AI專題演講
AI國際鏈結辦公室於11/22(四)舉辦一場AI專題演講!
演講詳細資訊如下，歡迎大家踴躍參加!
題目: AI Chip Design Challenges at the Edge – from Deep Learning Model to Hardware.
講者: Bike Xie, Director of engineering, Kneron Inc
時間: 11/22(四)10:00-11:20
地點: 國立清華大學 台達館106室
報名及詳細資訊: https://reurl.cc/yEv58
演講摘要: Since the remarkable success of AlexNet on the 2012 ImageNet competition, deep learning models and especially CNN models have become the architecture of the choice for many computer vision tasks. However, inference of a CNN model can be highly computational expensive, especially for the end-user devices, such as the internet of thing (IoT) devices, which have a very limited computing capability with low-precision arithmetic operators. A typical CNN model might require billions of multiply-accumulate operations (MACs), load millions of weights, and draw several watts power for a single inference. Limited computing resources and storage become the major obstacle to run computation-hungry CNN on IoT devices.
Many design techniques in the area of model structure, compiler, and hardware architecture are making it possible to deploy CNN models on edge devices. This report discusses the design challenges for AI chip at the edge and briefly introduces these design techniques. A well-designed small size model might only require much less storage and computation resource. Therefore, model compression techniques including pruning, quantization, model distillation become substantial to deploy CNN models on edge devices. Compiling CNN models to hardware instructions is another critical step. operation fusion, partition, and ordering might significantly improve the memory efficiency and model inference speed. Finally, hardware architecture for AI chip is currently one of the hottest topics in circuit design. Dedicated AI accelerators provide an opportunity to optimize the data movement in order to minimize memory access and maximize MAC efficiency.
田羽彣專案經理/黃敬婷專案經理
AI創新研究中心專案-國際鏈結計畫
Email:yuwen@mx.nthu.edu.tw/ huangct@mx.nthu.edu.tw
國立清華大學 資電館 438 室