUs) due to compression of doubles amplitudes.
• Enabled RR-CCSD calculations for systems up to 400 electrons and 1550 basis functions, exceeding canonical CCSD limits.
• Demonstrated RR-CCSD computations are faster than canonical CCSD for most molecules.
• Attained accuracy better than 0.1% error in correlation energy for large systems with ~95% compression.
• Predicted conformational energies within 0.1 kcal mol⁻¹.
• Found minimal errors (hundredths of an eV) in excitation energies using low-rank approximations.

Conclusions:

• The GPU-accelerated RR-CCSD method offers significant improvements in scalability and speed over canonical CCSD.
• RR-CCSD provides a highly accurate and efficient approach for electronic structure calculations of large chemical systems.
• This method opens new possibilities for studying complex molecular systems and properties.