Title: Evolution of Graph Neural Network Accelerators
Date: Wednesday, May 31, 2023, 10:30am.
Location: HBL 1102
Abstract: Presented by Kaustubh Shivdikar, this talk promises a deep exploration into the intriguing world of Graph Neural Networks (GNNs) and the concerted efforts being made to accelerate their operations. Showcasing the progression and transformative influence of GNNs in modern machine learning computing, Kaustubh brings a wealth of understanding from his cutting-edge research at Northeastern University.
In the initial phase of the talk, Kaustubh will address the architectural impact of GNN workloads across a multitude of platforms. This discussion will provide an extensive comprehension of how GNNs shape computational structures and influence operational efficiencies.
Delving further, the presentation will elucidate the intricate mechanisms of sparse matrix multiplication, focusing on the low-level implementations within the Basic Linear Algebra Subprograms (BLAS) library. These insights will highlight the significance of efficient sparse matrix operations in improving the performance of GNNs.
The final section of the talk will turn the spotlight on the microarchitectural enhancements being brought to state-of-the-art GPUs to propel GNN workloads. Kaustubh will expound on the groundbreaking advancements being made to augment GPU architectures, leading to faster and more efficient GNN operations.
Attendees of this talk will gain a comprehensive understanding of the current landscape of GNN accelerators, the key role of sparse matrix multiplication, and the forward-looking enhancements being made to GPUs. This presentation offers a glance into the exciting future of machine learning computing through the lens of GNN acceleration.
Bio: Kaustubh Shivdikar is a Ph.D. candidate in computer engineering at Northeastern University. His pioneering work in the NUCAR (Northeastern University Computer Architecture Research) lab, under the mentorship of Professor David Kaeli, is at the forefront of computer technology.
Kaustubh is primarily focused on creating and developing hardware accelerators for sparse graph workloads.
His proficiency extends across several niche sectors within computer engineering. Kaustubh’s work extends in the domains of Simulator Design, GNN Accelerators, Sparse Matrix Accelerators and recently, Homomorphic Encryption Accelerators.