Hi @dgarrido,
You’re correct that the 100% utilization shown in hailortcli monitor reflects the scheduler’s runtime occupancy - meaning the device is continuously busy processing your workload - but it doesn’t directly tell you how many of the 26 TOPS are actually being exercised by your model’s computations.
To get a clearer picture, you might try using the Hailo Profiler tool (available in the Dataflow Compiler / AI Software Suite), which reports GOPs per frame for your compiled model; multiplying that value by your measured FPS should give you an approximation of the actual TOPS being consumed during inference.
The profiler also breaks down internal resource utilization (MAC utilization, sub-cluster usage, memory bandwidth, etc.) which can help you understand how efficiently your YOLOv8s mapping uses the hardware.
Please see also here: OPs in Profiler - #2 by Nadav
Thanks,
Michael.