Performance of Coupled-Cluster Singles and Doubles on Modern Stream Processing Architectures

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• High-Performance Computing

Background:

• Coupled-cluster singles and doubles (CCSD) is a high-accuracy quantum chemistry method.
• Traditional CCSD implementations are computationally demanding, limiting system sizes.
• Advancements in GPU hardware offer potential for accelerating these calculations.

Purpose of the Study:

• To develop and optimize a new CCSD implementation for contemporary GPU architectures.
• To assess the performance and scalability of the new GPU-accelerated CCSD code.
• To compare the efficiency of GPU-based CCSD against traditional CPU-based approaches.

Main Methods:

• Implementation of CCSD utilizing optimized algorithms for NVIDIA V100 GPUs.
• Benchmarking the performance on systems with up to 100 atoms and 1300 basis functions.
• Comparative analysis against established massively parallel CPU implementations.

Main Results:

• A single node with 8 NVIDIA V100 GPUs can complete CCSD computations on ~100 atoms within 24 hours.
• This GPU performance surpasses that of over 64 multi-core CPU nodes.
• Demonstrates significant speedup and efficiency gains through GPU acceleration.

Conclusions:

• The new GPU-optimized CCSD implementation offers a substantial performance improvement.
• This advancement enables accurate quantum chemical calculations on larger systems more efficiently.
• GPU computing represents a powerful paradigm for advancing computational chemistry.