Ethos-U

Convert a Vela-compiled model for the Ethos-U NPU on atomiq110. heliaAOT consumes the command stream Vela produces; it does not call Vela.

Field	Value
Model	a compatible Vela-compiled LiteRT model you supply
Target	atomiq110
Shows	The Ethos-U operator, the driver dependency, and the platform gate
CI	Manual. The input needs Arm Vela, and the Ethos-U85 path this platform targets has no automated coverage (#78).

The model is whatever VELA_MODEL points at. Neither Vela nor a Vela artifact is supplied by this example. This is a converter-stage recipe, not an end-to-end Vela or board bring-up tutorial. Record your Vela version, accelerator configuration, input/output hashes and matching driver/platform versions before using it. A tested complete U85 setup is not established here.

Prerequisites

Use a repository checkout with Bash and the environment from Running examples: uv sync --frozen --group ci, then activate .venv. Supply a compatible Vela-compiled model and run the commands below from this example directory. This step converts on the host; building and running the module also requires the matching Ethos-U driver and target environment.

Run it

VELA_MODEL=/path/to/vela_model.tflite ./run.sh

config.yaml selects atomiq110 as the target. Vela has already partitioned the model and produced the accelerator command stream; selecting a target in heliaAOT does not offload an ordinary graph to the NPU.

Successful conversion writes the configured module under out/; inspect its ETHOS_U source and build dependencies. Compilation and actual accelerator execution are separate checks. The supplied model still has to satisfy the converter’s supported tensor/shape/operator contracts.

What happens

Every Ethos-U-supported subgraph arrives as a single ETHOS_U operator whose first input is the command stream. The generated module wraps the driver’s invoke call and wires up the command stream and the base addresses for inputs, outputs and Vela’s auxiliary buffers. Operators Vela did not take stay ordinary kernels.

Buffer alignment is handled for you. Where Vela emitted its own offline memory allocation, heliaAOT honours those offsets rather than re-planning the region, so the addresses the command stream was compiled against stay valid.

The driver is a hard dependency

A module containing ETHOS_U operators will not compile without the driver header on the include path; the generated translation unit stops with an #error. This is deliberate. A missing header used to compile a stub that returned success without touching the accelerator, so an integrator could ship a passing build whose accelerator never ran.

A non-functional stub is still available for host-side codegen tests and for bring-up without a driver tree, and it has to be asked for explicitly with -DHELIA_ETHOSU_ALLOW_STUB=1. Never use it in a build expected to run the model.

The platform gate

Converting a Vela model against a platform that declares no accelerator is rejected at conversion time. Simulation is the legitimate exception, most often an Arm fixed virtual platform that provides the accelerator itself. Opt out per operator, in the configuration file only:

operators:
  - type: ETHOS_U
    attributes:
      require_platform_npu: "false"

--operators is a nested-model flag and rejects inline values on the command line.

Support status

End-to-end tests exercise Ethos-U55 on an Arm Corstone-300 fixed virtual platform, with the gate opted out as above. The Ethos-U85 path that atomiq110 targets is not covered by CI, which is why this example is manual.

Next

• Targets and silicon for the platform families and the full Ethos-U section.