Pretty anecdotal use case, but when trying to do some type of neural architecture search for ML models that are ultimately going to be run on the edge, running

Pretty anecdotal use case, but when trying to do some type of neural architecture search for ML models that are ultimately going to be run on the edge, running some performance evaluation on a raspberry pi is a decent proxy. I have seen teams build their own type a raspberry clusters for this to be able to evaluate many models in parallel. It's basically a testing proxy for ARM architecture. But this thing only has 4 RPI per box so I don't quite understand which problem they are targeting either.