Running Neural Networks

[NMO]

Hi!

I just ordered a StereoPi. Can't wait til it arrives.
I wanted to ask if the performance of the CM4 is enough for running a computer vision model. To be more precise, I have an EfficientNet, which is optimized for running on edge devices. So in theory, a GPU is not strictly needed. However, I need to perform realtime object segmentation on the camera streams so I am not sure if the CM4 is powerful enough.
If I should find out, that the performance is not enough, would it be possible to replace the CM4 with a stronger Raspberry module?

[Realizator]

Hey NMO,

I don't have experience with EfficientNet. The "raw" ARM power of CM4 is not fantastic, but it can run some simple models. To estimate the performance, you can run the test on Pi4 or Google to evaluate the performance of a model similar to yours (since CM4 and Pi4 have the same CPU).

As an option, you can use an external NN accelerator (like Google Coral).

CM5 is (officially) not supported by StereoPi V2. Technically speaking, it works, but if you force CM5 to work at its maximum performance, it'll exceed the power limit that StereoPi V2 can safely provide. There are also some software tricks to be used to use CM5. Briefly speaking CM4 is the most balanced solution for StereoPi V2.