Single-precision CCSD and CCSD(T) Calculations with Density Fitting Approximations on Graphics Processing Units

It has been reported by our group that using single-precision data and consumer graphics processing units (GPUs) can significantly improve computation speed of CCSD (Coupled-Cluster approaches within the singles and doubles approximation) and CCSD(T) (CCSD approaches augmented by a perturbative treatment of triple excitations). However, CCSD(T) can only be employed for small molecules with about 300 similar to 400 basis functions when using consumer GPUs for acceleration without spatial symmetry due to the memory limitation of GPU. Using density-fitting approximation can significantly reduce the memory requirements in CCSD(T) calculations. In this paper, DF-CCSD(T) codes based on the density fitting approximation together with single precision data was developed. All the matrix contractions were performed employing GEMM in CUBLAS on GPU or in Intel MKL on CPU. The other operations such as matrix expansion and transpose were performed using OpenACC on GPU or OpenMP on CPU. Those codes can be applied to single point energies for systems with around 700 basis functions without spatial symmetry and to molecules with about 1700 basis functions with symmetry on a GPU with 24 Gb memory. The server employed in this work has an Intel I9-10900k CPU and a RTX3090 GPU. CCSD calculations with single-precision data on GPU are about 16 times faster and it is about 40 times faster for the (T) part compared with the calculations on CPU using double precision data on this server. Error introduced by single precision data is negligible. A code library that can employ GPU or CPU using either single precision or double precision data to perform matrix operations with spatial symmetry was also reported in this work. Direct product decomposition (DPD) method was employed to deal with spatial symmetry. Complexity of developing coupled cluster codes with spatial symmetry can be significantly reduced with this library. The computational accuracy of single precision DF-CCSD(T) was compared with CCSD(T)-F12a and DLPNO-CCSD(T). The results shown that DF-CCSD(T) would be more stable than the other two approaches in describing chemical properties.

Automated summary

This study demonstrates that using single-precision data and consumer GPUs can significantly improve the computation speed of CCSD and CCSD(T) approaches. The use of density-fitting approximation can reduce the memory requirements in CCSD(T) calculations. The developed DF-CCSD(T) codes can be applied to different systems on GPUs with specific memory requirements. A code library is also reported, which reduces the complexity of developing coupled cluster codes with spatial symmetry. The single precision DF-CCSD(T) approach is shown to be more stable in describing chemical properties compared to other methods.