Ramki Kannan is the Distinguished Research and Development Staff Member and Group Leader of the Discrete Algorithms group at Oak Ridge National Laboratory. His research expertise includes distributed machine learning and graph algorithms on HPC platforms, and their application to scientific data—with a focus on accelerating scientific discovery by reducing computation times from weeks to seconds.
He led the Distributed Accelerated Semiring All-Pairs Shortest Path (SNAPSHOT) project—a COVID-19 initiative that was a finalist for the Association for Computing Machinery’s Gordon Bell Award in both 2020 and 2022. Under his leadership, Summit achieved third place on the Graph500 benchmark using minimal resources—the first time an Oak Ridge Leadership Computing Facility system ranked on the Graph500 list.
He also led the team that demonstrated 1 ExaFLOPS on a Knowledge Graph AI application for the first time on Frontier, a project recognized with the UT-Battelle Research Accomplishment Award in 2023. He is currently the co-director of the DOE Mathematical Multifaceted Integrated Capability Center (MMICC) Sparsitute.
Additionally, along with Professor Anuj Karpatne of Virginia Tech and Professor Vipin Kumar of the University of Minnesota, he co-authored the book Knowledge-guided Machine Learning, published in 2022. With over 26 patents issued by the United States Patent and Trademark Office, he was recognized as an International Business Machines Corporation Master Inventor.
Presentation Title:
HyperNeuro: A High Performance Simulator for Neuromorphic Computing
Presentation Abstract:
As data generation accelerates exponentially, the continued training and scaling of deep learning models—such as large language models (LLMs), graph neural networks (GNNs), and vision transformers—is becoming increasingly unsustainable due to high energy consumption. In contrast, Spiking Neural Networks (SNNs) offer a promising, energy-efficient alternative for AI tasks, consuming only a fraction of the energy required by traditional models.
SNN training often relies on biologically inspired, local learning rules such as Spike-Timing-Dependent Plasticity (STDP). However, STDP-based training has not yet demonstrated scalability to networks with billions of neurons.
In this paper, we introduce HyperNeuro, an SNN simulator with native support for STDP-based learning. HyperNeuro employs efficient SpMV (sparse matrix–vector multiplication) and SADDMM (sparse-add-dense matrix–matrix multiplication) kernels to model both SNN dynamics and synaptic updates. We implement 1D and 2D variants of these kernels and scale the simulator to 8,112 nodes on the Frontier supercomputer, successfully simulating 33 billion neurons and 33 trillion synapses in 78.36 seconds using 3.5 PB of memory, achieving 65% parallel efficiency.