CRII: OAC: SEAM: Scalable and Efficient Adaptive Mixed-Precision Framework for Scientific and AI Workloads on GPUs

High-performance computing drives groundbreaking advancements in diverse domains by enabling researchers to analyze massive datasets and perform complex simulations with unparalleled speed and precision. SEAM supports these advances by introducing an adaptive mixed-precision framework to optimize performance further for scientific and artificial intelligence workloads across various heterogeneous systems. This approach fosters breakthroughs in fields ranging from climate science to healthcare and stands at the intersection of technological innovation, computational excellence, and practical application. By pushing the boundaries of computational capabilities, SEAM has the potential to benefit society broadly, from accelerating scientific discoveries to enhancing the tools available for national defense and economic competitiveness.

The high-performance computing landscape is significantly evolving in supporting energy-efficient, low-precision computations. This shift has prompted researchers to reassess traditional numerical algorithms, identifying areas where reduced precision can be employed without compromising the overall solution quality. The synergy between hardware advancements and algorithmic optimization is pivotal in addressing complex scientific challenges, particularly on heterogeneous GPU architectures that leverage Tensor Cores. SEAM develops a scalable and efficient adaptive mixed-precision framework tailored for scientific and artificial intelligence workloads on GPUs. This framework applies to fundamental operations like General Matrix Multiplication, Cholesky Decomposition, LU Decomposition, etc. In addition, SEAM is built on top of task-based runtime systems to ensure efficiency, scalability, and portability. It also features Julia interfaces for ease of use and is applied in various domains, including geospatial modeling, genome-wide association studies, and transformer-based foundation models on multiple GPU platforms. SEAM has significant implications for future research and development in high-performance computing, highlighting the transformative potential of mixed-precision arithmetic on GPUs and making progress toward a more sustainable, efficient, and accurate future for scientific and artificial intelligence applications.

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.

Award Number: 2451577
Principal Investigator: Qinglei Cao
Funds Obligated: $174,666
State: MO