Debunking the Myth of 100x GPU vs. CPU
Efficient Training and Prediction of Neural Networks on the CPU
Victor Jakubiuk, Chief Scientist, OnSpecta.com
6:30 Doors Open, Food & Networking
*** Please arrive by 7 PM due to Security ***
This talk attempts to debunk the GPU vs. CPU myth in deep neural networks applications, specifically for the connectomics problem in neuroscience. The talk will discuss how to performance engineer computer vision CNN networks (such as LeNet) with Caffe to achieve inference throughput on the order of 1 TB/hr on a multi-core Intel CPU, without the need for the GPU. The talk will also touch on other, related aspects of CPU engineering of a high-performance visual processing pipeline.
The talk is based on the author's paper published in PPoPP '17 - "A Multicore Path to Connectomics-on-Demand". Link to the paper: http://people.csail.mit.edu/yaronm/PPoPP17_Matveev_Meirovitch.pdf
Victor Jakubiuk is a scientist and an entrepreneur. Victor worked as a research scientist in a computational neuroscience lab at MIT CSAIL, where the talk's technology was developed. Currently he's the Chief Scientist of OnSpecta.com - a Palo Alto-based startup providing a fast CPU-based computational engine for neural networks. Victor holds a B.S. and M.S. from MIT and lives in San Francisco. In his spare time he loves triathlon and can be often spotted swimming in Aquatic Park.