Efficient Treatment of Large Active Spaces through Multi-GPU Parallel Implementation of Direct Configuration...

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Materials Science

Background:

• Accurate electronic structure calculations are crucial for understanding molecular properties.
• Large configuration interaction (CI) calculations are computationally demanding, limiting their application.
• Efficient algorithms are needed to handle complex molecular systems.

Purpose of the Study:

• To enhance a GPU-accelerated direct configuration interaction program for multi-device computation.
• To enable faster calculations of electronic properties and energy gradients.
• To facilitate routine simulations of molecules with large configuration spaces.

Main Methods:

• Extended a graphical processing unit (GPU)-accelerated direct configuration interaction program to multiple GPUs.
• Implemented parallel algorithms for large configuration spaces and reduced density matrix formation.
• Utilized NVIDIA P100 and 1080Ti GPUs for performance testing.

Main Results:

• Reduced iteration times for 165 million determinants to 3 seconds on NVIDIA P100 GPUs.
• Achieved 13-minute iteration times for 2.4 billion determinants (18 electrons in 18 orbitals) on six NVIDIA 1080Ti GPUs.
• Enabled fast analytical energy gradients and electronic property calculations.

Conclusions:

• The enhanced program significantly reduces computational cost for large CI calculations.
• This facilitates routine molecular dynamics, geometry optimizations, and absorption spectrum calculations.
• Demonstrated utility by calculating the absorption spectrum of diphenyl acetylene using floating occupation molecular orbital complete active space CI.