Accelerating Pythonic Coupled-Cluster Implementations: A Comparison Between CPUs and GPUs

Area of Science:

• Computational Chemistry
• High-Performance Computing
• Scientific Software Development

Background:

• Coupled-cluster (CC) methods are essential for accurate quantum chemistry simulations.
• Efficient tensor contraction is a major bottleneck in CC implementations.
• Leveraging Graphics Processing Units (GPUs) can significantly accelerate these computations.

Purpose of the Study:

• To benchmark Python routines for tensor contraction efficiency.
• To identify optimal tensor operations for coupled-cluster (CC) implementations.
• To accelerate bottleneck tensor operations using GPUs.

Main Methods:

• Benchmarking Python routines for time and memory requirements.
• Utilizing NVIDIA GPUs with the Compute Unified Device Architecture (CUDA) API and CuPy library.
• Implementing batch-wise GPU calculations due to memory limitations.
• Comparing CPU-only and CPU-GPU hybrid implementations.

Main Results:

• CuPy implementation achieved a 10-16x speed-up for bottleneck tensor contractions compared to 36 CPU cores.
• Hybrid CPU-GPU implementations showed a 3-4x speed-up over CPU-only calculations for CCSD and pCCD-LCCSD.
• Identified optimal tensor contraction strategies for GPU acceleration.

Conclusions:

• GPU acceleration using CuPy significantly enhances the performance of Pythonic coupled-cluster calculations.
• Batch-wise processing is a viable strategy to overcome GPU memory limitations.
• Hybrid CPU-GPU approaches offer substantial speed-ups for quantum chemistry simulations.