Computation of the Density Matrix in Electronic Structure Theory in Parallel on Multiple Graphics Processing Units.

Area of Science:

• Computational chemistry
• Electronic structure theory
• High-performance computing

Background:

• The computation of the density matrix is crucial in electronic structure theory.
• Previous work established a graphics processing unit (GPU) accelerated algorithm using the second-order spectral projection (SP2) method.
• Scaling this method to multiple GPUs is essential for handling larger and more complex systems.

Purpose of the Study:

• To efficiently parallelize the SP2 density matrix computation algorithm across multiple GPUs on a single compute node.
• To evaluate the performance gains and accuracy of the multi-GPU implementation.
• To compare the parallel GPU-based algorithm with single GPU and traditional CPU-based methods.

Main Methods:

• Parallel implementation of the SP2 algorithm across multiple GPUs.
• Utilizing the second-order spectral projection (SP2) method for density matrix computation.
• Performance and accuracy benchmarking against single GPU and multicore CPU implementations.

Main Results:

• Significant speed-ups were achieved with the parallel multi-GPU implementation compared to the single GPU version.
• The parallel implementation maintained the accuracy of the original SP2 method.
• The GPU-based approach demonstrated superior performance over traditional matrix diagonalization on multicore CPUs.

Conclusions:

• Efficient parallelization of the SP2 algorithm on multiple GPUs is feasible and highly effective.
• This approach offers substantial computational advantages for electronic structure calculations.
• The developed method provides a powerful tool for advancing computational chemistry research.