AI/ML training traditionally has been performed using floating point data formats, primarily because that is what was available. But this usually isn’t a viable option for inference on the edge, where ...