SCREAM Resonates In The Race For The Gordon Bell Climate Prize

The Bell will toll for some of the more interesting climate research projects on one of the world's most powerful supercomputers starting this year.

The new Gordon Bell Prize for Climate Modelling will be awarded annually for the next 10 years, starting in 2023, to acknowledge the efforts of climate scientists and software engineers in this domain.

Prize-winning submissions are expected to, among other things, highlight climate science and its allied fields, the effects of climate on society, and the use of supercomputers and parallel computing in addressing global climate issues.

A team at the Lawrence Livermore National Laboratory (LLNL) and seven other US Department of Energy (DOE) national laboratories are already among those shortlisted for the inaugural Association for Computing Machinery (ACM) Gordon Bell prize.

The nomination recognizes their work in operating a high-resolution global atmosphere model on one of the world's first exascale supercomputers. For those who don't follow supercomputing or climate modeling, whole Earth models are particularly challenging for a bevy of reasons – with model scalability only part of the equation.

The LLNL-led submission, overseen by the Energy Exascale Earth System Model (E3SM) primary computational scientist Mark Taylor, documents the historic performance of the Simple Cloud Resolving E3SM Atmosphere Model (SCREAM) on the first US exascale supercomputer, dubbed Frontier, at Oak Ridge National Laboratory, which has a computational capacity of 1.2 exaFLOPS (1.2 quintillion calculations per second).

SCREAM, spearheaded by LLNL researcher Peter Caldwell, integrates advanced parameters for fluid dynamics, microphysics, moist turbulence, and radiation, thereby establishing a comprehensive atmospheric general circulation model purposed for fine-resolution simulations on exascale systems.

• IBM, NASA emit actual open source AI model – for grokking Earth satellite images
• Processing data... in space: AWS powers Earth observation satellite payload
• A closer look at Harvard and Google's HPC heart research project
• Moore's Law is deader than corduroy bell bottoms. But with a bit of smart coding it's not the end of the road

One distinguishing aspect of SCREAM's development is its performance-portable design, which the team says enables ultra-detailed global atmospheric modelling, the first of its kind to function on an exascale supercomputer with a resolution finer than 5 kilometres. Specifically, this design allows it to operate effectively on both Nvidia and AMD GPU systems, surpassing one simulated year per day of throughput. Chew on that for a moment.

Due to a record-breaking run earlier this year, SCREAM advanced to the finalist position, with the updated submission exhibiting a 54 percent enhancement in speed, achieving a performance rate of 1.26 simulated years per day on 8,192 Frontier nodes.

Highlighting SCREAM's differentiation, Caldwell noted its construction using C++ and the integration of the Kokkos library, which facilitates efficient operation across various computer architectures. This architectural choice enabled faster operations on Frontier compared to other climate models, making it a valuable resource for other modelling centers aspiring to transition effectively to newer computing platforms.

Originating from the collaborative efforts of the E3SM team, a multi-lab DOE initiative led by LLNL scientist and well-known HPC and data science pioneer, David Bader, the project focuses on pioneering climate modelling, simulation, and forecasting for exascale machines.

The initiative is backed by the US DOE's Office of Biological and Environmental Research, including experts from various national laboratories such as LLNL, Sandia, and others. The nominated team from LLNL encompasses scientists including Aaron Donahue, Chris Terai, and Renata McCoy.

The team claims this development is a significant leap forward in climate modelling, marking a critical milestone for the E3SM project, which seeks to apply advanced computing science to climate simulation challenges. This, we're told, includes high-resolution simulations capable of representing large convective circulations and other vital atmospheric events, homing in on certain levels of accuracy missing in existing big climate models.

The reason why the focus is on resolution is because narrowing it down is essential for predicting crucial climate elements that could affect the US in the forthcoming years, including extreme weather events and sea level rises.

The prize, amounting to $10,000, courtesy of Gordon Bell, will be bestowed for the first time at the SC23 event in Denver (come and hang out with The Register and The Next Platform folks when you're there, would you?) and with nominations evaluated based on their potential to influence the climate modelling domain and associated fields. ®