Accelerating Large-Scale Atomistic td-DFTB Simulations Using GPU Offloading�With the DECODE Partnership (BES)
Scientific Achievement
Significance and Impact
Technical Approach
The computational time for a single time step is shown for four stages of the acceleration process. In the initial build, there is no GPU acceleration and the computation is dominated by slow matrix multiplication. A basic GPU offload radically speeds up the computation of the matrix multiplication but introduces some additional overhead. Optimizing this offloading reduces the overhead and results in a speed-up by a factor of 10 over the baseline implementation.
We accelerated time-dependent density functional tight binding (td-DFTB) simulations — a method used to study the electronic structure and properties of molecules and materials — by a factor of 10. Our new implementation efficiently leverages GPU resources to run simulations of large condensed matter systems containing thousands of atoms with favorable computational scaling as a function of system size.
PI(s)/Facility Lead(s): Mauro Del Ben (LBNL), Khaled Ibrahim (LBNL), Lenny Oliker (LBNL)
Collaborating Institutions: UC Riverside
ASCR Program: SciDAC
ASCR PM: Hal Finkel
Publication(s) for this work: Qiang Xu et al., “Velocity-Gauge Real-Time Time-Dependent Density Functional Tight-Binding for Large-Scale Condensed Matter Systems”, J. Chem. Theory Comput. 2023, 19, 22, 7989–7997, https://doi.org/10.1021/acs.jctc.3c00689
Despite its broad applicability, the high computational expense of standard time-dependent density functional theory prohibits its use for large systems. The computational efficiency of our td-DFTB implementation enables the simulation of electron dynamics of complex systems that are too large to handle otherwise, allowing researchers to investigate how complex systems like photoactive proteins or catalytic materials behave in their full chemical environments over long simulation windows.
�.