Open in tt-awesome →

tt-fluid

affiliated
by Moritz Thüning · C++ · GPL-3.0 ·
tt-fluid preview

OpenFOAM's steady incompressible solver (simpleFoam with k-omega SST) running entirely on Blackhole and Wormhole chips. Reads an ordinary OpenFOAM case, keeps every SIMPLE iteration resident on the Tensix grid with halo exchange between cores, and writes OpenFOAM time directories back. On the HPC motorbike benchmark a QuietBox 2 comes in at 2.34x lower cost per case than an 8x H100 server, with drag within the reference uncertainty.

LATEST v0.1.0 2026-10-19T04:03:24Z Release notes ↗
See all releases on GitHub ↗
# Changelog

All notable changes to this project are documented here. The format follows
[Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and the project uses [semantic versioning](https://semver.org/).

## [0.1.0] - 2026-10-19

First public release.

### Added
- `ttfluid` package and CLI: `ttfluid run` (device, after an emulator reference) and `ttfluid check` (CPU emulator only)
  for OpenFOAM simpleFoam cases with k-omega SST.
- numpy SimpleFoam (SIMPLE/SIMPLEC, k-omega SST) with a face-based and a device-layout (ELL) operator backend.
- Resident machine: tracing, RCB partition over chips and cores, scheduling, fusion, encoding, CPU emulator.
- Device runner and persistent kernels for Blackhole (p150, p300) and Wormhole, 1 to 32 chips, ttnn 0.68.0 (verified for this release on a p300, 1 and 2 chips).
- Large-case route: per-chip images, starting states from OpenFOAM time directories.
- OpenFOAM's pitzDaily tutorial, ready to run.
- CPU test suite against OpenFOAM v2412 references; CI with ruff, pytest and gitleaks.
- `bench/per_dollar.py` and the GPU baseline scripts (SPUMA build, run, bound() fix, Cd graders), with the raw data
  behind every published number in `results/`.
- Documentation: getting started, tutorial, architecture, performance methodology, accuracy, FAQ.

### Fixed before release (found by the clean-clone device check)
- `pip install -e '.[tenstorrent]'` now installs torch: the ttnn 0.68.0 wheel does not depend on it, and the device
  runner needs it.
- `ttfluid run --chips 2` on a p300 uses Ethernet links 0 and 1; the Wormhole default asked for link 3 and failed.
- pitzDaily did not converge on the device: the register allocator could place a neighbour gather's result on the
  tile it was still reading from, which the emulator cannot see. A gather's result now never shares a tile with a
  source, and lowering asserts it.
- A case whose `turbulenceProperties` selects a model other than kOmegaSST is refused; it used to run as SST.

[0.1.0]: https://github.com/moritztng/tt-fluid/releases/tag/v0.1.0
openfoam cfd rans simplefoam hpc turbulence cpp
wormhole blackhole quietbox