Abstract

Load-And-Go GPU Analytics: Overcoming the I/O Bottleneck for AI

In this session, learn how to efficiently arrange data for consumption by GPUs for analytics. We'll take a look at how combining several modular components and ideas can deliver fast AI performance by limiting the effect of I/O on data-intense queries and models. We'll show how arranging data for the GPU, combined with fast GPU compression, metadata mapping....


Session No: SIL8127
Speakers: Arnon Shimoni, Eyal Hirsch
Type:

Accelerated Data Analytics

Date: Thursday - October 18, 2018 04:00 PM - 04:45 PM
Location: Hall I
Topics: Deep Learning and AI, Accelerated Data Science
Industry: Entertainment

In this session, learn how to efficiently arrange data for consumption by GPUs for analytics. We'll take a look at how combining several modular components and ideas can deliver fast AI performance by limiting the effect of I/O on data-intense queries and models. We'll show how arranging data for the GPU, combined with fast GPU compression, metadata mapping, and several other techniques, can accelerate real database physical operators. We'll also show real examples of how GPU databases, powered by energy-efficient NVIDIA Tesla GPUs, can compete with much larger and more expensive tailored hardware MPP and distributed solutions. See how tools like TensorFlow, Spark MLib, R, and IBM's Watson interact with GPU Databases and benefit from analyzing and training on larger data sets than ever before.