Phaseless Auxiliary-Field Quantum Monte Carlo on Graphical Processing Units.

Area of Science:

• Computational Chemistry
• Quantum Many-Body Physics
• High-Performance Computing

Background:

• The phaseless Auxiliary-Field Quantum Monte Carlo (ph-AFQMC) method is a powerful tool for studying correlated electronic systems.
• Traditional ph-AFQMC implementations face computational challenges, limiting their application to smaller systems.

Purpose of the Study:

• To develop and implement a highly efficient GPU-accelerated version of the ph-AFQMC method.
• To significantly reduce computational time for ph-AFQMC calculations, enabling the study of larger and more complex systems.

Main Methods:

• Recasting the AFQMC method into matrix operations suitable for GPU parallelization.
• Utilizing custom Compute Unified Device Architecture (CUDA) kernels and the cuBLAS library for optimized matrix computations.
• Implementing algorithmic advances including batched matrix updates, density-fitting, and single-precision arithmetic.

Main Results:

• Achieving speed-ups of approximately two orders of magnitude per GPU compared to a single CPU core.
• Demonstrating near-unity parallel efficiency with multiple GPUs.
• Successfully applying the method to hydrogen chains and calculating ionization potentials of transition metal atoms.

Conclusions:

• The GPU-accelerated ph-AFQMC implementation offers a substantial performance improvement.
• The enhanced method broadens the scope of ph-AFQMC for realistic correlated electronic system studies.
• Combining GPU acceleration with correlated sampling improves both efficiency and accuracy.