skip to main content

Illinois ECE team named IEEE HPEC GraphChallenge Champions

10/21/2020

Allie Arp, CSL

Mert Hidayetoglu
Mert Hidayetoglu

Each year at the IEEE High Performance Extreme Computing Conference (HPEC), teams from academia, national laboratories, and industry pit big data graph analysis software systems they’ve developed against each other. In this year’s Sparse Deep Neural Network (SpDNN) Graph Challenge task, members of an Illinois ECE team were named champions.

The Sparse DNN Challenge draws upon the community’s latest understanding of issues surrounding machine learning and high-performance computing. It is reflective of the emerging sparse computation patterns of modern artificial intelligence solutions and their demanding requirements of current high-performance computing systems.

“We deal with algorithms every day in our research group and sometimes we get really frustrated and think we aren’t good enough with our small group in CSL,” said Illinois ECE  graduate student Mert Hidayetoglu, the team lead. “When we win this kind of competition, we are forced out of our bubble, and we realize we’re better than we thought, and that there are people who support us.”

Jinjun Xiong
Jinjun Xiong

The team’s winning entry builds on members’ deep expertise in computer architecture and graphics processing units (GPUs), algorithmic design, performance analysis, and rigorous implementation of high performing AI solutions. A key innovation of the new algorithms is the carefully crafted kernels and data layouts that fully utilize the on-chip memory bandwidth, which improves the throughput while reducing the required energy.

In order to demonstrate their algorithms, the group implemented them on the Oak Ridge National Lab’s Summit supercomputer (developed by IBM). The algorithms demonstrated an at-scale 180 TeraEdges/Second sustained inference throughput. A more detailed analysis, including performance benchmarking on 12 sparse deep neural network models of various sizes, can be found in their paper, “At-scale sparse deep neural network inference with efficient GPU implementation.”

“We have been participating in the HPEC Graph Challenge for the past four years on different challenge tasks,” said Illinois ECE Adjunct Research Professor Jinjun Xiong, IBM researcher and co-director of the IBM-ILLINOIS Center for Cognitive Computing Systems Research (C3SR), with which most of the team members are affiliated. “This year’s winning as Champion in the SpDNN task is really a culmination of our Center’s years of persistence in understanding the computation bottlenecks of modern AI and big graph analytics workloads, in addition to our relentless efforts to optimize systems and algorithms for highest possible performance.”

One unique trait that set the team apart from the rest crowd, Hidayetoglu believes, was the team’s commitment to test their innovation on one of the world’s most powerful supercomputers.

“It shows the algorithms we develop are battlefield tested,” Hidayetoglu said. “This international recognition shows we are doing state-of-the-art work. This gives more visibility and publicity to our Center.”

Wen-Mei W Hwu
Wen-Mei W Hwu

In addition to the influence of C3SR from several students and Illinois ECE Professor Emeritus Wen-Mei W Hwu, AMD Jerry Sanders Chair of Electrical and Computer Engineering, the team also had contributions from partners at IBM and NVIDIA that helped the team develop their now award-winning algorithms.

“What really excites me about winning is that it reinforces our belief that true innovation requires a diverse range of expertise and talents,” said Xiong. “And our C3SR center is an ideal place for that to happen.”

Other team members include Illinois ECE Affiliate Faculty Rakesh Nagi, industrial and enterprise systems engineering; Illinois ECE graduate students Carl William Pearson and Vikram Sharma Mailthody, and Eiman Ebrahimi, NVIDIA. The challenge is sponsored in part by IEEE, MIT, and Amazon. Hwu and Nagi are both affiliated with the CSL.

 

Check out the original article on the CSL site