The First OpenFOAM HPC Challenge (OHC-1)
First-ever OpenFOAM HPC Challenge analyzed 237 CFD simulations, showing software optimizations can cut energy use by 28%.
The OpenFOAM HPC Technical Committee (HPCTC) has published results from the inaugural OpenFOAM HPC Challenge (OHC-1), establishing the first comprehensive performance benchmark for the popular open-source computational fluid dynamics (CFD) software. The challenge aimed to capture a snapshot of OpenFOAM's performance on modern hardware by having participants run a standardized incompressible steady-state RANS simulation—the open-closed cooling DrivAer (occDrivAer) configuration. Contributors submitted results in two tracks: a hardware track using a reference setup to test raw system performance, and a software track allowing optimizations like modified solvers and accelerator offloading.
In total, 237 valid data points were submitted by 12 organizations, with 175 entries in the hardware track and 62 in the software track. The hardware track spanned an impressive range, covering 25 distinct CPU models from AMD, Intel, and ARM, with runs scaling from single-node setups to massive 256-node clusters using 32,768 CPU cores. Execution times varied dramatically from 7.8 minutes to 65.7 hours, with energy consumption ranging from 2.1 to 236.9 kWh. Analysis revealed that on-package high-bandwidth memory (HBM) is becoming crucial for single-node performance on next-generation CPUs.
The software track delivered even more significant insights, demonstrating that optimized code can substantially outperform even the best hardware. Software optimizations achieved up to 28% lower energy consumption per iteration, 17% higher maximum performance per node, and a remarkable 72% reduction in minimum time per iteration compared to the best hardware-track results. Full GPU ports and selective-memory optimizations proved particularly effective, dominating the performance range and showing clear paths for future CFD efficiency improvements.
- 237 CFD simulations benchmarked across 12 contributors, with hardware tests covering 25 CPU models and scaling to 256 nodes (32,768 cores)
- Software optimizations outperformed best hardware results with 28% lower energy per iteration and 72% shorter minimum time per iteration
- Full GPU ports and selective-memory optimizations led performance gains, while on-package HBM proved critical for next-gen CPU single-node performance
Why It Matters
Provides crucial performance benchmarks for engineering teams running large-scale CFD simulations, showing clear optimization paths for reducing compute costs and energy consumption.