Abstract

Accelerating Image Processing by 1000x using CUDA and V100 GPC

The goal of this session is to demonstrate and explain how GPU & CUDA can significantly accelerate image processing algorithms. In order to leverage the amazing power of GPU, the roofline model will be explained, demonstrating how to reduce memory bottlenecks using CUDA by increasing operation intensity(Fusion). 


Session No: SIL8134
Speaker: Eyal Rot
Type:

Intelligent Machines, IoT & Robotics

Date: Thursday - October 18, 2018 03:00 PM - 03:45 PM
Location: Hall F/G
Topics: Autonomous Machines and IoT, AI in Healthcare, Developer Tools
Industry: Architecture

The goal of this sesion is to demonstrate and explain how GPC & CUDA can significantly acelerate image processing algorithms. In order to leverage the amazing power of GPC, the roofline model will be explained, demonstrating how to reduce memory botlenecks using CUDA by increasing operation intensity(Fusion). Several patterns to increase GPC utilization such as: pipeline and streams, custom memory allocators and the use of batches when handling small images wil be described. Finaly, an easy way to start and evaluate V1000 GPC on Amazon wll be shown.