ECE graduate student Dian-Lun Lin has received the Champion Award in the prestigious MIT/Amazon Graph Challenge at the 2020 High-performance Extreme Computing Conference. The graph challenge aims to create an open environment reflective of emerging sparse Artificial Intelligence systems on high-performance computing platforms.
The paper, “A Novel Inference Algorithm for Large Sparse Neural Network using Task Graph Parallelism,” introduces a novel parallel approach to accelerate the inference of large-scale machine learning workloads up to 100 GB. Lin’s approach leverages a task graph-based parallel infrastructure with specially designed kernels to overcome the computational challenges of large sparse neural networks. The result scales to arbitrary model and data sizes under different numbers of graphics processing units.
“Because of this task graph technique, I can implement different parallel decomposition strategies and kernel algorithms to efficiently compute large-scale machine learning workloads,” said Dian-Lun. “It is really promising and we are extending our algorithms to a general-purpose task processing system that can benefit broader parallel applications.”
Advised by ECE Professor Tsung-Wei Huang, Lin will continue his research on parallel computing systems and applications as he works towards his Ph.D.