Mixed-Precision GPU Acceleration of RTE-RRTMGP-CPP: Towards Radiative Transfer Simulations on Exascale Supercomputers

Procedia Computer Science, vol. 286, pp. 95–105

Abstract

Radiative transfer is a fundamental component of weather and climate modeling, often accounting for a significant portion of total simulation time in Earth System Models (ESMs). RTE+RRTMGP-CPP is a C++ frontend as well as a CUDA-based implementation of the Radiative Transfer for Energetics (RTE) solver and the Rapid Radiative Transfer Model (RRTMGP), both widely used to compute longwave and shortwave radiative transfer in vertical columns of atmospheric models. However, despite existing GPU acceleration, high computational intensity remains a bottleneck for scalability on modern supercomputers. In this work, we present a scalable, mixed-precision implementation of RTE-RRTMGP-CPP. Driven by recent advances in machine learning, GPU manufacturers have integrated specialized low-precision hardware units offering significantly higher throughput and energy efficiency than traditional 64-bit floating-point numbers. By selectively integrating these low-precision data types and arithmetic in performance-critical GPU kernels of RTE-RRTMGP-CPP, we improve computational efficiency while maintaining scientific accuracy within an acceptable tolerance. Using automatic precision tuning, we are able to find multiple configurations of the application that provide Pareto-optimal trade-offs between numerical accuracy and execution speed. Our results show that mixed-precision tuning accelerates the radiative transfer solver, achieving speedups of up to 2.1 × on Leonardo (NVIDIA A100) and 2.2 × on LUMI (AMD MI250X) compared to double precision and carefully selected configurations offer attractive performance-accuracy trade-offs compared to uniform single precision. Overall, this work demonstrates how to exploit the low-precision capabilities of modern GPUs and prepare weather and climate models for upcoming exascale systems.

Authors 7

1. Stijn Heldens corresponding

   Netherlands eScience Center

   Affiliation as printed

   Netherlands eScience Center, Amsterdam, the Netherlands

2. Alessio Sclocco

   Netherlands eScience Center

   Affiliation as printed

   Netherlands eScience Center, Amsterdam, the Netherlands

3. Gijs van den Oord

   Netherlands eScience Center

   Affiliation as printed

   Netherlands eScience Center, Amsterdam, the Netherlands

4. Ben van Werkhoven Aachen

   Leiden University

   Affiliation as printed

   Leiden University, Leiden, the Netherlands

5. Chiel van Heerwaarden

   Wageningen University & Research

   Affiliation as printed

   Wageningen University & Research, Wageningen, the Netherlands

6. Erwan Raffin
   Affiliation as printed

   Center for Excellence in Performance Programming (CEPP), Bull, Rennes, France

7. Xavier Yepes-Arbós

   Barcelona Supercomputing Center

   Affiliation as printed

   Barcelona Supercomputing Center, Barcelona, Spain