Reducing Numerical Precision Requirements in Electronic Structure Calculations

The growing demand for deep learning compute resources has lead to the development of new kinds of low precision hardware. This talk explores how electronic structure calculation software can coexist on these AI-centric platforms. We will begin by examining in detail the numerical precision requirements of key computational kernels (including density functional theory and wavefunction methods). We will demonstrate that double precision often provides far more precision than is necessary. Then, we will present an approximation built from an error-free transformation of matrix multiplication that can exploit these relaxed requirements. This approximation will allow us to run calculations on high performance low precision hardware while obtaining the required precision (and no more).