Real-Space Density Functional Theory on Graphical Processing Units: Computational Approach and Comparison to Gaussian...

Area of Science:

• Computational Physics
• Materials Science
• Quantum Chemistry

Background:

• Density Functional Theory (DFT) is a cornerstone for simulating material properties.
• Real-space DFT methods offer an alternative to traditional basis set approaches.
• Accelerating computationally intensive DFT calculations is crucial for advancing materials discovery.

Purpose of the Study:

• To investigate the application of graphical processing units (GPUs) for accelerating real-space DFT calculations.
• To develop and implement an efficient data parallelism scheme for real-space DFT on GPUs.
• To evaluate the performance of the implemented scheme on current-generation GPUs.

Main Methods:

• Developed a data parallelism scheme tailored for real-space DFT procedures.
• Implemented the scheme within the Octopus open-source code.
• Benchmarked performance on AMD and Nvidia GPUs, comparing against CPU implementations and GPU Gaussian basis set codes.

Main Results:

• Achieved sustained performance of up to 90 GFlops on a single GPU.
• Demonstrated significant speed-up compared to the CPU version of the Octopus code.
• Showed that the real-space approach on GPUs can outperform GPU Gaussian basis set codes for certain systems.

Conclusions:

• The developed data parallelism scheme effectively accelerates real-space DFT calculations on GPUs.
• The Octopus code with this implementation offers a competitive and efficient platform for GPU-accelerated DFT simulations.
• Real-space DFT on GPUs presents a viable and powerful alternative for large-scale electronic structure calculations.