know-how

How we did it?

GPU

In the upcoming years the number of cores in computing devices will scale exponentially opposing the trend in the last years where the number of transistors incereased in a core. This requires new concepts to include into scientific (and other productional) software development. The experiments for the usage of GPU's in quantum chamistry has a long history in the last ten or so years. Our technology is unique ina way that.... 


Compiler technology

Quantum chemical integral calculation on Gaussian orbitals have a large number of slightly different expressions to evaluate. GPU's have a somewhat similar but in some parameteres different architecture. To optimize for GPU's well and to compile qc code, wa had to implement a full compiler framework to overcome the problems others faced in the last yaers. This made it possible to calculate on high angular momentum orbitals which was a neccessary prerequisite to user real world basis sets (e.g. cc-pVTZ).


Meta algorithm and machine learning.

BRUSH is an integral driver algorithm (like the well-known PRISM) it uses all the industry standard integral reduction algorithms (McMurchie-Davidson, Obara-Saika, HGP) and the quadrature based Rys algorithm for higher angular momentum orbitals. It uses a machine learning based (offline) benchmarking and training approach to choose the most efficient integral algorithm for a given quartet type (angular momentum and degree of contraction) for every GPU architecture we support. 
This machine learning and hardware specific approach means our technology edge causing magnitudes of speedup due to high variability of GPU architectures.