Researchers say China’s SD3 model ran continuously since 2021 and hit 205.7 simulated days/day

A team from TianJi Weather Science and Technology Company and the Chinese Academy of Sciences says in a new arXiv preprint that its global weather model has been running in continuous operations since 2021 and recently reached a new high-throughput benchmark at ultra-high resolution. The paper reports that the model, called Super Dynamics on the Cube, or SD3, has delivered five years of uninterrupted forecasting and achieved 205.7 simulated days per day at 1.5-kilometer global resolution on 24,576 GPUs. Those results, however, are author-reported claims in a preprint, not independently verified benchmark results or peer-reviewed findings.

That combination is what makes the paper notable. Global kilometer-scale weather models are a major goal in forecasting because they are convection-permitting, meaning they can explicitly simulate deep convective storms rather than relying as heavily on coarse approximations for those processes. But they are also extraordinarily expensive to run. Many high-profile demonstrations at this scale have been limited benchmark campaigns or short showcase runs. A claim that such a model has moved into uninterrupted operational use would be more consequential for forecasting than a one-off speed result alone.

The paper, “SD3: The First Operationally-Proven, Fully Heterogeneous-Accelerated Global Kilometer-Scale Model with Inline Physics Architecture,” was posted to arXiv on Oct. 29, 2026, as arXiv:2610.39334v1. The authors are affiliated with TianJi Weather Science and Technology Company in Beijing, the Institute of Atmospheric Physics at the Chinese Academy of Sciences, and the Institute of Software at the Chinese Academy of Sciences.

According to the preprint, SD3 is a global kilometer-scale model rebuilt “full-stack” in native CUDA C. It uses what the authors call an “Inline Physics” architecture, tightly coupling the model’s dynamical core with fast physical processes. The model is described as derived from the FV3 cubed-sphere dynamical core, a well-established modeling framework used in operational weather and climate systems.

The paper makes two headline claims. First, it says that since 2021, SD3 has run an uninterrupted production forecasting schedule in a stretched-grid setup averaging about 6 kilometers globally and 2.5 kilometers over China, with 95 vertical levels. The authors describe that as five years, or 1,825 days, of continuous operational forecasting. Second, the preprint says SD3 “achieved an unprecedented 205.7 simulated days per day at global 1.5 km resolution using 24576 GPUs” on a system the authors call China’s New-generation Intelligent Supercomputer, or CNIS.

The authors frame that performance as surpassing a 145.7 simulated-days-per-day figure they cite for a recent ICON run at similar ultra-high resolution. The paper also situates SD3 alongside earlier milestones such as SCREAM, a climate model that Lawrence Livermore National Laboratory said in 2023 was the first global cloud-resolving model to run on an exascale supercomputer and the first to exceed one simulated year per day. In other words, SD3 is being presented not just as another fast run, but as a model that combines very high speed with sustained operations.

Still, the benchmark claim should be treated cautiously. No independent public verification of the 205.7 simulated-days-per-day result was identified, and the paper has not yet gone through peer review. The evidence outside the preprint is narrower but still meaningful: The authors say SD3 forecast output is published through the AWS Open Data Program, and AWS lists a “TJweather Global-Regional Integrated Numerical Weather Prediction System Dataset (SD3)” as daily updated, managed by TianJi and licensed under CC BY 4.0. That does not verify the benchmark, but it does support the narrower point that SD3 forecast products are being published operationally. Other publications have also used SD3 in research contexts, though those sources likewise do not independently confirm the new performance result.

So the preprint stops short of establishing a verified record. But it does present a more interesting claim than a conventional supercomputing brag sheet: that a global, convection-permitting weather model may already have crossed from occasional demonstration into routine operations. The operational traces make that worth attention. The 205.7 simulated-days-per-day figure, for now, remains an author claim awaiting outside confirmation.