Benchmarks
Every benchmark is reported on two axes: accuracy against a reference simulator (OPM Flow today; others later) and runtime across single-CPU, multi-CPU, and GPU. Numbers are measured on real runs at matched core counts.
For physics-level validation of individual capabilities (dual porosity, wells, aquifers, …), see Physics Validation.
Accuracy vs. reference
Validated on the mass-conservation invariants (cumulative production) and on instantaneous field rates and pressure, against OPM Flow run on the same deck.
| Model | Cells | Reference | Accuracy metric | Result |
|---|---|---|---|---|
| SPE-1 | 300 | SPE ref / OPM | Cumulative production, field rates | Matches reference |
| SPE-9 | 9,000 | OPM Flow | FOPT / FWPT, mass balance | Mass exact; FOPT within ~0.8% |
| Norne | 44,431 | OPM Flow 2021.10 | Cumulative FOPT/FGPT/FWPT, FPR | Full-history match |
| SPE-10 | 1.12M | OPM Flow | Producer rates, FPR | Within ~1–3% |
Runtime by compute tier
BirSim runs the same physics on a pluggable CPU / MPI / GPU backend. Runtimes are at matched CPU core counts against OPM Flow; field-scale GPU results are being finalized.
| Model | Cells | Single-CPU | Multi-CPU (MPI) | GPU | Note |
|---|---|---|---|---|---|
| SPE-1 | 300 | < 1 s | — | in prep | Too small to parallelize |
| SPE-9 | 9,000 | 45 s | — | in prep | 2-core; up to 13.7× faster than the first build |
| Norne | 44,431 | — | 524 s @ 12 | in prep | ~2× OPM at matched cores (clean MPI-vs-MPI) |
| SPE-10 | 1.12M | — | beats OPM @ 12 | in prep | Faster than OPM at day-40 by ~2× |
"beats OPM @ 12" = faster than OPM Flow on 12 MPI ranks, same deck and host. GPU backend is in active development; per-benchmark GPU runtimes will populate this column as they are measured.
Methodology
BirSim is compared against OPM Flow, an open-source Eclipse-compatible simulator, run on the same input deck, the same host, and the same CPU core count. BirSim's wall time is measured directly; OPM's is taken from its own solver summary. Accuracy is validated on cumulative production (the mass-conservation invariants) plus instantaneous field rates and pressure.
Input deck files and exact command lines are available on request. As other reference simulators are added, accuracy will be reported against each.
Detailed reports
SPE-1: Three-Phase Blackoil
The Odeh (1981) comparative solution problem — gas injection into a 10×10×3 reservoir. The foundational validation case for any blackoil simulator.
SPE-9: Heterogeneous Blackoil
The Killough (1995) case — a highly heterogeneous permeability field, 24×25×15 grid, 25 producers and 1 injector, strong bubble-point nonlinearity.
Norne: Real Field, Corner-Point Grid
The Equinor Norne field — 44,431 active cells on a faulted corner-point grid, 36 wells, wet-gas PVT, WAG injection, ~9 years of production history.
SPE-10 Model 2: Million-Cell
The Christie–Blunt (2001) high-resolution model — 60×220×85 = 1.12M cells, extreme permeability contrast. A field-scale stress test.
More benchmarks & tiers coming
Field-scale GPU runtimes, SPE-11 (CO₂), and additional field models are in preparation. Request a specific case.