Is there any essential diferrence as far as training time is concerned between a CNN running on a FPGA and the same CNN accelerating on Google Coral Dev board? What I mean, is this: can I achieve faster training by using Google Coral Dev board instead of a FPGA?
I'm not sure about performance but ease of implementation between the Coral and an FPGA will be quite significant. The Coral is designed to work with existing frameworks like Tensorflow while the FPGA will have a lot of implementation overhead.