1Introduction 1.1 Background 1.2 Contents 2Key Attributes of AI Chips 2.1 Technology Overview 2.2 New computational paradigms 2.3 Training vs inference 2.4 Being able to handle big data 2.5 Precision for data representation 2.6 High configurability 2.7 Software toolchain 3Status Quo of AI Chips 3.1 Cloud AI computing 3.2 Edge AI computing 3.3 Collaboration between cloud and edge 4Technology Challenges of AI Chips 4.1 Von Neumann bottleneck4.2 Bottlenecks of CMOS process and device 5Architecture Design Trend of AI Chips 5.1 Cloud training and inference: big Storage, high performanceand scalability 5.2 Edge device: pushing efficiency to the extreme 5.3 Software-defined chips 6 Storage Technology of AI Chips 6.1 AI friendly memory 6.2 Commodity memory 6.3 On-Chip (Embedded) memory 6.4 Emerging memory CONTENTS Beijing Innovation Center for Future Chips (ICFC) 01 01 03 05 05 07 07 08 09 09 10 11 12 13 14 15 16 17 19 19 21 23 25 26 26 27 28 7 Emerging Computing Technologies 7.1 Near-Memory computing 7.2 In-memory computing 7.3 Artificial neural networks based on emerging non-volatilememory devices 7.4 Bio-inspired neural networks 7.5 Impact on circuit design 8 Neuromorphic Chip 8.1 Algorithm model of neuromorphic chip 8.2 Neuromorphic chip characteristics 8.2.1Scalable, highly parallel neural network interconnection 8.2.2Many-core architecture 8.2.3 Event-driven 8.2.4 Dataflow processing 8.3 Opportunities and challenges 9 Benchmarking with State-of-the-Art and Roadmap 10 Looking AheadReferences 29 29 30 31 32 34 35 36 37 37 38 39 39 40 41 45 47 。。。。。。