Accelerating CCSD(T) on Graphical Processing Units (GPUs)

TL;DR

This paper introduces a GPU-accelerated implementation of CCSD(T) in TeraChem, significantly reducing computation time and enabling high-accuracy quantum chemical calculations on larger systems.

Contribution

The paper presents a novel GPU-accelerated CCSD(T) implementation that achieves state-of-the-art performance and broadens the scope of feasible high-level quantum calculations.

Findings

Calculates (T) correction for 63-atom systems in under 8 hours.

Enables rapid CCSD(T)/CBS stacking energy calculations for DNA base pairs.

Achieves significant speedup over previous CPU-based implementations.

Abstract

Coupled cluster with singles, doubles and perturbative triples (CCSD(T)) often provides ground state correlation energies within "chemical accuracy," but suffers from high computational cost and steep scaling with system size. We present a GPU-accelerated implementation of CCSD(T) in the TeraChem software package. The new implementation achieves state-of-the-art performance, enabling the calculation of the (T) correction for a system with 63 atoms and more than 1000 basis functions in a little under 8 hours on a single node. Additionally, we demonstrate the utility of our optimized implementation for the rapid calculation of full CCSD(T)/CBS stacking energies for all ten unique DNA base pair stacked tetramers. We expect that the TeraChem CCSD(T) implementation will enable the rapid calculation of high-level data that was not previously accessible in a reasonable timeframe.