Skip to main content

axrunmodel

Runs a compiled model directly against the AIPU with minimal overhead — useful for measuring peak hardware throughput without the full inference pipeline.

axrunmodel <path/to/model.json>

Point it at the model.json file in your compiled model directory (see Model Formats).

Example:

axrunmodel build/yolov8l-coco-onnx/yolov8l-coco-onnx/1/model.json

axrunmodel uses the same input data for every frame, which removes input bottlenecks and measures maximum AIPU throughput in ideal conditions.

Output​

After the run completes, axrunmodel reports:

MetricWhat it means
Device FPS1 / execution_time at the device level, including data transfer. Not a throughput number — closer to "max instantaneous rate".
Host FPS1 / execution_time at the host level, including PCIe transfer overhead
System FPStotal_frames / total_time — the meaningful throughput number

Options​

OptionDefaultDescription
-d N, --devices NallComma-separated device indices to use (e.g., -d 0,1). Run axdevice to list device indices.
--seconds N10Run for N seconds
--aipu-cores Nall (4)Number of AIPU cores to use. Reduce to test single-core or leave headroom for other models.
--throttle-fps NnoneCap System FPS at N frames per second
--double-buffer / --no-double-bufferenabledEnable/disable double-buffering optimization
--input-dmabuf / --no-input-dmabufenabledEnable/disable DMA input buffers
--output-dmabuf / --no-output-dmabufenabledEnable/disable DMA output buffers
--show-bar-chartoffDisplay FPS per frame as a horizontal bar chart over time — useful for spotting ramp-up or variability
--show-histogramoffDisplay FPS distribution as a histogram — shows how consistent performance is

Full option list:

axrunmodel --help

See also​