Workload-Based Hardware Selection for Separate CFD and DEM Benchmarks on CPUs and GPUs

Industrial CFD and DEM benchmarks were reanalyzed to examine how workload characteristics and hardware constraints affect GPU performance in Ansys Fluent and Ansys Rocky. The dataset comprised separate CFD and DEM calculations, not coupled fluid–particle simulations. Fluent cases included a 24-million-cell combustor, a 4-million-cell mixing tank, and a 140-million-cell external-aerodynamics model evaluated using fixed iteration counts. Rocky cases covered spherical and polyhedral particles, particle-count and particle-size sweeps, and selected multi-GPU configurations. Recorded wall-clock times were used to calculate within-case speedups and scaling efficiencies, while out-of-memory (OOM) outcomes were treated separately as feasibility results. Peak FP32 throughput alone did not explain solver performance. Completed workloads showed different sensitivities to memory bandwidth, memory type, solver path, workload size, particle shape, and precision regime. Smaller workloads showed limited separation among high-performance GPUs, whereas larger or more demanding cases benefited more clearly from memory bandwidth and additional GPUs. OOM results showed that VRAM capacity acts primarily as a feasibility constraint, and multi-GPU execution both reduced runtime and enabled otherwise infeasible workloads. The results support a workload-aware hardware-selection framework in which memory feasibility is assessed before runtime and scaling. Because each configuration had a single recorded timing, detailed performance mechanisms and statistical rankings remain unresolved.