← All articles

🌍 Climate

32,768 Versions of the Weather in 34 Seconds: The Gordon Bell Winner Rewrites the Math

A Gordon Bell Prize-winning team ran 32,768 parallel versions of the atmosphere through the Frontier supercomputer in 34 seconds. That is 643 times the ensemble Europe's best forecast center runs, for about a third of one home's daily electricity.

Illustration: rows of supercomputer cabinets beneath a glowing holographic Earth with a hurricane spiral, trailed by dozens of translucent ghost Earths.

Thirty-four seconds. That is how long it took one of the world's most powerful supercomputers to generate 32,768 parallel versions of the entire atmosphere, each a complete picture of the planet's weather, each slightly different from the rest. On November 19, the team behind that run won the ACM Gordon Bell Prize for Climate Modelling at SC26 in Chicago, supercomputing's top climate honor, and the winning number was not a speed record but an ensemble size no forecast center on Earth has ever come close to running.

Why 32,768 Matters

Every forecast starts with a guess about the present. Satellites, buoys, aircraft, and ground stations sample only a fraction of the atmosphere, so the rest must be filled in, a job called data assimilation: combine physics simulation with patchy observations to estimate the true state of the sky. Get that starting picture wrong and even a perfect forecast model drifts from reality within days, which is why assimilation, not forecasting, has been the quiet bottleneck in weather science.

Because the atmosphere is chaotic and observations are incomplete, forecast centers run ensembles: dozens of slightly different starting states, each marched forward, with their spread measuring how much the forecast can be trusted. Operational systems have been stuck at tens to hundreds of members because each member costs a full physics simulation, and the paper says so plainly: "operational ensemble systems are typically limited to tens to hundreds of members because of computational and latency constraints, whereas much larger ensembles may be required to characterize complex dynamical systems, distribution tails, and rare events."

Europe's ECMWF, the most respected forecast center on the planet, runs 51 members for both its physics ensemble and its AI ensemble. STORM ran 32,768: divide one by the other and the gap is 643 to 1. Put differently, the world's best operational picture of forecast uncertainty samples 51 possible atmospheres; this run sampled 32,768 in about half a minute, and sampling error shrinks with the square root of the member count, so a 643-fold larger ensemble cuts that error roughly 25-fold.

What STORM Changed

STORM, short for Spatiotemporal Transformer for Observation-informed Reconstruction and Modeling, attacks in two places at once: first, it collapses the traditional two-stage assimilation workflow into one step: instead of running a physics ensemble forward and then correcting it with observations, a diffusion model samples the posterior state directly, conditioned on history and observations together, which removes the physics ensemble and the data movement between stages. Second, it fixes the attention math that has capped AI weather models: conventional global attention scales quadratically with space and time, so doubling resolution quadruples cost, while STORM's algorithm scales linearly, letting one model hold 20 billion tokens and 177,000 frames.

Those frame numbers deserve a second look: 177,000 hourly frames is 20.2 years of global weather, and 20 billion tokens is roughly 200,000 novels' worth of atmospheric state. Quadratic attention over 177,000 frames would need about 31 billion frame-pair interactions, while linear attention needs 177,000, a 177,000-fold cut and the arithmetic inside the paper's complexity claim.

The audit: every ratio below is our arithmetic on the paper's reported numbers
CalculationArithmeticResult
Ensemble multiple vs ECMWF32,768 / 51 membersAbout 643x
Share of Frontier used74,400 / 75,264 GPUs98.9%
Sustained vs peak BF166 / 14.25 ExaFLOPSAbout 42%
Ensemble energy, GPUs only4,096 x 280 W x 34 sAbout 11 kWh
Temporal context177,000 frames / 8,760 per year20.2 years
Attention pair cut177,000 squared vs 177,000177,000x fewer pairs

Frontier holds 9,408 nodes with four MI250X accelerators each, and each accelerator counts as two graphics compute dies, so the machine totals 75,264 GPUs: STORM's 74,400-GPU scaling run used 98.9 percent of the whole supercomputer at 96 to 99 percent strong-scaling efficiency. And 6 ExaFLOPs of sustained BF16 throughput is about 42 percent of the machine's theoretical peak: one MI250X package peaks at 383 teraflops, and 74,400 dies at 191.5 teraflops each top out near 14.25 ExaFLOPS.

Then the electricity bill. Four thousand ninety-six GPUs at 280 watts peak each draw about 1.15 megawatts, and 34 seconds at that draw is about 11 kilowatt-hours for the GPUs alone. An average American home buys 10,791 kilowatt-hours a year, about 29.6 a day, so one 32,768-member global ensemble costs roughly a third of a home's daily electricity. That figure excludes CPUs, memory, interconnect, and cooling, so treat it as a floor, but the order of magnitude is the point: the most expensive part of knowing how sure a forecast is now costs less power than running a dishwasher.

The Same Day, a Second Verdict

November 19 delivered a second, independent verdict on AI weather. University at Albany researchers analyzed Google DeepMind's WeatherNext Cyclone model across 385 tropical cyclones from June 2022 to May 2026: it beat NOAA's state-of-the-art hurricane model on intensity at every lead time from one to five days, matching human forecasters beyond the first day. STORM attacks the starting state. WeatherNext attacks the forecast itself. Two different teams, two different journals, one day: the machines are no longer the sideshow in weather prediction.

What You Can Do

If you check forecasts, look for the ensemble spread rather than the single line: a 51-member ensemble already tells you the odds, and STORM's point is that those odds could get 643 times sharper. If you work in energy, insurance, or emergency management, watch whether ECMWF or NOAA adopt one-stage generative assimilation, because the first operational deployment, not the prize, is the real milestone. If you run compute at scale, steal the 42-percent-of-peak figure: sustaining that fraction of BF16 throughput on 74,400 GPUs is a systems result as much as a weather result.

Limitations

This analysis leans on a preprint, arXiv:2604.16590 version 2 from August 31, 2026; the prize-winning version will appear in The International Journal of High Performance Computing Applications, so final peer review is pending. STORM does data assimilation, estimating the current state of the atmosphere, not forecasting, so a 32,768-member ensemble of todays is not 32,768 forecasts of next week, and its hurricane and reanalysis gains are the paper's own evaluations rather than an independent operational comparison. Six ExaFLOPs is BF16 half-precision throughput, not FP64 physics speed. Our 11-kilowatt-hour figure covers GPUs only at peak TDP and excludes CPUs, cooling, and interconnect. No operational center has adopted STORM.

The Strongest Case Against This Analysis

Here is the best case against taking this result as a forecast revolution. A 32,768-member ensemble is a research demonstration, not a forecast product, and operational centers run 51 members for reasons that have nothing to do with compute: each member must be high-resolution, low-latency, and maintained around the clock. Bigger ensembles shrink sampling error but do nothing about model bias, so if the model is wrong in the same direction 32,768 times, the ensemble is confidently wrong. The prize rewards computing rather than meteorology: STORM won for scaling, while the question that matters to a coastal resident, whether the hurricane hits their town, is still answered by the 51-member ensembles. Most of that is correct, so the honest reading is narrower: STORM removes the compute excuse for small ensembles, not the physics excuse, and it has not yet run operationally anywhere.

The Bottom Line

For thirty years, forecasters have known bigger ensembles mean better odds and settled for dozens because the physics would not allow more. STORM ran 32,768 versions of the atmosphere in 34 seconds on 98.9 percent of Frontier, at 42 percent of peak throughput, for about 11 kilowatt-hours of GPU electricity, and the prize it won says the compute problem is solved. What remains is the harder problem. Turning a demonstration into a forecast your town can trust still belongs to the 51-member ensembles running today.

Related

Sources: Wang et al., arXiv:2604.16590; ACM / HPCwire, Gordon Bell Prize for Climate Modelling; ECMWF, AIFS ENS operational; AMD, Instinct MI250X specifications; U.S. EIA, household electricity use; Vile et al., Geophysical Research Letters (2026), DOI 10.1029/2026gl123211. All ratios, energy figures, and comparisons are the author's own calculations on the cited numbers.