GPU acceleration of hybrid functional calculations in the SPARC electronic structure code.

Area of Science:

• Computational physics
• Materials science
• Quantum chemistry

Background:

• Electronic structure calculations are crucial for understanding material properties.
• Hybrid functional calculations in density functional theory are computationally demanding.
• Accelerating these calculations can enable larger and more complex simulations.

Purpose of the Study:

• To develop a Graphics Processing Unit (GPU)-accelerated version of the SPARC electronic structure code.
• To implement a batch variant of the Kronecker product-based linear solver for efficient hybrid functional calculations.
• To optimize the code for NVIDIA GPU architectures.

Main Methods:

• Developed a GPU-accelerated real-space SPARC code.
• Implemented a batch variant of the Kronecker product-based linear solver.
• Created a modular, math kernel-based implementation for hybrid functionals, offloading intensive operations to GPUs and remaining workload to CPUs.

Main Results:

• Achieved up to 8x speedup in node-hours and 80x in core-hours compared to CPU-only execution.
• Reduced time to solution to approximately 300 seconds for a metallic system with over 6000 electrons on V100 GPUs.
• Significantly decreased computational resource requirements for a given wall time.

Conclusions:

• GPU acceleration offers substantial performance gains for hybrid functional calculations in density functional theory.
• The developed code enables faster and more resource-efficient electronic structure simulations.
• This advancement can facilitate larger-scale materials modeling and discovery.