Mixed precision training [0] is a technique that mixes the use of full and half precision floating point numbers during training to reduce the memory bandwidth requirements and improve the computatio