AI is hot now because of deep learning superhuman result in both image and speech recognition. AI & Deep Learning will penetrate into every industry and drive electronics market growth. Deep Learning training requires high performance parallel computing to speed the process. GPU is currently the mostly adopted solution. Due to a lots of data movement required during training, memory bandwidth bottleneck is the biggest problem in Deep Learning, current solution could be co-package such as CoWoS, EMIB and 2.5D. Inference requires near zero latency & low power computation and require new chip architecture design. This is a hot area for IC design companies and start-ups.