Exascale computing is often described in terms of peak floating-point performance. But for applications such as weather and climate modeling, raw compute capacity does not tell the whole story. These simulations move large volumes of data through memory, and their performance depends on how effectively software and hardware work together. A roadmap developed around real model workloads offers a useful lesson for anyone planning high-performance computing (HPC) infrastructure: define success by the application’s useful output, not by a headline processor rating.
Why peak performance is an incomplete target
For decades, growth in transistor density and computer performance supported increasingly complex scientific simulations. Yet the relationship between peak speed and useful application performance varies by workload. Some scientific codes perform many operations on data already in memory. Weather and climate models, by contrast, often use grid-based numerical methods that repeatedly access neighboring values to approximate physical processes.
These methods typically have low arithmetic intensity: relatively few calculations are performed for each unit of data moved from memory. As a result, the speed of the floating-point units may not be the main constraint. The memory subsystem—its bandwidth and the efficiency with which a program uses it—can limit the rate at which a simulation advances.
This distinction matters beyond meteorology. The roadmap notes that similar performance challenges arise in other applications built around grid-based solvers of partial differential equations, including fluid dynamics, combustion, geophysics and solid mechanics. In each case, a system optimized for peak calculations may not deliver corresponding gains in time to solution.
Exascale computing needs an application-level goal
The roadmap sets a demanding target for global weather and climate simulations: approximately 1-kilometer horizontal resolution, 180 vertical levels, a 0.5-minute time resolution, and a throughput of one simulated year per wall-clock day (1 SYPD). The proposed configuration also includes coupled Earth-system components such as land, ocean, sea ice and ocean waves.
Higher resolution can help models represent smaller-scale processes. In the paper’s scientific context, kilometer-scale simulations were seen as an important intermediate step toward better representation of convective systems and other complex processes. The ambition was not simply to run a larger model, but to reach a useful simulation rate while accounting for more of the interacting Earth system.
The baseline results, based on systems available in 2016–2017, showed a substantial gap. A near-global COSMO simulation on 4,888 GPU-accelerated nodes achieved 0.043 SYPD at 0.93-kilometer grid spacing. A global IFS simulation on the same supercomputer, using CPUs, achieved 0.088 SYPD at 1.25-kilometer resolution. The authors estimated that these systems fell short of the target by factors of roughly 100 to 250, depending on the model and configuration.
Those figures were not a simple measure of hardware capability. The models had not been optimized for every aspect of the target workload, and the estimates included differences in resolution, time step, coupling and physical processes. The comparison nevertheless illustrated why a workload-specific baseline is essential: it shows the distance between current application throughput and the rate required for the intended scientific use.
Measure memory use, not just operations
To assess performance more directly, the authors proposed Memory Use Efficiency (MUE), a metric that combines two factors:
- Memory I/O efficiency compares the data transfers required by an implementation with the algorithm’s theoretical minimum.
- Bandwidth efficiency compares the achieved memory bandwidth with the bandwidth available from the system.
MUE is the product of these efficiencies. It helps distinguish between avoidable data movement in software and limits in how effectively the hardware memory subsystem serves the program. In the analyzed COSMO runs, MUE was 0.67, with memory I/O efficiency of 0.88 and bandwidth efficiency of 0.75. The analysis suggested that the code was relatively close to its data-movement lower bound, while the memory subsystem was not being used as effectively as possible for its access patterns.
That distinction is valuable for engineering teams. If an application already minimizes unnecessary transfers, further optimization may depend on processor architecture, memory behavior or the way software controls access to memory—not simply on rewriting numerical calculations. A useful performance metric should help identify which part of the system needs attention.
A roadmap combines software, hardware and scaling
The paper outlined several ways to reduce the measured performance gap. For COSMO, the authors identified inefficiencies associated with mapping a Cartesian grid onto a near-global latitude-longitude grid. A model designed around a suitable global grid could reduce redundant grid points and avoid some of the time-step constraints in the baseline setup.
They also pointed to improvements in memory bandwidth and parallel scaling. Their estimates suggested that addressing these factors, alongside software changes, could substantially narrow the shortfall. These were roadmap projections, not guaranteed results: achieving them depended on coordinated work across application design, programming frameworks, processors, accelerators and system configuration.
That coordination is the essence of codesign. Application developers need to characterize real workloads, while hardware designers need to understand their data access patterns. Domain-specific software frameworks can help port structured-grid methods across different architectures without treating every model as a one-off implementation.
Plan for data and energy as well as compute
Faster simulations also create operational challenges. The paper warned that storing all output from much higher-resolution climate runs using existing workflows could require an impractical expansion in storage. It therefore pointed to in-situ analysis and revised workflows: analyze data during execution and record simulation checkpoints rather than writing every result to storage for later processing. Such approaches can reduce storage traffic, but they require methods for recreating model trajectories.
Energy is another part of the performance equation. The roadmap compared the energy costs of its baseline simulations and argued for reaching higher throughput without a corresponding major increase in power consumption. For HPC planners, the implication is that performance targets should account for useful work, energy and data handling together.
Practical implications for HPC decision-makers
The central lesson is to evaluate computing systems against the behavior of the applications they must run. Peak FLOP/s remains informative, but it is not a substitute for measuring time to solution, memory efficiency, scalability, energy use and data-management requirements.
For teams developing or procuring infrastructure, this means establishing representative workload benchmarks early, profiling data movement, and involving application and hardware specialists in architecture decisions. The weather and climate roadmap demonstrates how a clearly defined scientific target can expose bottlenecks and guide software-hardware codesign. Exascale computing becomes meaningful not when a machine reaches a nominal speed, but when it delivers the required application throughput within practical power and data constraints.