CSR: Small: Unlocking Extreme Performance Scaling in Sparsified Graph Neural Network Models

Modern Artificial Intelligence (AI)-based applications, such autonomous systems, traffic forecasting, social media, drug discovery, and chip design handle increasingly large and evolving graph-based data. Efficient processing of graph-based problems is challenging because they involve complex mathematical operations that incur performance overhead on hardware processing units. Researchers have recently leveraged methods that reduce this computational complexity via pruning redundant computational tasks, but many challenges related to computer memory and task parallelism persist. This project devises novel mathematical operators that address the bottlenecks of graph-based AI applications to increase their performance. The project also develops computer science curriculum and provides student training through integration of research results and education. Moreover, through semiconductor industry collaborations, the project engages in disseminating its research outcomes, ensuring practical adoption and deployment of emerging AI applications, such as semiconductor chip design and autonomous systems, thus improving the U.S. AI infrastructure, with significant benefits to the economy and society. Efficient processing of graph models is challenging since the underlying computations require graph-proportional matrix operators. The strong input graph dependence has led to performance scaling and sustainability challenges for massively parallel hardware processing units. Although the research community has been attempting to reduce the computational complexity of graph processing operations by introducing sparsity in the model inputs, the resulting graph-proportional operators face underutilization of vector-level parallelism, data locality, and indirect memory access patterns, resulting in diminished hardware parallelism. The aggressively sparsified matrix operators exacerbate the computational structures and patterns in already unstructured and ultra-sparse inputs. This project devises novel matrix operators tailored to efficiently exploit extreme sparsity on highly vectorized and high-core-count processors to unlock sustainable and scalable performance. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.