For the Business Analytics Day 2020, a professor of mine, Prof. Dr. Peer Küppers and I presented an object detection prototype application.
While I had previously created the application as a university project, we never answered the question as to whether the Coral USB Accelerator was actually worth its $75 price tag. So, I set out to create a comparative demonstration between the Coral device and an overclocked Raspberry Pi 4 GB, which retailed for about $55.
I then created a video processing pipeline, in which a camera image was acquired, fed through the Coral TPU, and then presented with annotations. I did the same using the Raspberry Pi, but running the neural network on its CPU.
Even with 4 actively cooled cores overclocked to 2.0 GHz the Raspberry did not stand a chance. The TPU crunched through the frames at 10.5 times the speed, easily working through 30 frames per second.
However, the demo also clearly showed that the Raspberry could play to its strength as a universal computing platform. Even though the speed was slower, the quality of the individual inference shots was much higher, thanks to its capability to run an uncompressed neural network. The coral device on the other hand can only run compressed models, through a process called quantization. This noticeably affects the precision and accuracy.
We presented the annotated real-time video streams at a booth, and people could come up to test the system themselves. Since a live demo is not without risks, I was happy that the application worked as planned.