Within the scientific community the need for greater computational power is always increasing. As processor chip technology approaches the limit of Moore’s Law, it is becoming more and more difficult for computers to keep up with this demand. Gone are the days of increasing transistor counts in exchange for a satisfying boost to computer performance.
GPU acceleration has emerged as a viable alternative by offloading compute-intensive, highly parallelisable portions of code to dedicated hardware accelerators. However such improvements often require significant changes to the source code in order to take full advantage of the GPU capabilities; code needs to be ported, algorithms need to be redesigned and continual optimisation tweaks are required to eke out maximum possible performance.
At the Hartree Centre, we work on integrating GPU acceleration with existing high-performance computing codes to provide maximum computational performance on the latest IBM Power hardware. This enables exploration of more complex scientific problems, increasing the viable problem sizes and decreasing time to completion.