CFD simulation, run for you.
Send us your OpenFOAM case. We run it on a GPU-native solver on EuroHPC infrastructure and return standard results you can open in your own tools.
Surface pressure on 147,449,586 cells, solved across 16 A100s. Spalart-Allmaras, rendered from the written OpenFOAM fields.
Where it runs
MeLuXina — EuroHPC, Luxembourg
- —Operated by LuxProvide under the EuroHPC Joint Undertaking. Data stays on EU infrastructure — relevant for customers with data-residency requirements.
- —Parameter sweeps run as independent parallel jobs, not a sequential queue. Turnaround scales with the study size.
- —No vendor lock-in: output is standard OpenFOAM format, readable in ParaView, Ensight, or any post-processor you already use.
brae — H100 benchmarks vs. reference
26–30×
faster than OpenFOAM's own GPU offloads (AMGX, PETSc) on an H100, at matched accuracy — under 1% on the fields
3.9×
faster than SPUMA GPU port
2.5×
faster than 24-core CPU node
261/261
solver tests passing
Why us
Measured speedups, at matched accuracy
26–30× faster than OpenFOAM's own GPU offloads (AMGX, PETSc) on an H100, at matched accuracy — under 1% on the fields. Also 3.9× vs SPUMA and 2.5× vs a 24-core CPU node.
261/261 solver tests passing
The full test suite for brae passes. That is the regression gate before any customer case is run.
OpenFOAM in, OpenFOAM out
brae reads existing OpenFOAM cases and writes standard OpenFOAM results. No mesh conversion and no post-processing translation.
EuroHPC compute in Luxembourg
Cases run on MeLuXina, an EuroHPC system operated by LuxProvide. Including the 147,449,586-cell DrivAerML mesh solved across 16 A100s.
Why it is fast
brae — a GPU-native OpenFOAM-compatible solver
Most GPU-accelerated solvers offload only the linear solve to the GPU while the rest of the SIMPLE loop runs on CPU. brae keeps the entire SIMPLE loop resident on the GPU — eliminating the round-trips that dominate runtime at scale.
It reads existing OpenFOAM cases and writes standard OpenFOAM results. No mesh conversion, no post-processing translation.
Hardware note: gains require HBM-class memory bandwidth (H100, A100). On hardware without HBM, speedups are lower. Results are not bit-identical to OpenFOAM — GPU execution reorders floating-point sums by small amounts.
What we run
External and internal aerodynamics
Steady and transient flows around bodies or through ducts. Lift, drag, and pressure distribution from cases you already have in OpenFOAM format.
Hydraulics — valves and actuators
Flow resistance vs. temperature, end-position damping, and pressure-drop characterisation across operating conditions.
Transient and moving-boundary problems
Time-accurate simulations with deforming or sliding meshes, returned as standard OpenFOAM time directories.
Coupled fluid-structure-thermal
Problems where flow, structural deformation, and heat transfer cannot be solved in isolation.
Migration and benchmarking
Bring an existing OpenFOAM case. We reproduce your baseline first, then run it on brae so you have a direct before/after comparison at matched accuracy.
Parameter sweeps on EuroHPC
Independent parallel jobs on MeLuXina rather than a sequential queue — turnaround scales with the study, not with how many cases are waiting.
How we work
Send your case
Send an existing OpenFOAM case directory, or a geometry and operating conditions. We confirm scope and quote before starting.
Baseline first
We reproduce the result you already trust before changing anything. If we can't match your baseline, we tell you before running the matrix.
Results in your format
We run the full parameter matrix and return standard OpenFOAM results. Open them in ParaView or any tool you already use.
Honest limits
- —Results are not bit-identical to OpenFOAM. GPU execution reorders floating-point operations, so field values will differ by small amounts from a CPU reference. We quantify this on the baseline case before proceeding.
- —We validate against a case you already trust before running anything else. If you don't have an existing result to compare, we work out what validation looks like before quoting.
- —There are problem types brae does not yet cover. If your case requires a solver or physical model we don't have, we say so rather than attempting a workaround.
Ready to run your case?
Tell us about your simulation and we'll get back to you within one business day.
Request a quote →Or email info@nubri.co