rdk-model-benchmark

Run the benchmark script with your quantized model path, for example benchmark.sh --model /userdata/model.bin. It wraps the official hrt_model_exec perf tool and returns average latency in milliseconds and FPS as structured JSON.

Sweep the thread count parameter within the official range of 1 to 8 and compare FPS across runs. On dual-BPU boards like RDK X3 use core_id 0 so any core can be used, then report the best thread count found.

Yes, use baseline mode which automatically picks the first preinstalled model from /app/model/basic or /opt/hobot/model/<soc>/basic. The output marks the model source as baseline, and conclusions apply only to that model.

No, model conversion and quantization happen on the development host with the hb_mapper toolchain, not on the board. This skill only benchmarks already-quantized .bin or .hbm models and points you to the official toolchain docs for conversion.

Results are affected by board temperature and background load; near thermal thresholds the board may throttle and distort measurements. Cool the board before rerunning, monitor with diagnostic tools, and note temperature conditions when reporting.

Check your RDK OS version or install the tool from the official toolchain release package. The script reports the missing tool explicitly and never substitutes another benchmarking tool in its place.