The table below presents the performance benchmark results for the popular AI models and tasks such as object detection, face detection, and more on the raspberry pi 5 and Hailo-8L with batch size 8.
Network
Rpi5 Performance
Input resolution
yolov5s_personface
150.21
640x640x3
yolov6n
354.07
640x640x3
yolov8s
127.85
640x640x3
Yolov8s_pose
123.43
640x640x3
yolox_s_leaky
110.26
640x640x3
Yolov5n_seg
103.57
640x640x3
ssd_mobilenet_v2
145.42
300x300x3
efficientnet_edgetpu_l
97.48
300x300x3
efficientnet_edgetpu_m
242.92
240x240x3
efficientnet_edgetpu_s
373.96
224x224x3
Resnet_v1_50
257.56
224x224x3
clip_resnet_50
127.85
224x224x3
The results show the RPi5 and Hailo-8L offers excellent performance that should be more than sufficient for deploying these types of AI applications.
The frame rates are high enough to support real-time processing for most common use cases like security camers, smart home devices, robotics, etc. And these popular models are very efficient, so you still have headroom to run additional software on the Pi alongside them.
Please keep in mind that the performance results may vary depending on the software version being used.
The results presented here were generated using PCIe Gen 3 technology.
It’s worth noting that the difference in performance between the Hailo-8L with Raspberry Pi 5 and the Explorer model is due to the fact that the Raspberry Pi 5 and Hailo-8L are connected using a single-lane PCIe interface.
Thanks for publishing these results – I have built a couple modules using the Pimoroni Base Duo in order to pair the Hailo AI module with an NVMe drive, and one of the trade-offs regarding the NVMe drive is that the speed is supposedly halved due to sharing the PCIe bus with two devices. I am assuming that this means it could also reduce the Hailo module’s performance, but it would be interesting to run some tests to determine real world performance. Using your baseline numbers could be interesting. Do you have instructions for how to run this benchmark?
The results you see have been tested by running the following command:
hailortcli run {hef} --batch-size 8
If you want to explore more options for measurements, you can run the following command to see the available options:
hailortcli run -h
This will show you all the different options you can use when running the hailortcli run command, such as measuring latency and other performance metrics.
Please let me know if you have any other questions.
Hi Matt. It’s more basic than that. The Pimoroni Base Duo (as well as the Pineboard AI Bundle and probably other dual boards) has an ASM1182e PCIe switch onboard. This switch is limited to Gen2. So even though the Hailo8L M.2 and perhaps your SSD are capable of Gen3 you are going to be limited to Gen2 speeds. Don’t get something for nothing…
Hello @omria,
I am struggling to find the yolov5s_personface hef file for Hailo-8L.
The only one i found is only made for Hailo-8.
Where did you find it to perform this benchmark please?