Mixed Feelings About Mixed Precisions: Birds of a Feather at SC22!

Different formats for mixed-precision floating-point arithmetics

Abstract:

Come and share your experiences with the state-of-the-art of mixed-precision techniques! By wisely trading off accuracy, we can mitigate data movement overheads and increase performance of applications, including real-time adaptive optics simulations on ground-based telescopes and genome-wide association study for agricultural genomics.


Very Large Telescope
Deploying the MAVIS instrument on the Very Large Telescope with adaptive optics to outsmart the atmospheric turbulence and reveal the secrets of the Universe. (Credits: European Organisation for Astronomical Research in the Southern Hemisphere)

 

Wheat
Running Genome-Wide Association Study (GWAS) to "engineer the resistance of wild wheat into elite wheat cultivars to create crops that are genetically protected, without the need for pesticides." (Credits: Prof. Wulff, Center for Desert Agriculture, KAUST)

No free lunch, however: these optimizations require support from the software/hardware ecosystem and strong numerical validation. For instance, the tile Cholesky factorization is a fundamental matrix algorithm in solving covariance problems that capture strong and weak correlations between variables of interests (e.g., single nucleotide polymorphisms). Precision arithemetic can be determined using a tile-centric approach based on a mathematical formulae, while runtime systems can be in charge of scheduling the computational tasks and converting on-the-fly to the required precisions.

Mixed-precision Cholesky factorization
Dense Cholesky factorization || Mixed-precision band Cholesky factorization || Adaptive mixed-precision Cholesky factorization

This BoF invites the HPC community at large interested in applying mixed precisions into their workflows. We will briefly provide the context on this timely topic. We will engage with the audience via a set of questions and request experts from scientific applications/software libraries/hardware architectures (i.e., from national labs, academic, and vendors) to share their own perspectives. We will eventually gather feedback to define a roadmap moving forward.

 

Organizers:

- Hatem Ltaief, Principal Research Scientist, ECRC, KAUST

- Piotr Luszczek, Research Assistant Professor, ICL, University of Tennessee Knoxville  

 

SC22 BoF Date, Time and Location:

Tuesday, November 15th 12:15pm-1:15pm

Room D-172