Toward High‐Resolution Global Atmospheric Inverse Modeling Using Graphics Accelerators.

The ability of global transport models to go up in resolution becomes discriminating for greenhouse gas atmospheric inversions. This paper describes the porting on Graphics Processing Units of the global transport model currently used in the European operational Copernicus Atmosphere Monitoring Service (CAMS) for CO2 and N2O inversions. It represents an important milestone to achieve sub‐degree resolution. The code includes not only the direct model but also its tangent‐linear and its adjoint versions which are needed in variational inversions. Tests were carried out for CO2 at a resolution of 2.50° in longitude, 1.27° in latitude and 79 layers in the vertical, corresponding to 1,626,768 3D cells, 4.5 times more than the current standard resolution of the model used in the CAMS reanalyzes. A month's worth of computation of the tangent‐linear and of the adjoint versions now takes 2.5 min, including 50 s for reading meteorological data. Plain Language Summary: Atmospheric transport models are intensively used to infer global greenhouse gas emissions and removals, from atmospheric measurements: a single global analysis involves repeated and long transport simulations. This intensive use has limited the spatial resolution of such analyses despite an increasing need for national greenhouse gas budgets, despite an increasing number of corresponding space‐based observations at kilometer resolution, and despite an increasing number of high‐quality surface measurements made in sites marked by strong local influences. For the global transport model currently used in the European operational atmosphere monitoring service, our objective within the next 5 years is to make it reach a resolution of about 50 km over the whole globe. This paper describes an important milestone in this direction with the porting of this transport model on hardware components initially developed for video display but now used for high performance computing. Tests were carried out at a resolution of 2.50° in longitude, 1.27° in latitude and 79 layers in the vertical, corresponding to 4.5 times more 3D cells than the current standard resolution of the model. The direct model itself now takes less time than reading the input meteorological data. Key Points: The workload in a Eulerian transport model on a longitude‐latitude grid can be embarrassingly parallel enough to run efficiently on a Graphics Processing UnitThe transport model of the Laboratoire de Météorologie Dynamique was run in a new grid of 1,626,768 3D cellsThe calculation time in the direct version of the model is now less than the time to read the input meteorological data [ABSTRACT FROM AUTHOR]