SHF: Medium: Panther: New Highly Parallel Hardware and Software System for Graph-based Data Analytics to Improve Artificial Intelligence Reliability and Efficiency

Modern artificial intelligence (AI) techniques, including large language models (LLMs) like ChatGPT, have brought many benefits to our society, ushering in a new age of increased productivity and information accessibility. Despite these dramatic technological advances, modern AI techniques exhibit several drawbacks that limit their applicability and usability in many domains. They are trained on data that can easily become out-of-date in a world that currently generates over 400 million terabytes of new data every day. Modern AIs may generate answers that are difficult to explain or validate, and are therefore hard to trust. Finally, AIs are famously vulnerable to “hallucination,” producing answers that are simply wrong. The goal of this research is to address these shortcomings with a new computer system and software framework that enables efficient improvements to the reliability and applicability of modern AI. AI’s vulnerability to hallucination can be reduced using techniques that augment the context available to the AI with knowledge graphs. A knowledge graph is a way of representing information that represents not just individual data items, but connections between them. Graphs encode structure, hierarchy, and complex relationships, which, if accessible to an AI, can improve the correctness of its answers; at the same time, graphs can provide additional context which can help explain or validate answers, improving explainability. This research is necessary because current computer architectures and distributed computing platforms are not well suited to simultaneously supporting both LLM and large-scale graph computations. The GPU architectures that currently dominate AI are optimized for computations in which all data are laid out in a very regular, dense pattern, while graph computations have historically required a very different kind of optimization to support irregular data layouts. Additionally, advances in software are necessary to support multiple kinds of graph computations over distributed data to query the structure of graphs, analyze them, and make predictions based on those structures and analyses. This research will produce a system called Panther, that comprises a new, highly parallel architecture well-suited to both LLM and graph computations, a new memory system that efficiently supports the combination of large scale graphs and LLMs, and a distributed software framework and applications that collectively realize dramatic improvements for AI efficiency and reliability. We expect Panther to lay the groundwork for the next generation of high-performance trustworthy AI. 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.