Computational Efficiency and Performance Crosscut

Computational efficiency and performance are cornerstones of FASTMath. The Department of Energy has invested heavily in exascale high performance computing resources and continues to invest in more compute for the Genesis Mission.  The Computational Efficiency and Performance crosscut ensures that our algorithms and software enable application codes to fully utilize these resources while minimizing energy waste from unnecessary data movement and computation. We are organized around three themes which are broadly applicable in all our topic areas:

• Architecture awareness ensures we can take complete advantage of existing CPU/GPU supercomputers at DOE LCFs, and that we are poised to utilize new features coming in next-generation GPUs as well as emerging AI accelerators. Our team has experience working with hardware vendors to optimize algorithms and software for maximum efficiency and performance.

• Performance awareness is about taking an integrated approach to measuring performance, communicating performance data, and acting on performance data in an intelligent way. This spans continuous performance testing, use of autotuning tools, and researching AI-based approaches to dynamically steering algorithmic choices based on performance and problem data in situ.

• Mixed-precision algorithms and software deliver flexibility for application codes to reduce memory usage, data movement, and leverage lower precision hardware for increased throughput without sacrificing overall accuracy.  We are exploring mixed precision algorithms across space and time discretizations as well as our solvers.