Ask HN: Cheaply deploying CNNs on embedded devices

Is there a good way to cheaply deploy a CNN on a embedded device, such as a robot? Perhaps, are there cheap, arduino-like micro controllers with GPUs that are powerful enough to work for sophisticated models?

I guess the standard way is to make calls to some server that forward propagates the models, but that has its obvious pros and cons. I would like to have the benefits of on-device in real time.