<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE article PUBLIC "-//NLM//DTD Journal Publishing with OASIS Tables v3.0 20080202//EN" "https://jats.nlm.nih.gov/nlm-dtd/publishing/3.0/journalpub-oasis3.dtd">
<article xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:oasis="http://docs.oasis-open.org/ns/oasis-exchange/table" xml:lang="en" dtd-version="3.0" article-type="research-article">
  <front>
    <journal-meta><journal-id journal-id-type="publisher">GMD</journal-id><journal-title-group>
    <journal-title>Geoscientific Model Development</journal-title>
    <abbrev-journal-title abbrev-type="publisher">GMD</abbrev-journal-title><abbrev-journal-title abbrev-type="nlm-ta">Geosci. Model Dev.</abbrev-journal-title>
  </journal-title-group><issn pub-type="epub">1991-9603</issn><publisher>
    <publisher-name>Copernicus Publications</publisher-name>
    <publisher-loc>Göttingen, Germany</publisher-loc>
  </publisher></journal-meta>
    <article-meta>
      <article-id pub-id-type="doi">10.5194/gmd-19-10989-2026</article-id><title-group><article-title>GPU-accelerated atmospheric chemical kinetics using single precision in the ECHAM/MESSy (EMAC) model v2.55 with MEDINA v2.0</article-title><alt-title>MEDINA v2.0</alt-title>
      </title-group>
      <contrib-group>
        <contrib contrib-type="author" corresp="no" rid="aff1">
          <name><surname>Sophocleous</surname><given-names>Kyriacos</given-names></name>
          
        <ext-link>https://orcid.org/0000-0003-1306-3484</ext-link></contrib>
        <contrib contrib-type="author" corresp="no" rid="aff3">
          <name><surname>Kirfel</surname><given-names>Timo</given-names></name>
          
        <ext-link>https://orcid.org/0000-0001-6378-8745</ext-link></contrib>
        <contrib contrib-type="author" corresp="yes" rid="aff3">
          <name><surname>Taraborrelli</surname><given-names>Domenico</given-names></name>
          <email>d.taraborrelli@fz-juelich.de</email>
        <ext-link>https://orcid.org/0000-0003-2213-6307</ext-link></contrib>
        <contrib contrib-type="author" corresp="yes" rid="aff2">
          <name><surname>Christoudias</surname><given-names>Theodoros</given-names></name>
          <email>t.christoudias@cyi.ac.cy</email>
        <ext-link>https://orcid.org/0000-0001-9050-3880</ext-link></contrib>
        <aff id="aff1"><label>1</label><institution>Computation-based Science &amp; Technology Research Center, The Cyprus Institute, Nicosia, Cyprus</institution>
        </aff>
        <aff id="aff2"><label>2</label><institution>Climate and Atmosphere Research Center, The Cyprus Institute, Nicosia, Cyprus</institution>
        </aff>
        <aff id="aff3"><label>3</label><institution>Institute of Climate and Energy Systems - Troposphere (ICE-3), Forschungszentrum Jülich GmbH, Jülich, Germany</institution>
        </aff>
      </contrib-group>
      <author-notes><corresp id="corr1">Domenico Taraborrelli (d.taraborrelli@fz-juelich.de) and Theodoros Christoudias (t.christoudias@cyi.ac.cy)</corresp></author-notes><pub-date><day>18</day><month>November</month><year>2026</year></pub-date>
      
      <volume>19</volume>
      <issue>22</issue>
      <fpage>10989</fpage><lpage>11000</lpage>
      <history>
        <date date-type="received"><day>26</day><month>April</month><year>2026</year></date>
           <date date-type="rev-request"><day>30</day><month>July</month><year>2026</year></date>
           <date date-type="rev-recd"><day>18</day><month>October</month><year>2026</year></date>
           <date date-type="accepted"><day>4</day><month>November</month><year>2026</year></date>
      </history>
      <permissions>
        <copyright-statement>Copyright: © 2026 Kyriacos Sophocleous et al.</copyright-statement>
        <copyright-year>2026</copyright-year>
      <license license-type="open-access"><license-p>This work is licensed under the Creative Commons Attribution 4.0 International License. To view a copy of this licence, visit <ext-link ext-link-type="uri" xlink:href="https://creativecommons.org/licenses/by/4.0/">https://creativecommons.org/licenses/by/4.0/</ext-link></license-p></license></permissions><self-uri xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026.html">This article is available from https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026.html</self-uri><self-uri xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026.pdf">The full text article is available as a PDF file from https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026.pdf</self-uri>
      <abstract><title>Abstract</title>

      <p id="d2e126">Atmospheric chemical kinetics puts a major computational burden on Earth system models, yet its embarrassingly parallel nature of the problem structure makes it well suited to GPU acceleration. Here we present a single-precision floating point (FP32) GPU implementation of the MECCA chemical kinetics submodule in the ECHAM/MESSy Atmospheric Chemistry (EMAC) model. It is enabled by the MEDINA code-to-code compiler v2.0 and compatible with the Rosenbrock-family numerical solvers generated by the Kinetic Pre-Processor (KPP). The approach exploits FP32 arithmetic inside the GPU kernels while preserving double precision in the host EMAC workflow via explicit type conversion on the GPU.</p>

      <p id="d2e129">We evaluate the numerical fidelity and achieved performance in both an idealised box-model configuration and in full EMAC production simulations on the JUWELS Booster system. Across tested mechanisms and resolutions, FP32 reproduces the double-precision CPU reference to within <inline-formula><mml:math id="M1" display="inline"><mml:mo>&lt;</mml:mo></mml:math></inline-formula> 0.004 % in total-species-mass differences, with discrepancies confined to chemically depleted tracers near machine representation limits. In production-scale runs, FP32 further reduces the average chemical-kinetics execution time by about 30 % relative to the FP64 GPU implementation and lowers the memory footprint of FP-intensive data structures. Chemistry calculations in FP32 enhance scalability for high-resolution and chemically complex simulations.</p>
  </abstract>
    
<funding-group>
<award-group id="gs1">
<funding-source>European Commission</funding-source>
<award-id>856612</award-id>
<award-id>857645</award-id>
</award-group>
</funding-group>
</article-meta>
  </front>
<body>
      

<sec id="Ch1.S1" sec-type="intro">
  <label>1</label><title>Introduction</title>
      <p id="d2e148">Advances in numerical models have underpinned climate and atmospheric science over recent decades, allowing better understanding of the Earth system and improved  predictability under future climate change scenarios <xref ref-type="bibr" rid="bib1.bibx15" id="paren.1"/>. Coupled Earth System Models (ESMs) integrate interacting submodels for simulating different processes of the climate system. While ESMs have become more sophisticated, with greater capabilities and performance, in large part due to the advancements in CPU clock speeds and core counts, there are still limitations in complexity and spatial resolution for climate-scale simulations. These limitations are related to the time-to-solution requirements, power consumption demands, and storage and memory availability <xref ref-type="bibr" rid="bib1.bibx24 bib1.bibx10" id="paren.2"/>.</p>
      <p id="d2e157">A common ESM bottleneck is the simulation of atmospheric chemical kinetics <xref ref-type="bibr" rid="bib1.bibx5" id="paren.3"/>. The numerical integration of the stiff chemical ordinary differential equations (ODEs) remains computationally demanding despite recent optimization of the adaptive time-step control <xref ref-type="bibr" rid="bib1.bibx11" id="paren.4"/>. The problem is embarrassingly parallel allowing concurrent solving over model grid partitions <xref ref-type="bibr" rid="bib1.bibx1" id="paren.5"/>. Due to this property, offloading the chemical kinetics component of ESM simulations on Graphic Processing Units (GPU) accelerators can greatly reduce the required solution time. Such developments are documented in the literature for up to 20<inline-formula><mml:math id="M2" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula> speedup of the chemical kinetics computational kernel when compared to the equivalent CPU implementation (<xref ref-type="bibr" rid="bib1.bibx20" id="altparen.6"/>; <xref ref-type="bibr" rid="bib1.bibx3" id="altparen.7"/>).</p>
      <p id="d2e183">Architecturally, GPU processors are designed for massive parallel processing, featuring thousands of small, efficient cores optimised for handling simultaneous tasks. They utilise a Streaming Multiprocessor (SM) hierarchy, employing the Single Instruction, Multiple Threads (SIMT) model to execute data-intensive workloads. Because of their characteristics, GPUs are now widely adopted for heavy computing tasks including Machine Learning (ML), Artificial Intelligence (AI) and High Performance Computing (HPC) applications.</p>
      <p id="d2e186">The main limitation of GPUs is memory bandwidth and capacity, as they face severe performance degradation when workloads are memory-bound <xref ref-type="bibr" rid="bib1.bibx31 bib1.bibx10 bib1.bibx8" id="paren.8"/>. One under-exploited aspect of GPU design in this context, is that they deliver markedly higher floating point performance for 32-bit single precision (FP32) than 64-bit double precision (FP64) floating point operations, while also halving memory footprint. These benefits are especially valuable for the memory demanding, numerically intensive chemical kinetics simulations, and exploiting FP32 offers the prospect of additional speedups, reduced memory consumption, and improved energy efficiency.</p>
      <p id="d2e193">In this paper, we present the development of version 2.0 of the GPU-accelerated implementation of the MECCA atmospheric chemical kinetics submodule in EMAC (MEDINA) that now supports the use of single precision arithmetic (FP32) while managing to preserve numerical stability and accuracy. We evaluate the achieved computational performance (including time-to-solution, scaling and memory usage) and numerical accuracy of the single precision implementation against standard double-precision (FP64) CPU reference.</p>
<sec id="Ch1.S1.SS1">
  <label>1.1</label><title>Floating point representation</title>
      <p id="d2e203">In climate-chemistry modeling, physical parameters such as species concentrations, reaction constants, and sunlight intensity are all represented as floating point numbers, and simulating chemical processes involves extensive use of floating point operations.</p>
      <p id="d2e206">The IEEE 754 Standard defines floating point encoding and functionality <xref ref-type="bibr" rid="bib1.bibx14" id="paren.9"/>. The standard provides foundational formats to enable consistent calculations across platforms and sharing of floating point data while also specifies how floating point arithmetic results should be approximated. Floating point representation might have an impact on application accuracy when working with imprecise outcomes. In order to get the best performance with the precision necessary for any particular application, it is crucial to take into account a variety of factors in a heterogeneous computing environment where numerical operations are carried out on more than one hardware architectures.</p>
      <p id="d2e213">The fundamental binary floating point formats of 32 and 64 bits are called single and double precision floating point representations respectively. Under the IEEE 754 standard, double-precision (64-bit) floating-point arithmetic provides 15–17 significant decimal digits of precision and a dynamic range of approximately <inline-formula><mml:math id="M3" display="inline"><mml:mrow><mml:msup><mml:mn mathvariant="normal">10</mml:mn><mml:mrow><mml:mo>±</mml:mo><mml:mn mathvariant="normal">308</mml:mn></mml:mrow></mml:msup></mml:mrow></mml:math></inline-formula>, while single-precision (32-bit) arithmetic provides 7 significant decimal digits and a reduced dynamic range of approximately <inline-formula><mml:math id="M4" display="inline"><mml:mrow><mml:msup><mml:mn mathvariant="normal">10</mml:mn><mml:mrow><mml:mo>±</mml:mo><mml:mn mathvariant="normal">38</mml:mn></mml:mrow></mml:msup></mml:mrow></mml:math></inline-formula> <xref ref-type="bibr" rid="bib1.bibx14" id="paren.10"/>. Double precision has been commonly the representation used in Earth System Model simulations, as it is natively supported by recent generations of CPUs and enables greater numerical stability and precision. Lately, with the advent of GPU accelerators, the use of reduced precision has been increasingly explored in geoscientific modelling, to balance the trade-off between numerical accuracy and computational cost (<xref ref-type="bibr" rid="bib1.bibx13" id="altparen.11"/>; <xref ref-type="bibr" rid="bib1.bibx29" id="altparen.12"/>; <xref ref-type="bibr" rid="bib1.bibx28" id="altparen.13"/>).</p>
      <p id="d2e257">Even though GPUs support floating point numbers and operations in both single and double precision, they are optimised for single precision data structures arithmetic due to their historic development for graphics and very recently for machine-learning workloads. Using single precision can effectively halve the requirements compared to double precision for GPU global and cache memories, allowing more efficient data transfer over the memory hierarchy and the Arithmetic Logic Unit. Additionally, the amount of single precision cores is greater (typically double) compared to double precision cores for each multiprocessor that also contains in addition a set of 32-bit registers that are partitioned among the warps. Due to this architectural design, throughput of single precision instructions is much higher than the equivalent for double precision <xref ref-type="bibr" rid="bib1.bibx30" id="paren.14"/>.</p>

      <fig id="F1" specific-use="star"><label>Figure 1</label><caption><p id="d2e266">Single precision GPU-accelerated kernels with typecasting on CPU (Purple, Left Branch) or GPU (Cyan, Right Branch).</p></caption>
          <graphic xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026-f01.png"/>

        </fig>

</sec>
<sec id="Ch1.S1.SS2">
  <label>1.2</label><title>The MEDINA code</title>
      <p id="d2e283">The ECHAM/MESSy Atmospheric Chemistry (EMAC) model is a numerical chemistry and climate simulation system that includes sub-models describing tropospheric and middle atmosphere processes and their interaction with oceans, land and human influences <xref ref-type="bibr" rid="bib1.bibx16" id="paren.15"/>. It uses the second version of the Modular Earth Submodel System (MESSy2) to link multi-institutional computer codes <xref ref-type="bibr" rid="bib1.bibx17" id="paren.16"/>. The core atmospheric model is the 5th generation European Centre Hamburg general circulation model <xref ref-type="bibr" rid="bib1.bibx25" id="paren.17"><named-content content-type="pre">ECHAM5,</named-content></xref>. The physics subroutines of the original ECHAM code have been modularized and reimplemented as MESSy submodels and have continuously been further developed. Only the spectral transform core, the flux-form semi-Lagrangian large scale advection scheme, and the nudging routines for Newtonian relaxation are remaining from ECHAM.</p>
      <p id="d2e297">To compute the concentrations of and interactions between different species in the atmosphere, EMAC employs the module MECCA <xref ref-type="bibr" rid="bib1.bibx26" id="paren.18"/>, which incorporates the Kinetic Pre-Processor (KPP) general-purpose atmospheric chemistry open-source software tool <xref ref-type="bibr" rid="bib1.bibx27 bib1.bibx9" id="paren.19"/>. KPP numerically integrates atmospheric chemical kinetics, which are expressed as a coupled system of ordinary differential equations (ODE). KPP operates on chemical reactions written in a domain-specific language to output C and FORTRAN compatible code. The generated ODE solver enables temporal integration of the Jacobian representing the complete kinetic system.</p>
      <p id="d2e306">MEDINA is a software code that automatically generates Compute Unified Device Architecture <xref ref-type="bibr" rid="bib1.bibx21" id="paren.20"><named-content content-type="pre">CUDA</named-content></xref> kernels to numerically integrate atmospheric chemical kinetics in the EMAC model <xref ref-type="bibr" rid="bib1.bibx2" id="paren.21"/>. A compiler parses FORTRAN code generated by the Kinetic PreProcessor (KPP) and outputs CUDA API-compatible code to be compiled with the NVIDIA compiler suite. All Rosenbrock-family solvers <xref ref-type="bibr" rid="bib1.bibx12" id="paren.22"/> supported by KPP are supported by MEDINA.</p>
      <p id="d2e320">Performance evaluation of MEDINA v1.0, using Fermi and Pascal CUDA-enabled graphics processing unit (GPU) accelerators, shows achieved speed-ups of 4.5<inline-formula><mml:math id="M5" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula> and 20.4<inline-formula><mml:math id="M6" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula>, respectively, of the kernel execution time <xref ref-type="bibr" rid="bib1.bibx1" id="paren.23"/>. A node-to-node real-world production performance comparison shows a 1.75<inline-formula><mml:math id="M7" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula> speed-up over the non-accelerated model using the KPP three-stage Rosenbrock solver. In all cases, the accuracy and correctness of the accelerated implementation are evaluated by comparing to the CPU-only code of the application. The median relative difference is found to be less than <inline-formula><mml:math id="M8" display="inline"><mml:mrow><mml:msup><mml:mn mathvariant="normal">10</mml:mn><mml:mrow><mml:mo>-</mml:mo><mml:mn mathvariant="normal">10</mml:mn></mml:mrow></mml:msup></mml:mrow></mml:math></inline-formula> when comparing the output of the accelerated kernel the CPU-only code <xref ref-type="bibr" rid="bib1.bibx3" id="paren.24"/>. <xref ref-type="bibr" rid="bib1.bibx19" id="text.25"/> achieved much higher speedup (in the range 5.14<inline-formula><mml:math id="M9" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula> to 7.69<inline-formula><mml:math id="M10" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula>) for larger vectors of chemical species, resulting in greater computational load that can benefit from the capabilities of the GPUs.</p>
      <p id="d2e383">As proposed in <xref ref-type="bibr" rid="bib1.bibx28" id="text.26"/>, employment of single precision data structures for chemical kinetics simulations can effectively halve memory requirements without compromising the accuracy of results with respect to double precision, with maximum relative difference of <inline-formula><mml:math id="M11" display="inline"><mml:mo>∼</mml:mo></mml:math></inline-formula> 1 % in proof-of-concept initial investigations. Porting this concept to an architecture optimised for single precision, like GPUs, can thus be expected to improve performance in terms of time-to-solution, memory requirements and scalability.</p>
      <p id="d2e396">The release of MEDINA v2.0 has added capability to output CUDA code that supports offloading of computational kernels on GPUs that integrate atmospheric chemical kinetics in single precision. This paper presents the design and development of the software code (Sect. <xref ref-type="sec" rid="Ch1.S2.SS1"/>), experimental determination of the speed-up achieved against the GPU-accelerated double precision version of the software and against the original CPU code in a box model, and validation of the accuracy of the results, and profiling of the generated GPU-kernels (Sect. <xref ref-type="sec" rid="Ch1.S3.SS1"/>). Production-ready model simulations with MEDINA v2.0 in the EMAC model are also presented (Sect. <xref ref-type="sec" rid="Ch1.S3.SS2"/>) with evaluation of accuracy, speed-up and scalability on the Juelich Supercomputing Centre Juwels Booster HPC system (Sect. <xref ref-type="sec" rid="Ch1.S3.SS2.SSS1"/>).</p>
</sec>
</sec>
<sec id="Ch1.S2">
  <label>2</label><title>Methodology</title>
<sec id="Ch1.S2.SS1">
  <label>2.1</label><title>Implementation</title>
      <p id="d2e423">In the box model configuration, the whole simulation is output in single precision. In order to enable storing of data in single precision, but also keep the option of double precision, a new data type called <monospace>REAL</monospace> is introduced. <monospace>REAL</monospace> is interpreted as “double” by default. It can also be interpreted as “float” if at compilation the flag <monospace>__SINGLEPREC</monospace> is passed:

            <disp-formula id="Ch1.E1" content-type="numbered"><label>1</label><mml:math id="M12" display="block"><mml:mrow><mml:mi mathvariant="normal">REAL</mml:mi><mml:mo>=</mml:mo><mml:mfenced close="" open="{"><mml:mtable class="array" columnalign="right left"><mml:mtr><mml:mtd><mml:mrow><mml:mi mathvariant="normal">float</mml:mi><mml:mo>,</mml:mo></mml:mrow></mml:mtd><mml:mtd><mml:mtext>if (__SINGLEPREC) defined</mml:mtext></mml:mtd></mml:mtr><mml:mtr><mml:mtd><mml:mrow><mml:mi mathvariant="normal">double</mml:mi><mml:mo>,</mml:mo></mml:mrow></mml:mtd><mml:mtd><mml:mtext>otherwise</mml:mtext></mml:mtd></mml:mtr></mml:mtable></mml:mfenced></mml:mrow></mml:math></disp-formula>

          All floating point number data structures of the simulations have been transformed to <monospace>REAL</monospace>, thus according to the flag passed during compilation can either behave as double precision or single precision data structures during runtime. In addition, all special math functions are replaced by their single precision counterparts during the compilation.</p>

      <fig id="F2"><label>Figure 2</label><caption><p id="d2e472">Final concentrations (molec. cm<sup>−3</sup>) from the FP32 GPU implementation versus the FP64 CPU reference for all species using the three-stage Rosenbrock solver for the box model configuration.</p></caption>
          <graphic xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026-f02.png"/>

        </fig>

      <p id="d2e493">During integration of the new version of MEDINA with the EMAC model, one of the challenges is typecasting. As reaction rates, species concentrations and other variables used by chemical kinetics integration are in double precision in the rest of the modules, single precision chemical kinetics code cannot be straightforwardly embedded in the EMAC model. Hence, the data structures have to be typecast from double to single precision before the simulation of chemical kinetics and vice versa after the completion of the operation in order to ensure that the kernels are run in single precision arithmetic, while all the other submodules operate on the data in double precision.</p>

<table-wrap id="T1" specific-use="star"><label>Table 1</label><caption><p id="d2e499">Node configuration on the Juelich Supercomputing Centre Juwels Booster</p></caption><oasis:table frame="topbot"><oasis:tgroup cols="8">
     <oasis:colspec colnum="1" colname="col1" align="left"/>
     <oasis:colspec colnum="2" colname="col2" align="right"/>
     <oasis:colspec colnum="3" colname="col3" align="right"/>
     <oasis:colspec colnum="4" colname="col4" align="right" colsep="1"/>
     <oasis:colspec colnum="5" colname="col5" align="left"/>
     <oasis:colspec colnum="6" colname="col6" align="center"/>
     <oasis:colspec colnum="7" colname="col7" align="center"/>
     <oasis:colspec colnum="8" colname="col8" align="left"/>
     <oasis:thead>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1">CPU</oasis:entry>
         <oasis:entry colname="col2">Cores</oasis:entry>
         <oasis:entry colname="col3">RAM (GB)</oasis:entry>
         <oasis:entry colname="col4">TFlops</oasis:entry>
         <oasis:entry colname="col5">GPU</oasis:entry>
         <oasis:entry colname="col6">Cores</oasis:entry>
         <oasis:entry colname="col7">VRAM (GB)</oasis:entry>
         <oasis:entry colname="col8">TFlops</oasis:entry>
       </oasis:row>
     </oasis:thead>
     <oasis:tbody>
       <oasis:row>
         <oasis:entry colname="col1">AMD EPYC 7402</oasis:entry>
         <oasis:entry colname="col2">24</oasis:entry>
         <oasis:entry colname="col3">512</oasis:entry>
         <oasis:entry colname="col4">1.075</oasis:entry>
         <oasis:entry colname="col5">4 <inline-formula><mml:math id="M14" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula> NVidia A100</oasis:entry>
         <oasis:entry colname="col6">4 <inline-formula><mml:math id="M15" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula> 6912</oasis:entry>
         <oasis:entry colname="col7">4 <inline-formula><mml:math id="M16" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula> 40</oasis:entry>
         <oasis:entry colname="col8">9.7 (DP)/19.5 (SP)</oasis:entry>
       </oasis:row>
     </oasis:tbody>
   </oasis:tgroup></oasis:table></table-wrap>

      <p id="d2e607">For this reason, two approaches have been tested: <list list-type="order"><list-item>
      <p id="d2e612">Typecasting is performed on the host CPU and arguments are transferred to the device for the CUDA kernel in single precision. This implementation is optimal for data transfer on the GPU devices and can be easily unified with the EMAC model. In addition, it benefits from native single precision floating point numbers operations on the GPU, and lower memory requirements that enable better scalability and more complex and/or higher-resolution simulations. The drawback of this approach is that, due to the bulk of data required to be typecast, all the speed-up achieved by the use of single precision is contradicted by the time spent on the procedure of typecasting.</p></list-item><list-item>
      <p id="d2e616">Offloading the chemical kinetics data in double precision to the CUDA kernel and perform the typecast to single precision within the GPU. Performing the transformation on the device allows massive parallelisation for faster typecasting by alleviating the barrier that appears when typecasting on the CPU. The main drawback of this approach is that it requires <inline-formula><mml:math id="M17" display="inline"><mml:mrow><mml:mo>∼</mml:mo><mml:mn mathvariant="normal">1.5</mml:mn></mml:mrow></mml:math></inline-formula><inline-formula><mml:math id="M18" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula> memory footprint on the device as memory is allocated for the data that is passed to the kernel in double precision, with an additional half of the total memory also required to hold the variables in single precision prior to computation. The resulting species concentrations are also typecast back to double precision on the device prior to copying back to host. In this case two additional CUDA adapter functions have been implemented in the MEDINA kernel for typecasting: <list list-type="bullet"><list-item>
      <p id="d2e637"><monospace>__global__ void doubleToFloat(float *out, double *in, int n)</monospace></p></list-item><list-item>
      <p id="d2e642"><monospace>__global__ void floatToDouble(double *out, float *in, int n)</monospace></p></list-item></list></p></list-item></list> Both approaches are outlined in more detail in Fig. <xref ref-type="fig" rid="F1"/>.</p>
      <p id="d2e649">After extensively testing the two approaches, typecasting on CPU prior to GPU kernel execution was found to require overall more execution time. While this approach is straightforward to integrate and benefits from the reduction of memory requirements on the device side, it is inferior in terms of performance. Benchmark tests have shown that although the GPU kernel execution time was faster when compared to the typecasting-on-GPU approach, any speedup obtained was outweighed by the typecasting operations on the host CPU, rendering this approach inefficient. Thus, the benchmarking results presented (Sect. <xref ref-type="sec" rid="Ch1.S3"/>) are based on the second approach, where typecasting operations are offloaded to the GPU and are executed massively in parallel.</p>
</sec>
<sec id="Ch1.S2.SS2">
  <label>2.2</label><title>Model configuration</title>
      <p id="d2e662">MEDINA includes five numerical integrators commonly used  in atmospheric chemical kinetics models: the two-, three- and four- stage Rosenbrock family solvers (ros2, ros3, ros4) and the four- and six- stage Rodas family solvers (rodas3 and rodas4). The software also includes a unit test with a chemical mechanism in a box model configuration. This chemical mechanism includes 74 chemical species – 73 variable species and 1 fixed, and 77 chemical reactions and is reproduced a user defined number of times (<monospace>VL_GLO</monospace>) to mimic the column vector element computational structure of an atmospheric model. The box model simulation unit test is carried out for 30 d.</p>
      <p id="d2e668">For the production EMAC model runs, we applied EMAC (MESSy version 2.55.0) in the T42L90MA and T10690MA resolutions, i.e. with spherical truncations of T42 and T106 (corresponding to a quadratic Gaussian grid of approximately 2.8 and 1.1° in latitude and longitude respectively) and 90 vertical hybrid pressure levels up to 0.01 hPa. The chemical mechanism used includes 360 reactions and 159 chemical species, with 156 species active and 3 fixed based on the configuration described in detail in <xref ref-type="bibr" rid="bib1.bibx18" id="text.27"/>. With this configuration, simulations were run for one month (31 d). The experiments were carried out on the Juelich Supercomputing Centre Juwels Booster HPC nodes. The configuration of each Booster node is presented in Table <xref ref-type="table" rid="T1"/>.</p>
</sec>
</sec>
<sec id="Ch1.S3">
  <label>3</label><title>Results</title>
      <p id="d2e685">MEDINA has been extensively tested in terms of accuracy and performance. All the available solvers have been benchmarked for the non-accelerated CPU version and the GPU-accelerated versions of double and single precision. The box model configuration has been benchmarked on a single node of the Juwels Booster with the utilization of 1 (of the 4 available) GPU against a single CPU core. Scaling tests of the EMAC model have been conducted with up to 24 nodes with full available resources (24 CPU cores and 4 GPUs per node).</p>

<table-wrap id="T2"><label>Table 2</label><caption><p id="d2e691">Required execution time for the Integrator CUDA Kernel as output by NVIDIA Nsight Compute for all five supported numerical solvers in Double (DP) and Single Precision (SP).</p></caption><oasis:table frame="topbot"><oasis:tgroup cols="4">
     <oasis:colspec colnum="1" colname="col1" align="left"/>
     <oasis:colspec colnum="2" colname="col2" align="right"/>
     <oasis:colspec colnum="3" colname="col3" align="right"/>
     <oasis:colspec colnum="4" colname="col4" align="right"/>
     <oasis:thead>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1">Solver</oasis:entry>
         <oasis:entry colname="col2">DP (s)</oasis:entry>
         <oasis:entry colname="col3">SP (s)</oasis:entry>
         <oasis:entry colname="col4">Speedup (%)</oasis:entry>
       </oasis:row>
     </oasis:thead>
     <oasis:tbody>
       <oasis:row>
         <oasis:entry colname="col1">Ros2</oasis:entry>
         <oasis:entry colname="col2">2.40</oasis:entry>
         <oasis:entry colname="col3">1.69</oasis:entry>
         <oasis:entry colname="col4">42.01</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">Ros3</oasis:entry>
         <oasis:entry colname="col2">2.61</oasis:entry>
         <oasis:entry colname="col3">1.83</oasis:entry>
         <oasis:entry colname="col4">42.62</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">Ros4</oasis:entry>
         <oasis:entry colname="col2">3.22</oasis:entry>
         <oasis:entry colname="col3">2.22</oasis:entry>
         <oasis:entry colname="col4">45.05</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">Rodas3</oasis:entry>
         <oasis:entry colname="col2">2.60</oasis:entry>
         <oasis:entry colname="col3">1.79</oasis:entry>
         <oasis:entry colname="col4">45.25</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">Rodas4</oasis:entry>
         <oasis:entry colname="col2">4.93</oasis:entry>
         <oasis:entry colname="col3">3.42</oasis:entry>
         <oasis:entry colname="col4">44.15</oasis:entry>
       </oasis:row>
     </oasis:tbody>
   </oasis:tgroup></oasis:table></table-wrap>

<sec id="Ch1.S3.SS1">
  <label>3.1</label><title>Box model</title>
      <p id="d2e811">The accuracy of the results of the box model has been verified for the final concentrations of the species after 30 d simulations, by comparing the output by single precision GPU-accelerated simulations against the results of CPU-only reference simulations in double precision. Figure <xref ref-type="fig" rid="F2"/> shows on logarithmic axes the final concentrations of all species as output by single precision (FP32) GPU-accelerated code against the double precision (FP64) CPU reference. The points align along the <inline-formula><mml:math id="M19" display="inline"><mml:mrow><mml:mn mathvariant="normal">1</mml:mn><mml:mo>:</mml:mo><mml:mn mathvariant="normal">1</mml:mn></mml:mrow></mml:math></inline-formula> line across the full range spanning fifty orders of magnitude, indicating excellent agreement. We present here the results of the three-stage Rosenbrock solver (<monospace>Ros3</monospace>), but all the available solvers have been tested and the results are similar.</p>

<table-wrap id="T3" specific-use="star"><label>Table 3</label><caption><p id="d2e834">Memory transfer size between L1/L2 cache memories, L2/DRAM and L2 stitch edge cache memory in box model for Double and Single Precision using Ros3 solver.</p></caption><oasis:table frame="topbot"><oasis:tgroup cols="6">
     <oasis:colspec colnum="1" colname="col1" align="left"/>
     <oasis:colspec colnum="2" colname="col2" align="left"/>
     <oasis:colspec colnum="3" colname="col3" align="right"/>
     <oasis:colspec colnum="4" colname="col4" align="right"/>
     <oasis:colspec colnum="5" colname="col5" align="right"/>
     <oasis:colspec colnum="6" colname="col6" align="right"/>
     <oasis:thead>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1">Memory</oasis:entry>
         <oasis:entry colname="col2">Precision</oasis:entry>
         <oasis:entry colname="col3">Read (GB)</oasis:entry>
         <oasis:entry colname="col4">Write (GB)</oasis:entry>
         <oasis:entry colname="col5">Total (GB)</oasis:entry>
         <oasis:entry colname="col6">Reduction</oasis:entry>
       </oasis:row>
     </oasis:thead>
     <oasis:tbody>
       <oasis:row>
         <oasis:entry colname="col1">L1 <inline-formula><mml:math id="M20" display="inline"><mml:mo>↔</mml:mo></mml:math></inline-formula> L2</oasis:entry>
         <oasis:entry colname="col2">Double</oasis:entry>
         <oasis:entry colname="col3">1198.08</oasis:entry>
         <oasis:entry colname="col4">642.29</oasis:entry>
         <oasis:entry colname="col5">1840.37</oasis:entry>
         <oasis:entry colname="col6"/>
       </oasis:row>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1"/>
         <oasis:entry colname="col2">Single</oasis:entry>
         <oasis:entry colname="col3">1157.12</oasis:entry>
         <oasis:entry colname="col4">301.45</oasis:entry>
         <oasis:entry colname="col5">1458.57</oasis:entry>
         <oasis:entry colname="col6">20.75 %</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">L2 <inline-formula><mml:math id="M21" display="inline"><mml:mo>↔</mml:mo></mml:math></inline-formula> DRAM</oasis:entry>
         <oasis:entry colname="col2">Double</oasis:entry>
         <oasis:entry colname="col3">354.94</oasis:entry>
         <oasis:entry colname="col4">525.37</oasis:entry>
         <oasis:entry colname="col5">880.31</oasis:entry>
         <oasis:entry colname="col6"/>
       </oasis:row>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1"/>
         <oasis:entry colname="col2">Single</oasis:entry>
         <oasis:entry colname="col3">181.26</oasis:entry>
         <oasis:entry colname="col4">142.31</oasis:entry>
         <oasis:entry colname="col5">323.57</oasis:entry>
         <oasis:entry colname="col6">63.24 %</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">L2 stitch edge</oasis:entry>
         <oasis:entry colname="col2">Double</oasis:entry>
         <oasis:entry colname="col3"/>
         <oasis:entry colname="col4"/>
         <oasis:entry colname="col5">1167.36</oasis:entry>
         <oasis:entry colname="col6"/>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1"/>
         <oasis:entry colname="col2">Single</oasis:entry>
         <oasis:entry colname="col3"/>
         <oasis:entry colname="col4"/>
         <oasis:entry colname="col5">776.27</oasis:entry>
         <oasis:entry colname="col6">33.05 %</oasis:entry>
       </oasis:row>
     </oasis:tbody>
   </oasis:tgroup></oasis:table></table-wrap>

      <fig id="F3" specific-use="star"><label>Figure 3</label><caption><p id="d2e1022">Profiling analysis of GPU-accelerated chemical kinetics for the Box Model using the Ros3 solver and VL_GLO <inline-formula><mml:math id="M22" display="inline"><mml:mrow><mml:mo>=</mml:mo><mml:mn mathvariant="normal">576</mml:mn><mml:mspace linebreak="nobreak" width="0.125em"/><mml:mn mathvariant="normal">000</mml:mn></mml:mrow></mml:math></inline-formula>. <bold>(a)</bold> Time consumption of memory-related processes obtained with NVIDIA Nsight Systems, including memory allocation (malloc), host-to-device data transfer, typecasting, and device-to-host data transfer for single- (FP32) and double-precision (FP64). <bold>(b)</bold> L1/Tex Cache and L2 Cache Throughput and Hit Rates obtained with NVIDIA Nsight Compute for FP32 and FP64 implementations.</p></caption>
          <graphic xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026-f03.png"/>

        </fig>

      <p id="d2e1051">To quantify differences, we used the integrated unit test comparison post-processing routine in MEDINA, which computes, for each species, a relative difference between the GPU and CPU end-state and reports whether it exceeds a user-defined threshold. In FP64-vs-FP64 control runs, no species exceeded the nominal limit of 5 %, and in fact were at 0.0004 % or lower. In FP32-vs-FP64 comparisons, all species with non-negligible volume mixing ratios met the accuracy criterion, other than species that are chemically depleted and take values below the FP32 machine limit of representation. In those cases the absolute volume mixing ratio discrepancies are lower than 10<sup>−38</sup> molec. cm<sup>−3</sup>, and remain negligible compared to any atmospherically relevant value.</p>

      <fig id="F4" specific-use="star"><label>Figure 4</label><caption><p id="d2e1080">Relative difference (%) in total mass of species between the FP32 GPU implementation and the FP64 CPU reference in the EMAC model.</p></caption>
          <graphic xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026-f04.png"/>

        </fig>

      <p id="d2e1089">We used the NVIDIA Nsight Systems <xref ref-type="bibr" rid="bib1.bibx23" id="paren.28"/> and NVIDIA Nsight Compute <xref ref-type="bibr" rid="bib1.bibx22" id="paren.29"/> performance analysis tools for understanding and optimizing the performance of applications running on NVIDIA GPUs to gain more quantitative insights on performance, and identify possible bottlenecks. The GPU-accelerated box model code has been profiled on the HPC system for all five available solvers for both the FP64 and the FP32 implementations with grid-point vector length VL_GLO <inline-formula><mml:math id="M25" display="inline"><mml:mrow><mml:mo>=</mml:mo><mml:mn mathvariant="normal">576</mml:mn><mml:mspace width="0.125em" linebreak="nobreak"/><mml:mn mathvariant="normal">000</mml:mn></mml:mrow></mml:math></inline-formula>. As expected, most of the required simulation time (<inline-formula><mml:math id="M26" display="inline"><mml:mo lspace="0mm">∼</mml:mo></mml:math></inline-formula> 99 %) is spent in four procedures: allocation of memory (<monospace>cudaMalloc</monospace>), copying the data from the host to the device (<monospace>cudaMemcpy_HtoD</monospace>), integration of the sparse linear system (Rosenbrock solver), and copying the data back from the device to the host (<monospace>cudaMemcpy_DtoH</monospace>). Nsight Systems has shown that up to 94.5 % of the required runtime is spent on the integration and only <inline-formula><mml:math id="M27" display="inline"><mml:mo>∼</mml:mo></mml:math></inline-formula> 5 %–5.5 % of the time is required for the other processes. Profiling has also revealed improved kernel execution time for the Single Precision implementation compared to Double, with a speedup (Eq. <xref ref-type="disp-formula" rid="Ch1.E2"/>) of 42 %–45 % depending on the selected solver (Table <xref ref-type="table" rid="T2"/>).

            <disp-formula id="Ch1.E2" content-type="numbered"><label>2</label><mml:math id="M28" display="block"><mml:mrow><mml:mi>S</mml:mi><mml:mo>[</mml:mo><mml:mi mathvariant="italic">%</mml:mi><mml:mo>]</mml:mo><mml:mo>=</mml:mo><mml:mstyle displaystyle="true"><mml:mfrac style="display"><mml:mrow><mml:msub><mml:mi>t</mml:mi><mml:mi mathvariant="normal">DP</mml:mi></mml:msub><mml:mo>-</mml:mo><mml:msub><mml:mi>t</mml:mi><mml:mi mathvariant="normal">SP</mml:mi></mml:msub></mml:mrow><mml:mrow><mml:msub><mml:mi>t</mml:mi><mml:mi mathvariant="normal">SP</mml:mi></mml:msub></mml:mrow></mml:mfrac></mml:mstyle><mml:mo>×</mml:mo><mml:mn mathvariant="normal">100</mml:mn></mml:mrow></mml:math></disp-formula></p>
      <p id="d2e1179">Additionally, as shown in Table <xref ref-type="table" rid="T3"/>, the single precision implementation achieves a reduction in memory traffic of more than 63 % in L2 cache <inline-formula><mml:math id="M29" display="inline"><mml:mo>↔</mml:mo></mml:math></inline-formula> DRAM processes, more than 20 % in L1 cache <inline-formula><mml:math id="M30" display="inline"><mml:mo>↔</mml:mo></mml:math></inline-formula> L2 cache processes, and more than 33 % in L2 stitch edge cache memory. This happens not only because memory requirements are lower for single precision, but also because hit rates for both L1/Tex and L2 Cache memories are higher for single compared to double precision (Fig. <xref ref-type="fig" rid="F3"/>b).</p>

      <fig id="F5"><label>Figure 5</label><caption><p id="d2e1202">Final mass (kg) of species from the GPU single precision against GPU double precision simulation in the global EMAC model.</p></caption>
          <graphic xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026-f05.png"/>

        </fig>

      <p id="d2e1212">Figure <xref ref-type="fig" rid="F3"/> presents memory-related processes of GPU-accelerated chemical kinetics. For this analysis, the 3-step Rosenbrock-family (Ros3) solver was employed. The breakdown shows that the time spent in memory operations is slightly improved for FP32 and that typecasting consumes a marginal amount of the total time-to-solution (Fig. <xref ref-type="fig" rid="F3"/>a). Figure <xref ref-type="fig" rid="F3"/>b presents kernel-level memory metrics as captured by NVIDIA Nsight Compute. Single Precision manages better throughput for both L1/Tex and L2 Cache memories, while hit rates are improved by more than 10 %, indicating better utilisation of the GPU memory hierarchy.</p>
</sec>
<sec id="Ch1.S3.SS2">
  <label>3.2</label><title>EMAC model</title>
      <p id="d2e1230">This section evaluates the numerical accuracy of FP32 GPU implementation within the global EMAC model against the FP64 CPU and GPU versions. Figure <xref ref-type="fig" rid="F4"/> demonstrates the relative difference <inline-formula><mml:math id="M31" display="inline"><mml:mrow><mml:msub><mml:mi>M</mml:mi><mml:mi mathvariant="normal">i</mml:mi></mml:msub></mml:mrow></mml:math></inline-formula> in total atmospheric mass of individual chemical species between the accelerated GPU version of the code in single precision and the CPU reference version.

            <disp-formula id="Ch1.E3" content-type="numbered"><label>3</label><mml:math id="M32" display="block"><mml:mrow><mml:msub><mml:mi>M</mml:mi><mml:mi mathvariant="normal">i</mml:mi></mml:msub><mml:mo>[</mml:mo><mml:mi mathvariant="italic">%</mml:mi><mml:mo>]</mml:mo><mml:mo>=</mml:mo><mml:mstyle displaystyle="true"><mml:mfrac style="display"><mml:mrow><mml:msub><mml:mi>M</mml:mi><mml:mi mathvariant="normal">CPU</mml:mi></mml:msub><mml:mo>-</mml:mo><mml:msub><mml:mi>M</mml:mi><mml:mrow><mml:mi mathvariant="normal">GPU</mml:mi><mml:mi mathvariant="italic">_</mml:mi><mml:mn mathvariant="normal">32</mml:mn></mml:mrow></mml:msub></mml:mrow><mml:mrow><mml:msub><mml:mi>M</mml:mi><mml:mi mathvariant="normal">CPU</mml:mi></mml:msub></mml:mrow></mml:mfrac></mml:mstyle></mml:mrow></mml:math></disp-formula></p>
      <p id="d2e1290">The relative difference <inline-formula><mml:math id="M33" display="inline"><mml:mrow><mml:msub><mml:mi>M</mml:mi><mml:mi mathvariant="normal">i</mml:mi></mml:msub></mml:mrow></mml:math></inline-formula> in Eq. (<xref ref-type="disp-formula" rid="Ch1.E3"/>) is expressed as the difference between the mass output from the FP64 CPU simulation and FP32 GPU version over the mass computed by the reference CPU version. This analysis presents the 80 species with the highest relative differences. For all species, the maximum relative difference remains below 0.004 %, indicating excellent numerical agreement, well within the inherent systematic uncertainties in reaction rates. All other species with lower relative differences are presented in Fig. <xref ref-type="fig" rid="FA1"/>.</p>
      <p id="d2e1308">In Fig. <xref ref-type="fig" rid="F5"/> we compare the final mass of all species as output by Single Precision GPU version against the Double Precision GPU version using the global EMAC model in full production mode over a one-month simulation.</p>
      <p id="d2e1313">Again, we find consistent agreement across the range of masses spanning close to fifty orders of magnitude, with values almost perfectly aligned along the <inline-formula><mml:math id="M34" display="inline"><mml:mrow><mml:mn mathvariant="normal">1</mml:mn><mml:mo>:</mml:mo><mml:mn mathvariant="normal">1</mml:mn></mml:mrow></mml:math></inline-formula> equivalence. This result demonstrates that single-precision arithmetic preserves the numerical integrity of the GPU-accelerated chemical kinetics within EMAC. Results are essentially identical for FP32 against FP64 on GPUs, paving the way for use in real-world production simulations.</p>
      <p id="d2e1329">Moreover, we evaluate the spatial consistency of chemical species between the two precisions (Fig. <xref ref-type="fig" rid="F6"/>). The surface concentrations of key species, the hydroxyl radical (OH), the ozone (O<sub>3</sub>), and Nitrate (NO<sub>3</sub>) are compared along with their relative differences (Eq. <xref ref-type="disp-formula" rid="Ch1.E3"/>) at the end of the simulation. The spatial patterns for these species, are in excellent agreement between the two simulations, with relative differences remaining low across the domain and consistent with the species-wise mass differences shown in Fig. <xref ref-type="fig" rid="F4"/>. Elevated relative differences are confined to regions of very low concentrations, where absolute values are close to zero. These results confirm that single-precision GPU chemistry accurately reproduces the double-precision CPU solution.</p>

      <fig id="F6" specific-use="star"><label>Figure 6</label><caption><p id="d2e1358">Surface concentrations of OH, O<sub>3</sub>, and NO<sub>3</sub> for the single-precision GPU implementation (<bold>a, d, g</bold> respectively) the double-precision CPU reference (<bold>b, e, h</bold> respectively) and their relative differences (<bold>c, f, i</bold> respectively) in the EMAC model.</p></caption>
          <graphic xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026-f06.png"/>

        </fig>

<sec id="Ch1.S3.SS2.SSS1">
  <label>3.2.1</label><title>Strong scaling</title>
      <p id="d2e1401">A strong-scaling assessment of MEDINA v2.0 in full production EMAC simulations was conducted on the JUWELS Booster system <xref ref-type="bibr" rid="bib1.bibx4" id="paren.30"/> to evaluate the scalability and time-to-solution benefits of single-precision GPU chemistry. All experiments were carried out using the following configuration on JUWELS Booster: twenty-four (24) CPU cores and four (4) GPUs per node. For T42 resolution the experiments were carried out on 1, 2, 4, 6 and 8 JUWELS Booster nodes, while for T106 1, 2, 4, 8, 12, 16, 20 and 24 nodes were used.</p>
      <p id="d2e1407">Figure <xref ref-type="fig" rid="F7"/> presents the scaling of the execution time (<inline-formula><mml:math id="M39" display="inline"><mml:mrow><mml:msub><mml:mi>t</mml:mi><mml:mi mathvariant="normal">s</mml:mi></mml:msub></mml:mrow></mml:math></inline-formula>) associated with the MEDINA chemical kinetics integration for the FP32 and FP64 GPU implementations. The left panel (Fig. <xref ref-type="fig" rid="F7"/>a) corresponds to simulations at T42 horizontal resolution, while the right panel (Fig. <xref ref-type="fig" rid="F7"/>b) shows the results for the finer T106 resolution. Across all configurations and for both resolutions, the execution-time distributions for the FP32 implementation are consistently reduced compared to FP64, indicating markedly faster chemical kinetics integration when Single Precision is used. In terms of the average execution time (Eq. <xref ref-type="disp-formula" rid="Ch1.E4"/>), FP32 achieves a consistent <inline-formula><mml:math id="M40" display="inline"><mml:mo>∼</mml:mo></mml:math></inline-formula> 30 % improvement when compared to FP64, for all numbers of Nodes/GPUs used and for both model resolutions, demonstrating robust scalability of the single-precision implementation. This performance gain derives from increased throughput coming from the GPU architecture optimisation for single-precision floating-point operations.

              <disp-formula id="Ch1.E4" content-type="numbered"><label>4</label><mml:math id="M41" display="block"><mml:mrow><mml:mi mathvariant="normal">AET</mml:mi><mml:mo>=</mml:mo><mml:mstyle displaystyle="true"><mml:mfrac style="display"><mml:mrow><mml:msubsup><mml:mo>∑</mml:mo><mml:mrow><mml:mi>s</mml:mi><mml:mo>=</mml:mo><mml:mn mathvariant="normal">1</mml:mn></mml:mrow><mml:mrow><mml:msub><mml:mi>N</mml:mi><mml:mi mathvariant="normal">steps</mml:mi></mml:msub></mml:mrow></mml:msubsup><mml:msub><mml:mi>t</mml:mi><mml:mi mathvariant="normal">s</mml:mi></mml:msub></mml:mrow><mml:mrow><mml:msub><mml:mi>N</mml:mi><mml:mi mathvariant="normal">steps</mml:mi></mml:msub></mml:mrow></mml:mfrac></mml:mstyle></mml:mrow></mml:math></disp-formula></p>

      <fig id="F7" specific-use="star"><label>Figure 7</label><caption><p id="d2e1478">Distribution of execution time for the MEDINA chemical kinetics integration loop using FP32 and FP64 GPU implementations. <bold>(a)</bold> T42 resolution with experiments conducted on 1 to 8 nodes; <bold>(b)</bold> Results for T106 resolution using 1 to 24 nodes on the JUWELS system.</p></caption>
            <graphic xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026-f07.png"/>

          </fig>

      <p id="d2e1494">The overall strong-scaling performance for the chemical kinetics component of EMAC at T42 and T106 resolutions is presented in Fig. <xref ref-type="fig" rid="F8"/>. The speedup (<inline-formula><mml:math id="M42" display="inline"><mml:mrow><mml:msub><mml:mi>S</mml:mi><mml:mi>i</mml:mi></mml:msub></mml:mrow></mml:math></inline-formula>) is defined in Eq. (<xref ref-type="disp-formula" rid="Ch1.E5"/>) as the ratio of the time-to-solution on a given number of nodes (<inline-formula><mml:math id="M43" display="inline"><mml:mrow><mml:msub><mml:mi>t</mml:mi><mml:mi>i</mml:mi></mml:msub></mml:mrow></mml:math></inline-formula>) to the time-to-solution of the Double Precision implementation on a single node (<inline-formula><mml:math id="M44" display="inline"><mml:mrow><mml:msubsup><mml:mi>t</mml:mi><mml:mn mathvariant="normal">1</mml:mn><mml:mi mathvariant="normal">DP</mml:mi></mml:msubsup></mml:mrow></mml:math></inline-formula>). For both resolutions and all Node configurations, the FP32 GPU implementation scores higher overall speedup compared to FP64, with the performance gap increasing as more resources are provided.

              <disp-formula id="Ch1.E5" content-type="numbered"><label>5</label><mml:math id="M45" display="block"><mml:mrow><mml:msub><mml:mi>S</mml:mi><mml:mi>i</mml:mi></mml:msub><mml:mo>=</mml:mo><mml:mstyle displaystyle="true"><mml:mfrac style="display"><mml:mrow><mml:msub><mml:mi>t</mml:mi><mml:mi>i</mml:mi></mml:msub></mml:mrow><mml:mrow><mml:msubsup><mml:mi>t</mml:mi><mml:mn mathvariant="normal">1</mml:mn><mml:mi mathvariant="normal">DP</mml:mi></mml:msubsup></mml:mrow></mml:mfrac></mml:mstyle></mml:mrow></mml:math></disp-formula></p>

      <fig id="F8" specific-use="star"><label>Figure 8</label><caption><p id="d2e1567">Strong-scaling speedup of the EMAC model on JUWELS for T42 (left) and T106 (right) resolutions, comparing FP64 and FP32 GPU chemical kinetics, against the number of GPUs used.</p></caption>
            <graphic xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026-f08.png"/>

          </fig>

      <p id="d2e1576">Our results demonstrate that single-precision GPU chemical kinetics significantly improves both the performance and scalability of the EMAC ESM. The benefits are sustained for both resolutions and all node configurations, making FP32 GPU chemistry with MEDINA v.2.0 an effective approach for accelerating large-scale Earth system simulations without compromising numerical accuracy.</p>
</sec>
</sec>
</sec>
<sec id="Ch1.S4" sec-type="conclusions">
  <label>4</label><title>Conclusions</title>
      <p id="d2e1589">This paper presents the development of MEDINA v2.0, with the added feature to produce single-precision (FP32) CUDA kernels for atmospheric chemical kinetics ready to be integrated in the ECHAM/MESSy (EMAC) Earth system model. We evaluate the single-precision GPU implementation of the MECCA chemical kinetics submodule and demonstrate that it can reliably replace the FP64 GPU reference, delivering significant performance improvements while preserving numerical accuracy.</p>
      <p id="d2e1593">Accuracy assessments, for both the box-model configuration and full production-grade EMAC simulations, have shown excellent agreement between the newly developed FP32 GPU version of the code and the FP64 CPU and GPU references. Differences in species average mixing rations and total atmospheric mass remain below 0.004 % in all cases, with larger relative discrepancies to be found for species that are chemically depleted at the machine limit.</p>
      <p id="d2e1596">In terms of performance, the FP32 GPU achieves better speedup, scalability, and memory utilization compared to its FP64 counterpart. For the box-model configuration, single precision consistently achieves lower normalised execution times for all Rosenbrock-family solvers and problem sizes. By utilising NVidia profilers (Nvidia Nsight Compute and Nvidia Nsight Systems) we have shown that FP32 performes better in memory operations and cache efficiency, with higher throughput and hit rates for both L1/Tex and L2 cache memories. Although GPU typecasting introduces additional overhead, the enabled parallel execution manages to keep the required execution time low, preserving the gains of kernel execution in single precision.</p>
      <p id="d2e1599">In production-scale EMAC simulations on the JUWELS Booster system, the FP32 implementation achieves an average performance gain of approximately 30 % in chemical-kinetics execution time relative to the FP64 GPU version, for both T42 and T106 resolutions and for all node configurations. Strong-scaling experiments have demonstrated that the acceleration obtained is consistent and increases with system size, underlining the suitability of single-precision GPU chemistry for large-scale, high resolution simulations.</p>
      <p id="d2e1603">As Earth system models rely on heterogeneous GPU-accelerated architectures, reduction in computational cost and performance gain without accuracy compromising is essential. The presented approach demonstrates that single-precision GPU-accelerated chemical kinetics can be safely integrated into EMAC and provide a framework with better time-to-solution and memory management when compared to the nowadays standard Double Precision, improving efficiency and scalability of computationally intensive model components.</p>
</sec>

      
      </body>
    <back><app-group>

<app id="App1.Ch1.S1">
  <label>Appendix A</label><title>Additional Results</title>

      <fig id="FA1"><label>Figure A1</label><caption><p id="d2e1620">Relative difference (%) in total mass of additional species between the FP32 GPU implementation and the FP64 CPU reference in the EMAC model.</p></caption>
        
        <graphic xlink:href="https://gmd.copernicus.org/articles/19/10989/2026/gmd-19-10989-2026-f09.png"/>

      </fig>

</app>
  </app-group><notes notes-type="codedataavailability"><title>Code and data availability</title>

      <p id="d2e1635">The MEDINA FORTRAN to CUDA code-to-code compiler is developed in CUDA and Python and is included in the EMAC model. A consortium of institutions continuously develops the EMAC model. The usage of EMAC and access to the source code are licensed to all affiliates of institutions which are members of the MESSy Consortium. Institutions can become a member of the MESSy Consortium by signing the MESSy Memorandum of Understanding. More information can be found on the MESSy Consortium website (<uri>http://www.messy-interface.org</uri>, last access: 6 November 2026). The code presented here was developed based on MESSy version 2.55.0 and will be available in the next official release (version 2.56). The exact version of the MEDINA code used to produce the results used in this paper is archived on the Zenodo repository under the MIT lisence; DOI: <ext-link xlink:href="https://doi.org/10.5281/zenodo.19882619" ext-link-type="DOI">10.5281/zenodo.19882619</ext-link> <xref ref-type="bibr" rid="bib1.bibx6" id="paren.31"/>. All data sources for the results presented are available with <ext-link xlink:href="https://doi.org/10.5281/zenodo.21442083" ext-link-type="DOI">10.5281/zenodo.21442083</ext-link> <xref ref-type="bibr" rid="bib1.bibx7" id="paren.32"/>.</p>
  </notes><notes notes-type="authorcontribution"><title>Author contributions</title>

      <p id="d2e1661">Software, Investigation, Formal Analysis, Visualization: KS Writing – Original Draft: KS Writing – Review &amp; Editing: TC, DT, TK Resources: TC, DT Conceptualization, Methodology, Funding acquisition, Supervision: TC.</p>
  </notes><notes notes-type="competinginterests"><title>Competing interests</title>

      <p id="d2e1667">The contact author has declared that none of the authors has any competing interests.</p>
  </notes><notes notes-type="disclaimer"><title>Disclaimer</title>

      <p id="d2e1673">Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the funding agencies.Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. The authors bear the ultimate responsibility for providing appropriate place names. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.</p>
  </notes><ack><title>Acknowledgements</title><p id="d2e1682">This study has been partially funded by the European Union's Horizon 2020 research and innovation programme under grant agreement no. 856612 (EMME-CARE) and no. 857645 (NI4OS). The authors gratefully acknowledge the Gauss Centre for Supercomputing e.V. (<uri>https://www.gauss-centre.eu</uri>, last access: 6 November 2026) for funding this project by providing computing time through the John von Neumann Institute for Computing (NIC) on the GCS Supercomputer JUWELS <xref ref-type="bibr" rid="bib1.bibx4" id="paren.33"/> at Jülich Supercomputing Centre (JSC).</p></ack><notes notes-type="financialsupport"><title>Financial support</title>

      <p id="d2e1693">This research has been supported by the European Commission, Horizon 2020 Framework Programme (grant nos. 856612 and 857645).The article processing charges for this open-access publication were covered by the Forschungszentrum Jülich.</p>
  </notes><notes notes-type="reviewstatement"><title>Review statement</title>

      <p id="d2e1704">This paper was edited by Klaus Klingmüller and reviewed by Christian Guzman and one anonymous referee.</p>
  </notes><ref-list>
    <title>References</title>

      <ref id="bib1.bibx1"><label>Alvanos and Christoudias(2017a)</label><mixed-citation>Alvanos, M. and Christoudias, T.: GPU-accelerated atmospheric chemical kinetics in the ECHAM/MESSy (EMAC) Earth system model (version 2.52), Geosci. Model Dev., 10, 3679–3693, <ext-link xlink:href="https://doi.org/10.5194/gmd-10-3679-2017" ext-link-type="DOI">10.5194/gmd-10-3679-2017</ext-link>, 2017a.</mixed-citation></ref>
      <ref id="bib1.bibx2"><label>Alvanos and Christoudias(2017b)</label><mixed-citation>Alvanos, M. and Christoudias, T.: MEDINA: MECCA Development in Accelerators – KPP Fortran to CUDA Source-to-Source Pre-processor, Journal of Open Research Software, 5, 13, <ext-link xlink:href="https://doi.org/10.5334/jors.158" ext-link-type="DOI">10.5334/jors.158</ext-link>, 2017b.</mixed-citation></ref>
      <ref id="bib1.bibx3"><label>Alvanos and Christoudias(2019)</label><mixed-citation>Alvanos, M. and Christoudias, T.: Accelerating Atmospheric Chemical Kinetics for Climate Simulations, IEEE T. Parall. Distr., 30, 2396–2407, <ext-link xlink:href="https://doi.org/10.1109/TPDS.2019.2918798" ext-link-type="DOI">10.1109/TPDS.2019.2918798</ext-link>, 2019.</mixed-citation></ref>
      <ref id="bib1.bibx4"><label>Alvarez(2021)</label><mixed-citation>Alvarez, D.: JUWELS cluster and booster: exascale pathfinder with modular supercomputing architecture at juelich supercomputing centre, Journal of Large-Scale Research Facilities, 7, A183–A183, <ext-link xlink:href="https://doi.org/10.17815/jlsrf-7-183" ext-link-type="DOI">10.17815/jlsrf-7-183</ext-link>, 2021.</mixed-citation></ref>
      <ref id="bib1.bibx5"><label>Christou et al.(2016)Christou, Christoudias, Morillo, Alvarez, and Merx</label><mixed-citation>Christou, M., Christoudias, T., Morillo, J., Alvarez, D., and Merx, H.: Earth system modelling on system-level heterogeneous architectures: EMAC (version 2.42) on the Dynamical Exascale Entry Platform (DEEP), Geosci. Model Dev., 9, 3483–3491, <ext-link xlink:href="https://doi.org/10.5194/gmd-9-3483-2016" ext-link-type="DOI">10.5194/gmd-9-3483-2016</ext-link>, 2016.</mixed-citation></ref>
      <ref id="bib1.bibx6"><label>Christoudias(2026)</label><mixed-citation>Christoudias, T.: MEDINA v.2, Zenodo [code], <ext-link xlink:href="https://doi.org/10.5281/zenodo.19882619" ext-link-type="DOI">10.5281/zenodo.19882619</ext-link>, 2026.</mixed-citation></ref>
      <ref id="bib1.bibx7"><label>Christoudias and Sophocleous(2026)</label><mixed-citation>Christoudias, T. and Sophocleous, K.: GPU-accelerated atmospheric chemical kinetics using single precision in the ECHAM/MESSy (EMAC) model v2.55 with MEDINA v2.0: Data sources, Zenodo [data set], <ext-link xlink:href="https://doi.org/10.5281/zenodo.21442083" ext-link-type="DOI">10.5281/zenodo.21442083</ext-link>, 2026.</mixed-citation></ref>
      <ref id="bib1.bibx8"><label>Christoudias et al.(2021)Christoudias, Kirfel, Kerkweg, Taraborrelli, Moulard, Raffin, Azizi, Oord, and Werkhoven</label><mixed-citation>Christoudias, T., Kirfel, T., Kerkweg, A., Taraborrelli, D., Moulard, G.-E., Raffin, E., Azizi, V., van den Oord, G., and van Werkhoven, B.: GPU Optimizations for Atmospheric Chemical Kinetics, in: The International Conference on High Performance Computing in Asia-Pacific Region, Association for Computing Machinery, New York, NY, USA, 136–138, <ext-link xlink:href="https://doi.org/10.1145/3432261.3439863" ext-link-type="DOI">10.1145/3432261.3439863</ext-link>, 2021.</mixed-citation></ref>
      <ref id="bib1.bibx9"><label>Damian et al.(2002)Damian, Sandu, Damian, Potra, and Carmichael</label><mixed-citation>Damian, V., Sandu, A., Damian, M., Potra, F. A., and Carmichael, G. R.: The kinetic preprocessor KPP – a software environment for solving chemical kinetics, Comput. Chem. Eng., 26, 1567–1579, <ext-link xlink:href="https://doi.org/10.1016/S0098-1354(02)00128-X" ext-link-type="DOI">10.1016/S0098-1354(02)00128-X</ext-link>, 2002.</mixed-citation></ref>
      <ref id="bib1.bibx10"><label>Dongarra et al.(2020)Dongarra, Hittinger, Bell, Chacon, Falgout, Heroux, Hovland, Ng, Webster, and Wild</label><mixed-citation>Dongarra, J., Hittinger, J., Bell, J., Chacon, L., Falgout, R., Heroux, M., Hovland, P., Ng, E., Webster, C., and Wild, S.: Applied mathematics research for exascale computing, SIAM Rev., 62, 3–64, <ext-link xlink:href="https://doi.org/10.2172/1149042" ext-link-type="DOI">10.2172/1149042</ext-link>, 2020.</mixed-citation></ref>
      <ref id="bib1.bibx11"><label>Dreger et al.(2025)Dreger, Kirfel, Pozzer, Rosanka, Sander, and Taraborrelli</label><mixed-citation>Dreger, R., Kirfel, T., Pozzer, A., Rosanka, S., Sander, R., and Taraborrelli, D.: Optimized step size control within the Rosenbrock solvers for stiff chemical ordinary differential equation systems in KPP version 2.2.3_rs4, Geosci. Model Dev., 18, 4273–4291, <ext-link xlink:href="https://doi.org/10.5194/gmd-18-4273-2025" ext-link-type="DOI">10.5194/gmd-18-4273-2025</ext-link>, 2025.</mixed-citation></ref>
      <ref id="bib1.bibx12"><label>Hairer and Wanner(1991)</label><mixed-citation>Hairer, E. and Wanner, G.: Solving Ordinary Differential Equations II: Stiff and Differential-Algebraic Problems, Springer, Berlin, <ext-link xlink:href="https://doi.org/10.1007/978-3-662-09947-6" ext-link-type="DOI">10.1007/978-3-662-09947-6</ext-link>, 1991.</mixed-citation></ref>
      <ref id="bib1.bibx13"><label>Higham(2002)</label><mixed-citation>Higham, N. J.: Accuracy and Stability of Numerical Algorithms, SIAM, <ext-link xlink:href="https://doi.org/10.1137/1.9780898718027" ext-link-type="DOI">10.1137/1.9780898718027</ext-link>, 2002.</mixed-citation></ref>
      <ref id="bib1.bibx14"><label>IEEE(2008)</label><mixed-citation>IEEE: IEEE Standard for Floating-Point Arithmetic, Tech. rep., IEEE, Piscataway, NJ, USA, <ext-link xlink:href="https://doi.org/10.1109/IEEESTD.2008.4610935" ext-link-type="DOI">10.1109/IEEESTD.2008.4610935</ext-link>, 2008.</mixed-citation></ref>
      <ref id="bib1.bibx15"><label>Flato et al.(2013)Flato, Marotzke, Abiodun, Braconnot, Chou, Collins, Cox, Driouech, Emori, Eyring et al.</label><mixed-citation>Flato, G., Marotzke, J., Abiodun, B., Braconnot, P., Chou, S. C., Collins, W., Cox, P., Driouech, F., Emori, S., Eyring, V., Forest, C., Gleckler, P., Guilyardi, E., Jakob, C., Kattsov, V., Reason, C., and Rummukainen, M.: Evaluation of Climate Models, in: Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change, edited by: Stocker, T. F., Qin, D., Plattner, G.-K., Tignor, M., Allen, S. K., Boschung, J., Nauels, A., Xia, Y., Bex, V., and Midgley, P. M., Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA,  ISBN 9781107661820, <ext-link xlink:href="https://doi.org/10.1017/CBO9781107415324.020" ext-link-type="DOI">10.1017/CBO9781107415324.020</ext-link>, 2013.</mixed-citation></ref>
      <ref id="bib1.bibx16"><label>Jöckel et al.(2006)Jöckel, Tost, Pozzer, Brühl, Buchholz, Ganzeveld, Hoor, Kerkweg, Lawrence, and Sander</label><mixed-citation>Jöckel, P., Tost, H., Pozzer, A., Brühl, C., Buchholz, J., Ganzeveld, L., Hoor, P., Kerkweg, A., Lawrence, M. G., Sander, R., Steil, B., Stiller, G., Tanarhte, M., Taraborrelli, D., van Aardenne, J., and Lelieveld, J.: The atmospheric chemistry general circulation model ECHAM5/MESSy1: consistent simulation of ozone from the surface to the mesosphere, Atmos. Chem. Phys., 6, 5067–5104, <ext-link xlink:href="https://doi.org/10.5194/acp-6-5067-2006" ext-link-type="DOI">10.5194/acp-6-5067-2006</ext-link>, 2006.</mixed-citation></ref>
      <ref id="bib1.bibx17"><label>Jöckel et al.(2010)Jöckel, Kerkweg, Pozzer, Sander, Tost, Riede, Baumgaertner, Gromov, and Kern</label><mixed-citation>Jöckel, P., Kerkweg, A., Pozzer, A., Sander, R., Tost, H., Riede, H., Baumgaertner, A., Gromov, S., and Kern, B.: Development cycle 2 of the Modular Earth Submodel System (MESSy2), Geosci. Model Dev., 3, 717–752, <ext-link xlink:href="https://doi.org/10.5194/gmd-3-717-2010" ext-link-type="DOI">10.5194/gmd-3-717-2010</ext-link>, 2010.</mixed-citation></ref>
      <ref id="bib1.bibx18"><label>Jöckel et al.(2016)Jöckel, Tost, Pozzer, Kunze, Kirner, Brenninkmeijer, Brinkop, Cai, Dyroff, Eckstein et al.</label><mixed-citation>Jöckel, P., Tost, H., Pozzer, A., Kunze, M., Kirner, O., Brenninkmeijer, C. A. M., Brinkop, S., Cai, D. S., Dyroff, C., Eckstein, J., Frank, F., Garny, H., Gottschaldt, K.-D., Graf, P., Grewe, V., Kerkweg, A., Kern, B., Matthes, S., Mertens, M., Meul, S., Neumaier, M., Nützel, M., Oberländer-Hayn, S., Ruhnke, R., Runde, T., Sander, R., Scharffe, D., and Zahn, A.: Earth System Chemistry integrated Modelling (ESCiMo) with the Modular Earth Submodel System (MESSy) version 2.51, Geosci. Model Dev., 9, 1153–1200, <ext-link xlink:href="https://doi.org/10.5194/gmd-9-1153-2016" ext-link-type="DOI">10.5194/gmd-9-1153-2016</ext-link>, 2016.</mixed-citation></ref>
      <ref id="bib1.bibx19"><label>Kerkweg et al.(2025)Kerkweg, Kirfel, Do, Griessbach, Jöckel, and Taraborrelli</label><mixed-citation>Kerkweg, A., Kirfel, T., Do, D. H., Griessbach, S., Jöckel, P., and Taraborrelli, D.: The MESSy DWARF (based on MESSy v2.55.2), Geosci. Model Dev., 18, 1265–1286, <ext-link xlink:href="https://doi.org/10.5194/gmd-18-1265-2025" ext-link-type="DOI">10.5194/gmd-18-1265-2025</ext-link>, 2025.</mixed-citation></ref>
      <ref id="bib1.bibx20"><label>Linford et al.(2009)</label><mixed-citation>Linford, J. C., Michalakes, J., Vachharajani, M., and Sandu, A.: Multi-core acceleration of chemical kinetics for simulation and prediction, in: Proceedings of the Conference on High Performance Computing Networking, Storage and Analysis, SC ’09, Association for Computing Machinery, New York, NY, USA, 11 pp., <ext-link xlink:href="https://doi.org/10.1145/1654059.1654067" ext-link-type="DOI">10.1145/1654059.1654067</ext-link>, 2009.</mixed-citation></ref>
      <ref id="bib1.bibx21"><label>NVIDIA(2013)</label><mixed-citation>NVIDIA: CUDA C Programming Guide, <uri>https://academicweb.nd.edu/~zxu2/acms60212-40212/CUDA_C_Programming_Guide.pdf</uri> (last access: 6 November 2026), 2013.</mixed-citation></ref>
      <ref id="bib1.bibx22"><label>NVIDIA(2023a)</label><mixed-citation>NVIDIA: NVIDIA Nsight Compute User Guide, <uri>https://docs.nvidia.com/nsight-compute/2023.1/NsightCompute/index.html</uri>  (last access: 6 November 2026), 2023a.</mixed-citation></ref>
      <ref id="bib1.bibx23"><label>NVIDIA(2023b)</label><mixed-citation>NVIDIA: NVIDIA Nsight Systems User Guide, <uri>https://docs.nvidia.com/nsight-systems/UserGuide/</uri>  (last access: 6 November 2026), 2023b.</mixed-citation></ref>
      <ref id="bib1.bibx24"><label>Palmer(2014)</label><mixed-citation>Palmer, T. N.: More reliable forecasts with less precise computations: A fast-track route to cloud-resolved weather and climate simulators?, Philos. T. R. Soc. A, 372, <ext-link xlink:href="https://doi.org/10.1098/rsta.2013.0391" ext-link-type="DOI">10.1098/rsta.2013.0391</ext-link>, 2014.</mixed-citation></ref>
      <ref id="bib1.bibx25"><label>Roeckner et al.(2006)Roeckner, Brokopf, Esch, Giorgetta, Hagemann, Kornblueh, Manzini, Schlese, and Schulzweida</label><mixed-citation>Roeckner, E., Brokopf, R., Esch, M., Giorgetta, M., Hagemann, S., Kornblueh, L., Manzini, E., Schlese, U., and Schulzweida, U.: Sensitivity of simulated climate to horizontal and vertical resolution in the ECHAM5 atmosphere model, J. Climate, 19, 3771–3791, <ext-link xlink:href="https://doi.org/10.1175/JCLI3824.1" ext-link-type="DOI">10.1175/JCLI3824.1</ext-link>, 2006.</mixed-citation></ref>
      <ref id="bib1.bibx26"><label>Sander et al.(2019)Sander, Baumgaertner, Cabrera-Perez, Frank, Gromov, Grooß, Harder, Huijnen, Jöckel, Karydis, Niemeyer, Pozzer, Riede, Schultz, Taraborrelli, and Tauer</label><mixed-citation>Sander, R., Baumgaertner, A., Cabrera-Perez, D., Frank, F., Gromov, S., Grooß, J.-U., Harder, H., Huijnen, V., Jöckel, P., Karydis, V. A., Niemeyer, K. E., Pozzer, A., Riede, H., Schultz, M. G., Taraborrelli, D., and Tauer, S.: The community atmospheric chemistry box model CAABA/MECCA-4.0, Geosci. Model Dev., 12, 1365–1385, <ext-link xlink:href="https://doi.org/10.5194/gmd-12-1365-2019" ext-link-type="DOI">10.5194/gmd-12-1365-2019</ext-link>, 2019.</mixed-citation></ref>
      <ref id="bib1.bibx27"><label>Sandu and Sander(2006)</label><mixed-citation>Sandu, A. and Sander, R.: Technical note: Simulating chemical systems in Fortran90 and Matlab with the Kinetic PreProcessor KPP-2.1, Atmos. Chem. Phys., 6, 187–195, <ext-link xlink:href="https://doi.org/10.5194/acp-6-187-2006" ext-link-type="DOI">10.5194/acp-6-187-2006</ext-link>, 2006.</mixed-citation></ref>
      <ref id="bib1.bibx28"><label>Sophocleous and Christoudias(2022)</label><mixed-citation>Sophocleous, K. and Christoudias, T.: Reduced-Precision Chemical Kinetics in Atmospheric Models, Atmosphere, 13, 1418, <ext-link xlink:href="https://doi.org/10.3390/atmos13091418" ext-link-type="DOI">10.3390/atmos13091418</ext-link>, 2022.</mixed-citation></ref>
      <ref id="bib1.bibx29"><label>Váňa et al.(2017)</label><mixed-citation>Váňa, F., Düben, P., Lang, S., Palmer, T., Leutbecher, M., Salmond, D., and Carver, G.: Single precision in weather forecasting models: An4 evaluation with the IFS, Mon. Weather Rev., 145, 495–502, <ext-link xlink:href="https://doi.org/10.1175/MWR-D-16-0228.1" ext-link-type="DOI">10.1175/MWR-D-16-0228.1</ext-link>, 2017.</mixed-citation></ref>
      <ref id="bib1.bibx30"><label>Whitehead and Fit-Florea(2017)</label><mixed-citation>Whitehead, N. and Fit-Florea, A.: Precision &amp; Performance: Floating Point and IEEE-754 Compliance for NVIDIA GPUs, Tech. rep., NVIDIA, <uri>https://developer.download.nvidia.com/assets/cuda/files/NVIDIA-CUDA-Floating-Point.pdf</uri>  (last access: 6 November 2026), 2017.</mixed-citation></ref>
      <ref id="bib1.bibx31"><label>Williams et al.(2009)Williams, Waterman, and Patterson</label><mixed-citation>Williams, S., Waterman, A., and Patterson, D.: Roofline: An insightful visual performance model for multicore architectures, Commun. ACM, 52, 65–76, <ext-link xlink:href="https://doi.org/10.1145/1498765.1498785" ext-link-type="DOI">10.1145/1498765.1498785</ext-link>, 2009.</mixed-citation></ref>

  </ref-list></back>
    <!--<article-title-html>GPU-accelerated atmospheric chemical kinetics using single precision in the ECHAM/MESSy (EMAC) model v2.55 with MEDINA v2.0</article-title-html>
<abstract-html/>
<ref-html id="bib1.bib1"><label>Alvanos and Christoudias(2017a)</label><mixed-citation>
      
Alvanos, M. and Christoudias, T.: GPU-accelerated atmospheric chemical kinetics in the ECHAM/MESSy (EMAC) Earth system model (version 2.52), Geosci. Model Dev., 10, 3679–3693, <a href="https://doi.org/10.5194/gmd-10-3679-2017" target="_blank">https://doi.org/10.5194/gmd-10-3679-2017</a>, 2017a.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib2"><label>Alvanos and Christoudias(2017b)</label><mixed-citation>
      
Alvanos, M. and Christoudias, T.: MEDINA: MECCA Development in Accelerators –
KPP Fortran to CUDA Source-to-Source Pre-processor, Journal of Open Research
Software, 5, 13, <a href="https://doi.org/10.5334/jors.158" target="_blank">https://doi.org/10.5334/jors.158</a>, 2017b.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib3"><label>Alvanos and Christoudias(2019)</label><mixed-citation>
      
Alvanos, M. and Christoudias, T.: Accelerating Atmospheric Chemical Kinetics
for Climate Simulations, IEEE T. Parall. Distr., 30, 2396–2407,
<a href="https://doi.org/10.1109/TPDS.2019.2918798" target="_blank">https://doi.org/10.1109/TPDS.2019.2918798</a>, 2019.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib4"><label>Alvarez(2021)</label><mixed-citation>
      
Alvarez, D.: JUWELS cluster and booster: exascale pathfinder with modular
supercomputing architecture at juelich supercomputing centre, Journal of
Large-Scale Research Facilities, 7, A183–A183,
<a href="https://doi.org/10.17815/jlsrf-7-183" target="_blank">https://doi.org/10.17815/jlsrf-7-183</a>, 2021.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib5"><label>Christou et al.(2016)Christou, Christoudias, Morillo, Alvarez, and
Merx</label><mixed-citation>
      
Christou, M., Christoudias, T., Morillo, J., Alvarez, D., and Merx, H.: Earth system modelling on system-level heterogeneous architectures: EMAC (version 2.42) on the Dynamical Exascale Entry Platform (DEEP), Geosci. Model Dev., 9, 3483–3491, <a href="https://doi.org/10.5194/gmd-9-3483-2016" target="_blank">https://doi.org/10.5194/gmd-9-3483-2016</a>, 2016.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib6"><label>Christoudias(2026)</label><mixed-citation>
      
Christoudias, T.: MEDINA v.2, Zenodo [code], <a href="https://doi.org/10.5281/zenodo.19882619" target="_blank">https://doi.org/10.5281/zenodo.19882619</a>, 2026.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib7"><label>Christoudias and Sophocleous(2026)</label><mixed-citation>
      
Christoudias, T. and Sophocleous, K.: GPU-accelerated atmospheric chemical
kinetics using single precision in the ECHAM/MESSy (EMAC) model v2.55 with
MEDINA v2.0: Data sources, Zenodo [data set], <a href="https://doi.org/10.5281/zenodo.21442083" target="_blank">https://doi.org/10.5281/zenodo.21442083</a>, 2026.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib8"><label>Christoudias et al.(2021)Christoudias, Kirfel, Kerkweg, Taraborrelli,
Moulard, Raffin, Azizi, Oord, and Werkhoven</label><mixed-citation>
      
Christoudias, T., Kirfel, T., Kerkweg, A., Taraborrelli, D., Moulard, G.-E.,
Raffin, E., Azizi, V., van den Oord, G., and van Werkhoven, B.: GPU
Optimizations for Atmospheric Chemical Kinetics, in: The International
Conference on High Performance Computing in Asia-Pacific Region, Association for Computing Machinery, New York, NY, USA,
136–138, <a href="https://doi.org/10.1145/3432261.3439863" target="_blank">https://doi.org/10.1145/3432261.3439863</a>, 2021.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib9"><label>Damian et al.(2002)Damian, Sandu, Damian, Potra, and
Carmichael</label><mixed-citation>
      
Damian, V., Sandu, A., Damian, M., Potra, F. A., and Carmichael, G. R.: The
kinetic preprocessor KPP – a software environment for solving chemical
kinetics, Comput. Chem. Eng., 26, 1567–1579,
<a href="https://doi.org/10.1016/S0098-1354(02)00128-X" target="_blank">https://doi.org/10.1016/S0098-1354(02)00128-X</a>, 2002.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib10"><label>Dongarra et al.(2020)Dongarra, Hittinger, Bell, Chacon, Falgout,
Heroux, Hovland, Ng, Webster, and Wild</label><mixed-citation>
      
Dongarra, J., Hittinger, J., Bell, J., Chacon, L., Falgout, R., Heroux, M.,
Hovland, P., Ng, E., Webster, C., and Wild, S.: Applied mathematics research
for exascale computing, SIAM Rev., 62, 3–64,
<a href="https://doi.org/10.2172/1149042" target="_blank">https://doi.org/10.2172/1149042</a>, 2020.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib11"><label>Dreger et al.(2025)Dreger, Kirfel, Pozzer, Rosanka, Sander, and
Taraborrelli</label><mixed-citation>
      
Dreger, R., Kirfel, T., Pozzer, A., Rosanka, S., Sander, R., and Taraborrelli, D.: Optimized step size control within the Rosenbrock solvers for stiff chemical ordinary differential equation systems in KPP version 2.2.3_rs4, Geosci. Model Dev., 18, 4273–4291, <a href="https://doi.org/10.5194/gmd-18-4273-2025" target="_blank">https://doi.org/10.5194/gmd-18-4273-2025</a>, 2025.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib12"><label>Hairer and Wanner(1991)</label><mixed-citation>
      
Hairer, E. and Wanner, G.: Solving Ordinary Differential Equations II: Stiff
and Differential-Algebraic Problems, Springer, Berlin,
<a href="https://doi.org/10.1007/978-3-662-09947-6" target="_blank">https://doi.org/10.1007/978-3-662-09947-6</a>, 1991.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib13"><label>Higham(2002)</label><mixed-citation>
      
Higham, N. J.: Accuracy and Stability of Numerical Algorithms, SIAM,
<a href="https://doi.org/10.1137/1.9780898718027" target="_blank">https://doi.org/10.1137/1.9780898718027</a>, 2002.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib14"><label>IEEE(2008)</label><mixed-citation>
      
IEEE: IEEE Standard for Floating-Point Arithmetic, Tech. rep., IEEE,
Piscataway, NJ, USA, <a href="https://doi.org/10.1109/IEEESTD.2008.4610935" target="_blank">https://doi.org/10.1109/IEEESTD.2008.4610935</a>, 2008.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib15"><label>Flato et al.(2013)Flato, Marotzke, Abiodun, Braconnot, Chou, Collins,
Cox, Driouech, Emori, Eyring et al.</label><mixed-citation>
      
Flato, G., Marotzke, J., Abiodun, B., Braconnot, P., Chou, S. C., Collins, W., Cox, P., Driouech, F., Emori, S., Eyring, V., Forest, C., Gleckler, P., Guilyardi, E., Jakob, C., Kattsov, V., Reason, C., and Rummukainen, M.: Evaluation of Climate Models, in: Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change, edited by: Stocker, T. F., Qin, D., Plattner, G.-K., Tignor, M., Allen, S. K., Boschung, J., Nauels, A., Xia, Y., Bex, V., and Midgley, P. M., Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA,  ISBN 9781107661820, <a href="https://doi.org/10.1017/CBO9781107415324.020" target="_blank">https://doi.org/10.1017/CBO9781107415324.020</a>, 2013.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib16"><label>Jöckel et al.(2006)Jöckel, Tost, Pozzer, Brühl, Buchholz,
Ganzeveld, Hoor, Kerkweg, Lawrence, and Sander</label><mixed-citation>
      
Jöckel, P., Tost, H., Pozzer, A., Brühl, C., Buchholz, J., Ganzeveld, L., Hoor, P., Kerkweg, A., Lawrence, M. G., Sander, R., Steil, B., Stiller, G., Tanarhte, M., Taraborrelli, D., van Aardenne, J., and Lelieveld, J.: The atmospheric chemistry general circulation model ECHAM5/MESSy1: consistent simulation of ozone from the surface to the mesosphere, Atmos. Chem. Phys., 6, 5067–5104, <a href="https://doi.org/10.5194/acp-6-5067-2006" target="_blank">https://doi.org/10.5194/acp-6-5067-2006</a>, 2006.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib17"><label>Jöckel et al.(2010)Jöckel, Kerkweg, Pozzer, Sander, Tost,
Riede, Baumgaertner, Gromov, and Kern</label><mixed-citation>
      
Jöckel, P., Kerkweg, A., Pozzer, A., Sander, R., Tost, H., Riede, H., Baumgaertner, A., Gromov, S., and Kern, B.: Development cycle 2 of the Modular Earth Submodel System (MESSy2), Geosci. Model Dev., 3, 717–752, <a href="https://doi.org/10.5194/gmd-3-717-2010" target="_blank">https://doi.org/10.5194/gmd-3-717-2010</a>, 2010.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib18"><label>Jöckel et al.(2016)Jöckel, Tost, Pozzer, Kunze, Kirner,
Brenninkmeijer, Brinkop, Cai, Dyroff, Eckstein et al.</label><mixed-citation>
      
Jöckel, P., Tost, H., Pozzer, A., Kunze, M., Kirner, O., Brenninkmeijer, C. A. M., Brinkop, S., Cai, D. S., Dyroff, C., Eckstein, J., Frank, F., Garny, H., Gottschaldt, K.-D., Graf, P., Grewe, V., Kerkweg, A., Kern, B., Matthes, S., Mertens, M., Meul, S., Neumaier, M., Nützel, M., Oberländer-Hayn, S., Ruhnke, R., Runde, T., Sander, R., Scharffe, D., and Zahn, A.: Earth System Chemistry integrated Modelling (ESCiMo) with the Modular Earth Submodel System (MESSy) version 2.51, Geosci. Model Dev., 9, 1153–1200, <a href="https://doi.org/10.5194/gmd-9-1153-2016" target="_blank">https://doi.org/10.5194/gmd-9-1153-2016</a>, 2016.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib19"><label>Kerkweg et al.(2025)Kerkweg, Kirfel, Do, Griessbach, Jöckel, and
Taraborrelli</label><mixed-citation>
      
Kerkweg, A., Kirfel, T., Do, D. H., Griessbach, S., Jöckel, P., and Taraborrelli, D.: The MESSy DWARF (based on MESSy v2.55.2), Geosci. Model Dev., 18, 1265–1286, <a href="https://doi.org/10.5194/gmd-18-1265-2025" target="_blank">https://doi.org/10.5194/gmd-18-1265-2025</a>, 2025.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib20"><label>Linford et al.(2009)</label><mixed-citation>
      
Linford, J. C., Michalakes, J., Vachharajani, M., and Sandu, A.: Multi-core acceleration of chemical kinetics for simulation and prediction,
in: Proceedings of the Conference on High Performance Computing Networking, Storage and Analysis, SC ’09, Association for
Computing Machinery, New York, NY, USA, 11 pp., <a href="https://doi.org/10.1145/1654059.1654067" target="_blank">https://doi.org/10.1145/1654059.1654067</a>, 2009.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib21"><label>NVIDIA(2013)</label><mixed-citation>
      
NVIDIA: CUDA C Programming Guide,
<a href="https://academicweb.nd.edu/~zxu2/acms60212-40212/CUDA_C_Programming_Guide.pdf" target="_blank"/> (last access: 6 November 2026),
2013.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib22"><label>NVIDIA(2023a)</label><mixed-citation>
      
NVIDIA: NVIDIA Nsight Compute User Guide,
<a href="https://docs.nvidia.com/nsight-compute/2023.1/NsightCompute/index.html" target="_blank"/>  (last access: 6 November 2026),
2023a.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib23"><label>NVIDIA(2023b)</label><mixed-citation>
      
NVIDIA: NVIDIA Nsight Systems User Guide,
<a href="https://docs.nvidia.com/nsight-systems/UserGuide/" target="_blank"/>  (last access: 6 November 2026),
2023b.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib24"><label>Palmer(2014)</label><mixed-citation>
      
Palmer, T. N.: More reliable forecasts with less precise computations: A
fast-track route to cloud-resolved weather and climate simulators?,
Philos. T. R. Soc. A, 372,
<a href="https://doi.org/10.1098/rsta.2013.0391" target="_blank">https://doi.org/10.1098/rsta.2013.0391</a>, 2014.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib25"><label>Roeckner et al.(2006)Roeckner, Brokopf, Esch, Giorgetta, Hagemann,
Kornblueh, Manzini, Schlese, and Schulzweida</label><mixed-citation>
      
Roeckner, E., Brokopf, R., Esch, M., Giorgetta, M., Hagemann, S., Kornblueh,
L., Manzini, E., Schlese, U., and Schulzweida, U.: Sensitivity of simulated
climate to horizontal and vertical resolution in the ECHAM5 atmosphere model,
J. Climate, 19, 3771–3791, <a href="https://doi.org/10.1175/JCLI3824.1" target="_blank">https://doi.org/10.1175/JCLI3824.1</a>, 2006.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib26"><label>Sander et al.(2019)Sander, Baumgaertner, Cabrera-Perez, Frank,
Gromov, Grooß, Harder, Huijnen, Jöckel, Karydis, Niemeyer, Pozzer,
Riede, Schultz, Taraborrelli, and Tauer</label><mixed-citation>
      
Sander, R., Baumgaertner, A., Cabrera-Perez, D., Frank, F., Gromov, S., Grooß, J.-U., Harder, H., Huijnen, V., Jöckel, P., Karydis, V. A., Niemeyer, K. E., Pozzer, A., Riede, H., Schultz, M. G., Taraborrelli, D., and Tauer, S.: The community atmospheric chemistry box model CAABA/MECCA-4.0, Geosci. Model Dev., 12, 1365–1385, <a href="https://doi.org/10.5194/gmd-12-1365-2019" target="_blank">https://doi.org/10.5194/gmd-12-1365-2019</a>, 2019.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib27"><label>Sandu and Sander(2006)</label><mixed-citation>
      
Sandu, A. and Sander, R.: Technical note: Simulating chemical systems in Fortran90 and Matlab with the Kinetic PreProcessor KPP-2.1, Atmos. Chem. Phys., 6, 187–195, <a href="https://doi.org/10.5194/acp-6-187-2006" target="_blank">https://doi.org/10.5194/acp-6-187-2006</a>, 2006.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib28"><label>Sophocleous and Christoudias(2022)</label><mixed-citation>
      
Sophocleous, K. and Christoudias, T.: Reduced-Precision Chemical Kinetics in
Atmospheric Models, Atmosphere, 13, 1418, <a href="https://doi.org/10.3390/atmos13091418" target="_blank">https://doi.org/10.3390/atmos13091418</a>, 2022.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib29"><label>Váňa et al.(2017)</label><mixed-citation>
      
Váňa, F., Düben, P., Lang, S., Palmer, T., Leutbecher, M., Salmond, D., and Carver, G.: Single precision in weather forecasting models: An4
evaluation with the IFS, Mon. Weather Rev., 145, 495–502, <a href="https://doi.org/10.1175/MWR-D-16-0228.1" target="_blank">https://doi.org/10.1175/MWR-D-16-0228.1</a>, 2017.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib30"><label>Whitehead and Fit-Florea(2017)</label><mixed-citation>
      
Whitehead, N. and Fit-Florea, A.: Precision &amp; Performance: Floating Point and
IEEE-754 Compliance for NVIDIA GPUs, Tech. rep., NVIDIA,
<a href="https://developer.download.nvidia.com/assets/cuda/files/NVIDIA-CUDA-Floating-Point.pdf" target="_blank"/>  (last access: 6 November 2026),
2017.

    </mixed-citation></ref-html>
<ref-html id="bib1.bib31"><label>Williams et al.(2009)Williams, Waterman, and
Patterson</label><mixed-citation>
      
Williams, S., Waterman, A., and Patterson, D.: Roofline: An insightful visual
performance model for multicore architectures, Commun. ACM, 52,
65–76, <a href="https://doi.org/10.1145/1498765.1498785" target="_blank">https://doi.org/10.1145/1498765.1498785</a>, 2009.

    </mixed-citation></ref-html>--></article>
