Lecture by Max Breughe: Optimizing deep learning workloads on the GPU

19-12-2018 from 11:00 to 12:00
iGent, meeting room 1.1
Lieven Eeckhout
Add to my calendar

Dr. Max Breughe (NVIDIA) will talk about optimizing deep learning workloads on GPUs.

GPUs are powerful workhorses that made their way from rendering video games into the HPC world with applications in various fields, such as astrophysics and protein folding. Today, the massive parallelism of GPUs enables us to build very complex deep learning (DL) algorithms in fields such as natural language processing (speech recognition, natural language translation), medical imaging (automating tumor detection, etc.), graphics (anti aliasing), automotive (self-driving cars), and many more.
In this talk I will discuss how the Nvidia ecosystem optimizes the inferencing part of deep learning: i.e., once a DL network has been successfully trained, how can it be deployed with maximum performance? Specifically, I will be using a Neural Machine Translation network throughout the talk and will discuss software optimizations, as well as pre-silicon steps involved in creating today's best performing inferencing hardware.

Max works as a Sr. Performance Architect at Nvidia, where he focuses on end-to-end performance optimization of deep learning workloads. In particular, Max is involved in accelerating software libraries for inferencing, such as TensorRT, as well as pre-silicon performance analysis of Nvidia's future GPUs.
In 2014, Max graduated with his PhD in computer architecture at the Ghent University under supervision of prof. Lieven Eeckhout. After his PhD, Max worked as a performance architect on Samsung's Exynos processors in Austin, Texas. He was involved in the development of branch predictors, automation of pre-silicon performance analysis, RTL correlation and was also in charge of an abstract performance simulator.