SHF: Precision Emulation on AI Chips (PEACH): Challenges, Opportunities, and Impact on Scientific Computing

Current trends in Artificial Intelligence (AI) are driving changes in the hardware available on the market for general purpose computing. One major change is the increased emphasis on performance at relatively low accuracy, which is a typical setup for the vast majority of AI algorithms. However, this change will affect other scientific domains, often requiring highly accurate simulations of natural phenomena. As a result, scientific software can either evolve to benefit from the new hardware design trends, or push back against prioritizing reduced accuracy at the hardware level. Some research efforts investigate novel algorithms for maintaining high accuracy on hardware primarily designed for AI workloads. Such efforts are called "emulation algorithms". This project aims to study the impact of emulation algorithms on scientific software requiring high-precision computations, and whether emulation can completely replace native computations at high accuracy. It seeks to define the performance-accuracy tradeoff on several numerical linear algebra algorithms and provide insights on the numerical behavior of emulation on various workloads that could expose shortcomings, ultimately leading to potential improvements over the current state of the art. The project also explores possible opportunities for improving existing algorithms using emulation, such as mixed-precision linear solvers. Graphics Processing Units (GPUs) are the primary focus of this project, as they dominate the AI market. The projects seeks the effective use of hardware accelerators for low-precision matrix multiplications, such as NVIDIA's Tensor Core technology, in various compute workloads spanning linear solvers, eigenvalue and SVD problems, batch linear algebra, and other algorithms. 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: 2551234 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Ahmad Ahmad | Institution: University of Tennessee Knoxville, KNOXVILLE, TN | Award Amount: $800,000 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2551234 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2551234.html