Accelerating finite-element-based projector augmented-wave density functional theory calculations with scalable GPU-centric computational methods

TL;DR

This paper introduces GPU-optimized finite-element PAW methods enabling scalable, accurate large-scale DFT calculations on exascale systems, significantly reducing computation time compared to traditional approaches.

Contribution

The authors develop a GPU-centric finite-element PAW DFT implementation with novel algorithms and mixed-precision strategies for scalable, accurate electronic-structure calculations on exascale architectures.

Findings

Achieves up to 8x and 20x CPU-GPU speedups on Intel and AMD GPUs.

Reduces time-to-solution by nearly 8x for 10,000-electron systems compared to plane-wave PAW methods.

Demonstrates scalability to 130,000-electron systems, suitable for exascale computing.

Abstract

Accurate large-scale Kohn-Sham density functional theory (DFT) calculations are essential for modeling complex material systems, including interfaces, defects, nanoclusters, and twisted two-dimensional heterostructures. Achieving chemical accuracy at scales of $10^4$-$10^5$ electrons with practical time-to-solution, however, remains challenging for existing DFT implementations. We present GPU-centric computational methods and algorithmic innovations within a finite-element (FE) discretized projector augmented-wave (PAW) formulation (PAW-FE) for accurate, efficient, and scalable electronic-structure calculations on modern exascale systems. The FE discretization, developed within a collinear spin formalism, accommodates generic boundary conditions and employs multi-resolution quadrature for accurate evaluation of atom-centered PAW integrals on coarse grids. The resulting generalized Hermitian eigenproblem is solved using residual-based Chebyshev filtered subspace iteration (R-ChFSI). Exploiting R-ChFSI's tolerance to inexact matrix-multivector products, we employ an approximate inverse PAW overlap matrix, mixed-precision arithmetic (FP32/TF32), and low-precision nearest-neighbor communication (BF16) during filtered subspace construction, along with block-wise computation-communication overlap to reduce cost while preserving robustness. These strategies yield up to $8\times$ and $20\times$ CPU-GPU speedups on Intel and AMD GPU architectures, respectively. Compared to plane-wave PAW methods, PAW-FE achieves close to 8$\times$ reduction in time-to-solution for 10,000-electron systems on NVIDIA GPUs, with larger gains at scale, and around 6$\times$ over norm-conserving FE approaches. We demonstrate scalability to 130,000-electron systems, establishing PAW-FE as an exascale-ready method for chemically accurate first-principles simulations.