GPU acceleration of local and semilocal density functional calculations in the SPARC electronic structure code

Area of Science:

• Computational materials science
• Quantum chemistry
• High-performance computing

Background:

• Density functional theory (DFT) is crucial for electronic structure calculations.
• Real-space methods like SPARC offer advantages for complex systems.
• Computational cost limits the scale of DFT simulations.

Purpose of the Study:

• To develop and implement a Graphics Processing Unit (GPU)-accelerated version of the real-space SPARC code.
• To enhance the performance of Kohn-Sham density functional theory (KS-DFT) calculations.
• To reduce the computational resources and time required for electronic structure simulations.

Main Methods:

• Developed a modular, math-kernel-based implementation for NVIDIA GPUs.
• Offloaded computationally intensive operations to GPUs, retaining others on Central Processing Units (CPUs).
• Utilized representative bulk and slab models for performance evaluation.

Main Results:

• Achieved speedups of up to 6x in node hours and 60x in core hours compared to CPU-only execution.
• Reduced calculation time to under 30 seconds for a metallic system exceeding 14,000 electrons.
• Demonstrated significant reductions in computational resource requirements.

Conclusions:

• GPU acceleration of the SPARC code dramatically improves computational efficiency for DFT.
• The developed implementation enables faster and more resource-efficient electronic structure calculations.
• This advancement facilitates larger and more complex materials science simulations.