Simulation and runtimes

Compile, then execute

Vela compiles a LiteRT model for the configured Ethos-U target. A Vela-only experiment stops before simulation. An FVP experiment additionally builds a runtime and runs it on Arm’s Corstone Fixed Virtual Platform to collect counters.

The FVP follows the configured NPU target. Its role here is to simulate NPU execution and memory traffic. The host CPU runs the harness; its execution time is not an accurate Cortex-M55 timing model. Results are simulator estimates, not measurements from a connected Ambiq board.

Choose a runtime

Engine	Role
LiteRT	Model execution through the LiteRT runtime integration
heliaRT	Model execution through the HELIA runtime integration
heliaAOT	Execution using generated ahead-of-time code; requires the optional AOT dependencies

Support depends on the target, backend and memory placement combination. Validate your configuration and use the runtime combinations supplied with your evaluation. A compiler accepting a memory placement does not mean the FVP can execute it exactly; unsupported combinations are rejected.

Prepare the tools

uv run helx doctor
uv run helx setup

Setup manages the required dependencies and requests acceptance of Arm’s FVP licence. On Apple silicon, choose native macOS tools with only the FVP in a Linux container, or run the full toolchain in the supplied dev container. The FVP is not native on macOS in either path. See both installation options. On Windows use the supplied dev container for full runs.

For command options, see CLI reference.