Leveraging Mixed Precision and Emerging Low Precision Operators in HPC Application
Birds of a Feather
AI/Machine Learning/Deep Learning
Performance Analysis and Optimization
TimeWednesday, June 19th1:45pm - 2:45pm CEST
DescriptionWith ever-increasing need for computation of scientific applications, major energy constraints, and new application domains such as machine learning, general purpose processor arithmetic is entering a new burst of evolution. Novel hardware and software solutions are explored to leverage emerging low precision operators. Indeed, a better fit in precision usage can provide higher flops, higher energy efficiency, lower communication bandwidth and volume needs. To fully exploit this potential optimization, all the application design ecosystem needs profound evolution with new numerical algorithms, low and mixed precision libraries, and dedicated tools for validating, optimizing, and debugging precision issues. This BoF session will present four technical contributions to this topic on different aspects: AI, linear algebra libraries, tools, and applications. This presentation will be followed by an interactive round-table with the speakers.
Research Engineer Royal Society University Research Fellow at ECMWF