A 1.5-kilometer forecast of Earth—run faster than real time
A global weather model at 1.5-kilometer resolution has to calculate the atmosphere in extraordinary detail—and do it fast enough to be useful. SD3's authors report that their model simulated 205.7 days of global weather per day of computing, using 24,576 GPUs. That is not a claim that tomorrow's forecast is perfect, or that every part of Earth was modeled uniformly at 1.5 km for five years. It is a striking computational result, paired with a separate claim of years of operational forecasting at a coarser resolution.
The distinction matters. The paper describes continuous twice-daily operations since 2021 on a grid stretched to about 2.5 km over China and averaging 6 km globally. The 1.5 km result is a newer, large-scale performance benchmark. Together, the claims point toward a central challenge in weather and climate science: not only whether a model can briefly run at extraordinary resolution, but whether it can keep doing useful, stable work over time.
The Science
Most global weather and climate models divide the planet into grid cells, then calculate how air, water, heat and momentum move between them. Many established models use grid spacing of roughly 10 to 100 km. At those scales, a thunderstorm is smaller than a cell or occupies only a few cells, so its effects must be represented with approximations called parameterizations. Those approximations can become less reliable as grid spacing shrinks—a transitional range known as the “grey zone,” where storms are partly resolved but still partly represented by simplified formulas.
At around 1 to 4 km, models can begin to represent deep convection more directly. Convection is the vertical movement of warm, moist air that helps form clouds and intense rain. But finer grids do not automatically deliver better forecasts. They demand much more computation, and the model must handle rapid vertical accelerations and fast-moving sound waves. That makes numerical stability harder to maintain and can force shorter calculation steps.
SD3—Super Dynamics on the Cube—builds on the FV3 atmospheric model and its cubed-sphere grid. Instead of wrapping the globe in latitude and longitude lines, a cubed sphere maps it onto six connected faces. That avoids the tiny, tightly packed grid cells near the poles that can constrain time steps in a conventional latitude-longitude grid. The approach also divides naturally into regions that can be assigned across many processors.
The paper's most distinctive design choice is “Inline Physics.” Some atmospheric processes, including cloud microphysics, turbulent drag and scale-aware moist convection, are placed closely inside the model's main dynamical calculations rather than being treated as wholly separate steps. The goal is to make the handoff between the large-scale flow and fast local processes more physically consistent, particularly at resolutions where convection sits in the grey zone. The authors say the design conserves mass, momentum and total moist energy; the provided text does not include detailed conservation-error measurements.
SD3 also reworks much of the model for accelerator hardware. The authors describe a reconstruction of core dynamics and physical schemes in CUDA C, along with a mixed-precision strategy: many computationally intensive variables use 32-bit rather than 64-bit arithmetic. In principle, that can increase throughput, but lower precision can also introduce numerical errors. The paper says physical consistency was retained, although the supplied text does not report a detailed precision-sensitivity analysis.
The computing work is not just a matter of putting old code on new chips. GPU systems are highly parallel and can be slowed by data movement, memory limits and communication between nodes. SD3 uses memory pools to reduce repeated allocation, processes memory-intensive radiation calculations in tiles, and overlaps some communication with computation. The authors also describe a software abstraction intended to let the code compile against NVIDIA CUDA or AMD's HIP/ROCm interface. The paper presents this as portable; the evidence provided here gives performance results on one named system, not a cross-vendor benchmark.
What They Found
The headline result is the 1.5 km run: 205.7 simulated days per day on 24,576 GPUs. In plain terms, the model advanced through nearly seven months of simulated atmospheric time during one day of wall-clock time. The benchmark excludes initialization, and the paper says it includes the main dynamics and physical parameterizations. It is still a major computational achievement, but it is not equivalent to seven months of verified forecasts: model speed and forecast accuracy are different measures.
The authors compare that throughput with two prominent systems. They report 145.7 simulated days per day for ICON at 1.25 km on 20,480 NVIDIA GH200 GPUs, and 458.2 for SCREAM at 3.25 km on 32,768 AMD MI250X GPUs. SD3 therefore exceeds ICON's reported rate, as the paper emphasizes, but it does not have the highest raw throughput in this table: SCREAM's number is larger. These are not identical experiments. Resolution, hardware, model configuration and benchmark practices differ, so the comparison offers context rather than a definitive ranking.
Reported peak throughput across three kilometer-scale models
Peak simulated days per day (SDPD) reported in the paper. These are different model resolutions, systems and benchmarks, so the bars are not a controlled head-to-head comparison.
| Label | Value |
|---|---|
| SCREAM · 3.25 km | 458.2 SDPD |
| ICON · 1.25 km | 145.7 SDPD |
| SD3 · 1.5 km | 205.7 SDPD |
A second argument in the paper is that long, stable operations deserve to count alongside peak speed. The authors introduce accumulated degrees of freedom per hour, or DOF·hr, which combines the number of grid points, vertical layers and prognostic variables with the duration of a run. They estimate about 1.13 × 10^12 DOF·hr for a three-hour ICON benchmark and 1.11 × 10^14 for a 40-day SCREAM experiment. For SD3's operational configuration, they report 1.53 × 10^16 DOF·hr over five years.
That is a substantial difference in accumulated workload. But DOF·hr is a metric proposed by the paper's authors, not a standard measure of forecast quality or model reliability. Its value depends on what counts as a degree of freedom and how configurations are compared. It helps make duration visible; it cannot by itself establish that one model is more scientifically accurate.
Cumulative simulation workload reported by the paper
Accumulated degrees of freedom per hour (DOF·hr), a workload measure the authors use to emphasize integration duration and scale. Values for SCREAM and ICON are given as approximate estimates in the paper.
| Label | Value |
|---|---|
| ICON | 1,130,000,000,000 DOF·hr |
| SCREAM | 111,000,000,000,000 DOF·hr |
| SD3 operational run | 15,300,000,000,000,000 DOF·hr |
The paper reports strong-scaling tests, in which more GPUs are applied to the same problem. At 12 km, throughput rose from 375.5 simulated days per day on 48 GPUs to 1,492.2 on 384 GPUs, with a reported parallel efficiency of 49.7%. At 1.5 km, it rose from 66.7 on 3,072 GPUs to 205.7 on 24,576, with about 38.5% efficiency. In both cases, adding GPUs increased throughput, but not in direct proportion to the number added; communication becomes a larger share of the work as the workload per GPU falls.
SD3 strong-scaling throughput at two resolutions
Reported endpoints as GPU count increases in each experiment. The two resolution series use different GPU-count ranges; the chart shows the reported endpoints, not a controlled resolution comparison.
| Label | Value |
|---|---|
| 12 km · 48 GPUs | 375.5 SDPD |
| 12 km · 384 GPUs | 1,492.2 SDPD |
| 1.5 km · 3,072 GPUs | 66.7 SDPD |
| 1.5 km · 24,576 GPUs | 205.7 SDPD |
On forecast skill, the paper describes a hybrid approach that combines physical simulations and artificial-intelligence model outputs to create large-scale atmospheric guidance. For Northern Hemisphere 500-hPa geopotential height—a measure of the height of a standard pressure surface and a useful indicator of broad circulation—the authors report anomaly correlation above 0.6 at day 10 across 730 cases from March 2025 to February 2026. They also say SD3 outperformed ECMWF's IFS and NCEP's GFS across all lead times. The text supplied does not give the plotted values or uncertainty, so the magnitude and robustness of that advantage cannot be assessed here.
The paper also shows a multiscale case study of Super Typhoon Ragasa and a comparison of simulated cloud fields with satellite images and an astronaut photograph. Such visual comparisons can show whether a model produces plausible structures, but they are not a substitute for quantitative verification across many storms and conditions.
Why This Changes Things
Why spend so much computing power on smaller grid cells? Rain and wind extremes are shaped by processes that unfold over short distances: a line of thunderstorms, a mountain-driven downpour, or the rotating eyewall of a tropical cyclone. A global model that resolves more of those features could make it easier to study where hazards form, how they interact with larger weather patterns and how those patterns may shift in a warming climate.
That possibility is especially important for risk planning. Flood defenses, emergency preparations and infrastructure decisions depend on more than average temperature. They depend on the local character of intense rainfall, storm tracks and other extremes. Better-resolved simulations could help researchers investigate those details consistently across the globe, rather than relying only on limited regional experiments. But a finer grid is an opportunity, not a guarantee: errors in cloud physics, initial conditions or large-scale circulation can persist, and higher resolution can expose rather than eliminate model weaknesses.
SD3's emphasis on sustained operations addresses a real blind spot in headline supercomputer demonstrations. A short “hero run” can establish that a model reaches a speed under favorable conditions. Operational forecasting adds recurring deadlines, data input and output, repeated initialization and the risk of a failure that interrupts service. The authors' description of years of twice-daily production is relevant evidence of engineering maturity—though the paper should distinguish that operational configuration from the newer 1.5 km benchmark, as it does in the technical details.
Its “inline” coupling also targets a genuine scientific problem. If small-scale physical processes are handled as disconnected add-ons, the model can struggle where a process is neither wholly unresolved nor wholly resolved. Bringing selected fast processes into closer contact with the dynamics may reduce that mismatch. Whether it improves forecasts is an empirical question, however: the paper would be stronger with controlled comparisons against the same model configuration without inline physics, plus detailed conservation and precipitation-error statistics.
What's Next
The headline speed is impressive, but the next test is reproducibility. The paper's accessible results describe CNIS hardware and broad software choices, yet do not provide enough information here to independently reproduce every benchmark: for example, complete run configurations, timing variability, code availability and a verified second-hardware performance result. A single compile-time abstraction may simplify porting, but it does not alone prove equal behavior or efficiency across accelerator ecosystems.
Scientific validation deserves similar depth. The paper's 730-case day-10 circulation result is promising, and the storm and cloud images make the work tangible. Readers would still need numerical scores, baselines, uncertainty estimates and a clearer account of case selection to judge how much skill improved, for which regions and variables, and whether the improvement holds across seasons and extremes. The abstract's assertion of five years of validation against diverse observations is broader than the specific evidence reproduced in the supplied text.
There is also a climate question beyond forecast skill. Operational weather runs test whether a model can produce near-term conditions on schedule; climate projections require long integrations and careful treatment of changing boundary conditions and coupled Earth-system processes. SD3's mixed-layer ocean component and atmospheric focus are important steps, but the paper does not establish that the reported benchmark is already a complete, validated Earth-system digital twin.
If the results hold up under fuller independent testing, SD3 points toward a future in which kilometer-scale global simulations are no longer rare supercomputer stunts. The practical prize is not a more impressive number of simulated days. It is a reliable way to ask sharper questions about storms, rainfall and a changing climate—and to make the answers available in time to matter.