Quantum Chemical Calculations Using Accelerators: Migrating Matrix Operations to the NVIDIA Kepler GPU and the Intel...

Area of Science:

• Computational Chemistry
• High-Performance Computing
• Computer Architecture

Background:

• Modern computers utilize heterogeneous architectures with general-purpose processors and specialized accelerators (e.g., NVIDIA Kepler GPU, Intel Phi).
• Accelerators provide significant performance advantages over host processors, but integrating legacy applications into these environments is complex.
• Matrix operations, particularly double-precision general matrix multiply, are fundamental to electronic structure calculations in quantum chemistry.

Purpose of the Study:

• To investigate the migration of matrix operations from quantum chemical calculations to GPU and Intel Phi accelerators.
• To explore adaptive strategies for utilizing host processors and accelerators based on problem size.
• To evaluate the performance of GPUs and Intel Phi for matrix multiplications in computational chemistry.

Main Methods:

• Implementation of double precision general matrix multiply operations for quantum chemical calculations on GPU and Phi platforms.
• Development of a system to dynamically allocate computations between the host processor and accelerators.
• Performance benchmarking of data transfer rates over PCI-e and matrix multiplication efficiency.

Main Results:

• The NVIDIA Kepler GPU demonstrated superior performance for both square and non-square matrix multiplications compared to the Intel Phi.
• For data transfers over PCI-e, the GPU achieved optimal performance for data sizes up to 4096 MB, with high upload (5-5.6 GB/s) and download (5.4-6.3 GB/s) rates.
• The adaptive approach successfully managed computations across host and accelerator resources.

Conclusions:

• GPUs are highly effective accelerators for computationally intensive matrix operations in quantum chemistry, outperforming Intel Phi.
• Efficient data transfer mechanisms are crucial for maximizing performance in heterogeneous computing environments.
• Further research into optimizing legacy code migration for accelerators is warranted for advancing computational chemistry capabilities.