迈向无处不在的人工智能时代.pdf

1、Toward the Era of AI EverywhereLOKWON KIMCEODEEPX2 2023 DEEPX Co.,Ltd.Rise of Edge AIEverything with wheels has been demonstrated to move autonomously.All camera or sensor companies demonstrated object detection solutions.45 2023 DEEPX Co.,Ltd.Blocking Points forAI EverywhereMismatch Between AI and

2、Edge devices Processing AI requires greater data and computations than any other algorithm created by humans.AI algorithms are rapidly evolving and becoming more intelligent,adding new mathematical operators and requiring more data and computations.By nature,AI is ill-suited for edge device applicat

3、ions.Need a solution that goes beyond limits of theory!2023 DEEPX6#1:Conventional Solution 2023 DEEPX7#2:Edge Applications Require SOTA AI ModelsSome NPU solutions improved performance and power efficiency,but 8#3:A 1%Accuracy Loss Is Too Much!Using 8-bit integer instead of 32-bit floating point is

4、the key for power and area efficiency on resource-constrained edges.Normally,results in AI accuracy drop compared to GPUs.With GPUs,edge application developers create AI models for their own applications.The actual intelligence of the application is determined by edge NPUs.This incurs another loop o

5、f the AI accuracy validation for commercial applications which incurs longer development time and effort.A 1%AI accuracy drop is too much for easy deployment of edge AI.2023 DEEPX9#4:A Single One-Size-Fits-All Solution?There are various AI edge device applications.Each edge device AI application has

6、 specific function,performance,and cost requirements.A family of chips with different functions,price,and performance is necessary for different applications.#5:Developing Edge AI Software TakesToo Much Time and Effort Edge AI application developers demand an easy-to-use SDK.(In