The lowest precision is using binary. Look for "binaryConnect" for research. Also the group behind pulpino did a chip implememtation of binaryConnect with great

The lowest precision is using binary. Look for "binaryConnect" for research. Also the group behind pulpino did a chip implememtation of binaryConnect with great results. Another possibility is the work of Jennifer hassler from Georgia tech on analog neural networks , including a roadmap of the possibilities. Those are possibly the most optimal theoretically , but it's not certain that they'll work.