Recent advances in graphics processing units (GPUs) have substantially broadened the applicability of computational chemistry. Nevertheless, the double-precision (FP64) operations conventionally required in quantum chemistry remain inefficient on artificial intelligence (AI)-focused GPUs, which are predominantly optimized for lower precision arithmetic. Moreover, the limited memory capacity of such GPUs necessitates algorithmic adaptations to traditional approaches in density functional theory (DFT). In this work, we propose a systematic mixed precision strategy for the iterative matrix diagonalization procedure, which constitutes the principal bottleneck in real-space DFT calculations. By selectively employing single-precision (FP32) in most computational steps and even adopting brain floating point (BF16) in the preconditioning stage, we demonstrate that numerical accuracy can be preserved with negligible degradation. Building on these observations, we developed a mixed precision eigensolver and validated its performance across material systems spanning a wide range of sizes. Our results show that the proposed method achieves up to 10× speedup in the diagonalization step relative to FP64, while extending the feasible system size by approximately 50%. Comprehensive validation across multiple GPU architectures further confirms that AI-focused GPUs can deliver performance comparable to that of high-end high-performance computing (HPC)-focused GPUs, thereby substantially improving accessibility to large-scale electronic structure simulations.
Woo et al. (Wed,) studied this question.