The elevated frequency and severity of maximum climate and local weather occasions might take one million lives and price $1.7 trillion yearly by 2050, in accordance with the Munich Reinsurance Firm.
This underscores a vital want for correct climate forecasting, particularly with the rise in extreme climate occurrences reminiscent of blizzards, hurricanes and heatwaves. AI and accelerated computing are poised to assist.
Greater than 180 climate modeling facilities make use of sturdy excessive efficiency computing (HPC) infrastructure to crunch conventional numerical climate prediction (NWP) fashions. These embody the European Middle for Medium-Vary Climate Forecasts (ECMWF), which operates on 983,040 CPU cores, and the U.Okay. Met Workplace’s supercomputer, which makes use of greater than 1.5 million CPU cores and consumes 2.7 megawatts of energy.
Rethinking HPC Design
The worldwide push towards power effectivity is urging a rethink of HPC system design. Accelerated computing, harnessing the ability of GPUs, provides a promising, energy-efficient different that quickens computations.
NVIDIA GPUs have made a big influence on globally adopted climate fashions, together with these from ECMWF, the Max Planck Institute for Meteorology, the German Meteorological Service and the Nationwide Middle for Atmospheric Analysis.
GPUs improve efficiency as much as 24x, enhance power effectivity, and scale back prices and area necessities.
“To make dependable climate predictions and local weather projections a actuality inside energy funds limits, we depend on algorithmic enhancements and {hardware} the place NVIDIA GPUs are a substitute for CPUs,” stated Oliver Fuhrer, head of numerical prediction at MeteoSwiss, the Swiss nationwide workplace of meteorology and climatology.
AI Mannequin Boosts Pace, Effectivity
NVIDIA’s AI-based weather-prediction mannequin FourCastNet provides aggressive accuracy with orders of magnitude higher pace and power effectivity in contrast with conventional strategies. FourCastNet quickly produces week-long forecasts and permits for the era of huge ensembles — or teams of fashions with slight variations in beginning circumstances — for high-confidence, excessive climate predictions.
For instance, primarily based on historic information, FourCastNet precisely predicted the temperatures on July 5, 2018, in Ouargla, Algeria — Africa’s hottest recorded day.
Utilizing NVIDIA GPUs, FourCastNet rapidly and precisely generated 1,000 ensemble members, outpacing conventional fashions. A dozen of the members precisely predicted the excessive temperatures in Algeria primarily based on information from three weeks earlier than it occurred.
This marked the primary time the FourCastNet staff predicted a high-impact occasion weeks upfront, demonstrating AI’s potential for dependable climate forecasting with decrease power consumption than conventional climate fashions.
FourCastNet makes use of the most recent AI advances, reminiscent of transformer fashions, to bridge AI and physics for groundbreaking outcomes. It’s about 45,000x sooner than conventional NWP fashions. And when skilled, FourCastNet consumes 12,000x much less power to provide a forecast than the Europe-based Built-in Forecast System, a gold-standard NWP mannequin.
“NVIDIA FourCastNet opens the door to using AI for all kinds of purposes that can change the form of the NWP enterprise,” stated Bjorn Stevens, director of the Max Planck Institute for Meteorology.
Increasing What’s Attainable
In an NVIDIA GTC session, Stevens described what’s attainable now with the ICON local weather analysis device. The Levante supercomputer, utilizing 3,200 CPUs, can simulate 10 days of climate in 24 hours, Stevens stated. In distinction, the JUWELS Booster supercomputer, utilizing 1,200 NVIDIA A100 Tensor Core GPUs, can run 50 simulated days in the identical period of time.
Scientists want to examine local weather results 300 years into the long run, which suggests programs must be 20x sooner, Stevens added. Embracing sooner expertise like NVIDIA H100 Tensor Core GPUs and easier code might get us there, he stated.
Researchers now face the problem of hanging the optimum steadiness between bodily modeling and machine studying to provide sooner, extra correct local weather forecasts. A ECMWF weblog printed final month describes this hybrid method, which depends on machine studying for preliminary predictions and bodily fashions for information era, verification and system refinement.
Such an integration — delivered with accelerated computing — might result in vital developments in climate forecasting and local weather science, ushering in a brand new period of environment friendly, dependable and energy-conscious predictions.
Study extra about how accelerated computing and AI increase local weather science by way of these assets:
