Here we show that over 15 large scale deep neural networks can be trained in mixed IEEE float16 (multiplication) + IEEE float32 (addition) with no loss in accur

Here we show that over 15 large scale deep neural networks can be trained in mixed IEEE float16 (multiplication) + IEEE float32 (addition) with no loss in accuracy. We are happy to answer questions about this work.