Throughput-oriented and Accuracy-aware DNN Training with BFloat16 on GPU | IEEE Conference Publication | IEEE Xplore