Benchmark performance differs from Hailo Model Zoo results on Raspberry Pi 5 + Hailo-8L

General Discussion
,
1

Hi,

I tested several models using the Hailo Model Zoo on a Raspberry Pi 5 with a Hailo-8L module, but the benchmark results(FPS,batch size=1) are significantly different from the numbers shown in the official Model Zoo benchmarks.

I would like to ask what hardware and software environment was used for the official benchmark measurements.

Specifically, could you share details such as:

  • Host hardware specifications (CPU, RAM, PCIe , etc.)

  • OS version and kernel version

  • Whether preprocessing and postprocessing were included in the FPS measurements

  • Input source used during benchmarking (camera, video file, synthetic input, etc.)

I would appreciate any clarification on the benchmark environment or tips for reproducing the published performance numbers.

Thanks!

2

Hi @minjoo_kim,

The main reason you’re likely seeing lower FPS on a Raspberry Pi 5 + Hailo-8L is the PCIe bandwidth difference - the official Model Zoo benchmarks are measured on a system with PCIe Gen3 x2 (two lanes), while the RPi5 connects at PCIe Gen3 x1 (one lane), giving you roughly half the host-to-device bandwidth. The published FPS numbers are “hw_only” measurements from hailortcli benchmark , meaning they reflect pure device inference throughput with synthetic data (no camera, no pre/post-processing involved), so on a x1 link the NPU stays fully utilized but processes fewer frames per second simply because data arrives slower. You might try running hailortcli benchmark your_model.hef on your RPi5 and comparing the “hw_only” vs “streaming” FPS - if those two numbers are close, that confirms the PCIe pipe is the bottleneck rather than the device itself, and your Hailo-8L is performing as expected for a x1 configuration.

Thanks,
Michael.