Person:

Edgar, Richard G

Two new tools for the acceleration of computational chemistry codes using graphical processing units (GPUs) are presented. First, we propose a general black-box approach for the efficient GPU acceleration of matrix−matrix multiplications where the matrix size is too large for the whole computation to be held in the GPU’s onboard memory. Second, we show how to improve the accuracy of matrix multiplications when using only single-precision GPU devices by proposing a heterogeneous computing model, whereby single- and double-precision operations are evaluated in a mixed fashion on the GPU and central processing unit, respectively. The utility of the library is illustrated for quantum chemistry with application to the acceleration of resolution-of-the-identity second-order Møller−Plesset perturbation theory calculations for molecules, which we were previously unable to treat. In particular, for the 168-atom valinomycin molecule in a cc-pVDZ basis set, we observed speedups of 13.8, 7.8, and 10.1 times for single-, double- and mixed-precision general matrix multiply (SGEMM, DGEMM, and MGEMM), respectively. The corresponding errors in the correlation energy were reduced from −10.0 to −1.2 kcal mol(^{-1}) for SGEMM and MGEMM, respectively, while higher accuracy can be easily achieved with a different choice of cutoff parameter.

The Murchison Widefield Array (MWA) is a next-generation radio telescope currently under construction in the remote Western Australia Outback. Raw data will be generated continuously at 5 GiB s(^{−1}), grouped into 8 s cadences. This high throughput motivates the development of on-site, real time processing and reduction in preference to archiving, transport and off-line processing. Each batch of 8 s data must be completely reduced before the next batch arrives. Maintaining real time operation will require a sustained performance of around 2.5 TFLOP s(^{−1}) (including convolutions, FFTs, interpolations and matrix multiplications). We describe a scalable heterogeneous computing pipeline implementation, exploiting both the high computing density and FLOP-per-Watt ratio of modern GPUs. The architecture is highly parallel within and across nodes, with all major processing elements performed by GPUs. Necessary scatter-gather operations along the pipeline are loosely synchronized between the nodes hosting the GPUs. The MWA will be a frontier scientific instrument and a pathfinder for planned peta- and exascale facilities.