Local DrivAer65M benchmark | 2026-11-11 | 65,334,765 cells
OpenFOAM v2606, double precision, A100 PCIe 80 GB x4 (4 MPI ranks), and Xeon Gold 6430 x2 (64 physical cores / MPI ranks). Restart at 400; timing includes MPI startup/shutdown and four writes.
Each CSV row is one SIMPLE iteration. Time offsets use a monotonic clock relative to MPI launch. For the restart window subtract end_offset_s at step 410 from the value at 510. For the public comparison from zero, if present, use step 110 minus step 10. Force coefficients are dimensionless; residuals and continuity are taken from solver output. Each field's n column is its inner iteration count.
benchmark-data.json reports total time, initial and ordinary step times and the unweighted relative L2 differences at step 840 after alignment by global cell index. Two GPU runs and one CPU reference. Independent validation is against the preceding GPU run. The publicly reported SPUMA times originate from Bna et al., CPC 321, 110009, DOI 10.1016/j.cpc.2025.110009, Table 13; their relative stopping tolerances differ from our fixed inner counts. The separate mesh download is https://zenodo.org/records/17063427 .
Source and patch package for the A100 target are recorded in the companion A100 experiment archive. This dataset contains measurements, not a full solver installation.
