Partager cette page :

Number Format Optimization Problems and How to Solve Them

le 10 avril 2024


Campus de Beaulieu Salle i-51 - bât. 12D

Intervention de Silviu-Ioan Filip, chargé de recherche à l'INRIA Rennes Bretagne Atlantique dans l'équipe Taran à l'IRISA, dans le cadre des séminaires du département Informatique.


We live in an age where it has become paramount to optimize as much as possible the execution of numerical kernels with respect to performance and energy efficiency. This is probably nowhere more obvious today than in the design and deployment of so-called large language models, whose sizes can go up to hundreds of GB and that require training times in the order of days and weeks on powerful cloud hardware. One of the most effective ways of controlling this complexity is through carefully selecting the arithmetic formats and precisions (e.g., IEEE-754 floating-point formats) that are used during computations and memory transfers. Making these choices is however not trivial, leading to interesting numerical optimization and hardware design questions. In this talk I will introduce some of these problems that have shown up in my own work and ways to address them, using applications from Computer Arithmetic, Digital Signal Processing, and Machine Learning.
Formation, Recherche - Valorisation
Killian Barrere

Mise à jour le 30 janvier 2024