Accelerating Correlated Quantum Chemistry Calculations Using Graphical Processing Units and a Mixed Precision Matrix...

Area of Science:

• Computational Chemistry
• High-Performance Computing
• Quantum Chemistry

Background:

• Computational chemistry often faces performance bottlenecks due to large matrix operations.
• Standard single-precision computations on GPUs can lead to significant accuracy loss.
• Handling large matrices that exceed GPU memory limits is a challenge.

Purpose of the Study:

• To develop novel GPU-accelerated tools for computational chemistry.
• To address the limitations of matrix-matrix multiplication for large datasets on GPUs.
• To enhance the accuracy of computations on single-precision GPU devices.

Main Methods:

• A black-box approach for GPU acceleration of out-of-core matrix-matrix multiplications.
• A heterogeneous computing model combining single-precision GPU and double-precision CPU operations for mixed-precision calculations.
• Application to resolution-of-the-identity second-order Møller-Plesset perturbation theory (RI-MP2) calculations.

Main Results:

• Significant speedups observed for general matrix multiply (GEMM) operations: 13.8x (single-precision), 7.8x (double-precision), and 10.1x (mixed-precision) for a large molecule.
• Reduction in correlation energy errors from -10.0 kcal mol(-1) to -1.2 kcal mol(-1) using the mixed-precision approach.
• Demonstrated ability to perform previously intractable RI-MP2 calculations for large molecular systems.

Conclusions:

• The developed GPU acceleration tools provide substantial performance gains for computational chemistry.
• The mixed-precision approach effectively balances accuracy and computational speed, enabling more complex simulations.
• These advancements open possibilities for treating larger and more complex molecular systems in quantum chemistry.