ENGAGE: Energy-efficient Novel Algorithms and Architectures for Graph Learning

To reign in the intractable energy consumption with large AI models, we propose to create a pioneering full-stack co-design methodology for graph learning that is ubiquitous in science applications. We propose novel paradigms of graph learning using Spiking Neural Networks (SNNs) and novel hardware beyond digital systems to reduce energy usage by orders of magnitude in comparison to graph neural (GNNs), without compromising predictive performance. In fact, our methodology may solve some of the algorithmic issues related to GNNs. Our approach innovates through five key strategies: 1) Developing science-aware SNN architectures tailored for graph learning rather than mimicking existing GNN frameworks, 2) Implementing advanced training algorithms that extend beyond traditional gradient-descent methods, such as spike timing dependent plasticity (STDP), to enhance learning efficiency and reduce energy costs, 3) Powerful simulators that bridge the gap between logical models and hardware realities for SNNs for co-design that can optimize for both performance and energy usage, 4) Scaling the simulations to millions of nodes/neurons on leadership computing facilities leveraging deep expertise in exascale computing, and 5) Designing new energy-efficient hardware that surpasses traditional digital complementary metal-oxide semiconductor (CMOS), accommodating common algorithmic patterns in graph learning. Our approach diverges markedly from current popular deep learning practices on GPUs and opens up a new landscape for research on native learning on novel hardware. We aim to establish an open, scalable ecosystem for building highly efficient graph learning algorithms and hardware tailored for scientific applications supported by the Department of Energy (DOE), and we believe the lessons learned can be translated to other AI domains such as foundational models with long-lasting impact for energy-efficient AI systems.