Collaborative Research: Frameworks: Framework for Advanced (Multi)Linear Infrastructure in Engineering and Science (FAMLIES)

Discovery in science, machine learning, and artificial intelligence (AI) invariably employs computations with matrices and their higher-dimensional extensions, tensors. Examples include the design of new antibiotics or cancer drugs, the development of quantum computers which can rapidly solve even the harder problems in chemistry and physics, or the construction of novel artificial intelligence architectures which can work more effectively alongside human researchers while reducing energy and hardware costs. These operations typically demand a major part of the computational resources, necessitating the use of fast computers and high-performance software. Convenience and flexibility are also of great importance, as cutting-edge applications often require new functionality and rapid development. The project investigates and delivers a new, adaptable framework for a broad class of matrix and tensor computations, targeting the entire high-performance hardware stack while vertically integrating the software layers. This supports innovation in science and engineering and the application of advanced models to real-world problems. The software is available under open-source license. It is designed to conveniently support existing and future computational tools, while reducing barriers to entry and facilitating the training of the next generation of computational and data scientists. Dense linear algebra software libraries, developed over the past four decades, have had an arguably unparalleled impact on scientific computing and, more recently, machine learning, data science, and AI. While much innovation has happened over this time, the fundamental approach and exported interfaces have changed little. The Framework for Advanced (Multi)Linear Infrastructure in Engineering and Science (FAMLIES) project leverages highly successful prior research and development, sponsored by the National Science Foundation and industry, to develop, design, and deploy a new, vertically integrated dense matrix and tensor software stack. The library targets the entire hardware stack, including single and multi-core, GPU-accelerated, and massively parallel compute environments. It is simultaneously backward compatible via its support of widely used interfaces and forward compatible because it is a framework for synthesizing new functionality. The effort builds on decades of experience by the research team turning fundamental research on the systematic derivation of algorithms into practical software for these domains. This project implements key linear algebra and tensor operations, highlighting the flexibility and effectiveness of the new framework. The software is shared via GitHub, allowing contribution from and dissemination to the broader community. 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. NSF Award ID: 2513929 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Tze Meng Low | Institution: Carnegie Mellon University, PITTSBURGH, PA | Award Amount: $1,395,001 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2513929 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2513929.html