Date&Time : 

2021 / 03 / 12 (Fri) 14:20 ~ 16:20

Location : 

Delta Building R216

Speaker : 

闕志達 教授 (Prof. Tzi-Dar Chiueh)

台灣大學電機工程學系 (Dept. of Electrical Engineering, NTU)

Topic : 

FloatSD Technology: Energy-Efficient Deep CNN Training and Inference

Abstract : 

The FloatSD technology exploits the (im)precision-tolerant characteristic of neural network training and adopts only a couple of non-zero digits in a neural network weight, reducing the convolution multiplication to addition of two shifted multiplicands (partial products). Furthermore, the mantissa field and the exponent field of neuron activations and gradients during training are also quantized. The 4-bit FloatSD technology was successfully applied to training three popular CNNs, namely MNIST, CIFAR-10, and ImageNet (Resnet18/Resnet50) and obtained accuracy results very close to or even better than CNNs trained by the 32-bit floating-point arithmetic.

The FloatSD technology has also been applied to build a PE-based CNN training hardware accelerator SOC in 28nm CMOS technology as well as Xilinx FPGA based training accelerator IP. The FloatSD technology can significantly reduces the power consumption needed for CNN training, while achieving great saving in memory capacity and IO bandwidth.

Bio :

Tzi-Dar Chiueh received the B.S. and Ph.D. in electrical engineering from National Taiwan University and California Institute of Technology in 1983 and 1989, respectively. He is now a Distinguished Professor in the Department of Electrical Engineering and Graduate Institute of Electronics Engineering at National Taiwan University. His research interests include algorithm, architecture, and integrated circuits for baseband communication systems and neural network IC design.

Dr. Chiueh was the recipient of the Outstanding Research Award from National Science Council, Taiwan in 2004-2007. In 2009, he received the Outstanding Industry Contribution Award from the Ministry of Economic Affairs, Taiwan. In 2017, he also received the Outstanding Technology Transfer Contribution Award from the Ministry of Science and Technology, Taiwan. Prof. Chiueh is an IEEE Fellow.