Accelerating Quantum Chemistry with Emulated Double-Precision Matrix Multiplication on GPUs
Giuseppe M. J. Barca, Cole Brower
Abstract
The evaluation of exact exchange and the treatment of electron correlation remain major computational challenges in electronic-structure theory. Over the past decade, GPU computing has substantially accelerated these calculations, extending the scale and complexity of electronic-structure problems that can be addressed computationally. More recently, GPU architectures developed to meet the demands of artificial intelligence have introduced new opportunities for acceleration, with Tensor Cores providing reduced-precision arithmetic throughput far exceeding that of native double-precision (FP64) operations. Exploiting this performance advantage while preserving the numerical accuracy required in quantum chemistry remains a significant challenge. Here, we investigate the Ozaki-I and Ozaki-II floating-point emulation schemes in the RI-HF, RI-MP2, and RI-CCSD implementations of EXESS on NVIDIA GB200 GPUs. We demonstrate that Ozaki-II at 52-bit precision achieves numerical accuracy comparable to native FP64, delivering complete calculation speedups of up to 1.26× (cc-pVDZ) and 1.50× (cc-pVTZ) for RI-HF, 2.55× and 3.01× for RI-MP2, and 3.11× and 2.15× for RI-CCSD. The energy to solution improves by 1.22× and 2.21× for the representative RI-HF and RI-MP2 cc- pVTZ calculations and by 2.37× for the RI- CCSD cc-pVDZ calculation. Reducing the precision to 47 bits provides further acceleration while maintaining total energy deviations below 0.1 kcal mol -1 across all systems examined.