1 Introduction1 1.1 Research Background and Significance1 1.1.1 Development Trends of Neural Network1 1.1.2 Requirements of NN Processor2 1.1.3 Energy-Efficient NN Processors4 1.2 Summary of the Research Work6 1.2.1 Overall Framework of the Research Work6 1.2.2 Main Contributions of This Book7 1.3 Overall Structure of This Book8 References9 2 Basics and Research Status of Neural Network Processors13 2.1 Basics of Neural Network Algorithms13 2.2 Basics of Neural Network Processors16 2.3 Research Status of Digital-Circuits-Based NN Processors18 2.3.1 Data Reuse18 2.3.2 Low-Bit Quantization20 2.3.3 NN Model Compression and Sparsity21 2.3.4 Summary of Digital-Circuits-Based NN Processors23 2.4 Research Status of CIM NN Processors23 2.4.1 CIM Principle24 2.4.2 CIM Devices25 2.4.3 CIM Circuits26 2.4.4 CIM Macro27 2.4.5 Summary of CIM NN Processors28 2.5 Summary of This Chapter28 References29 3 Energy-Efficient NN Processor by Optimizing Data Reuse for Specific Convolutional Kernels33 3.1 Introduction33 3.2 Previous Data Reuse Methods and the Constraints33 3.3 The KOP3 Processor Optimized for Specific Convolutional Kernels35 3.4 Processing Array Optimized for Specific Convolutional Kernels36 3.5 Local Memory Cyclic Access Architecture and Scheduling Strategy39 3.6 Module-Level Parallel Instruction Set and the Control Circuits40 3.7 Experimental Results41 3.8 Conclusion44 References45 4 Optimized Neural Network Processor Based on Frequency-Domain Compression Algorithm47 4.1 Introduction47 4.2 The Limitations of Irregular Sparse Optimization and CirCNN Frequency-Domain Compression Algorithm47 4.3 Frequency-Domain NN Processor STICKER-T50 4.4 Global-Parallel Bit-Serial FFT Circuits52 4.5 Frequency-Domain 2D Data-Reuse MAC Array55 4.6 Small-Area Low-Power Block-Wise TRAM59 4.7 Chip Measurement Results and Comparison62 4.8 Summary of This Chapter69 References69 5 Digital Circuits and CIM Integrated NN Processor71 5.1 Introduction71 5.2 The Advantage of CIM Over Pure Digital Circuits71 5.3 Design Challenges for System-Level CIM Chips74 5.4 Sparse CIM Processor STICKER-IM78 5.5 Structural Block-Wise Weight Sparsity and Dynamic Activation Sparsity79 5.6 Flexible Mapping and Scheduling and Intra/Inter-Macro Data Reuse81 5.7 Energy-Efficient CIM Macro with Dynamic ADC Power-Off85 5.8 Chip Measurement Results and Comparison88 5.9 Summary of This Chapter92 References93 6 A “Digital+CIM” Processor Supporting Large-Scale NN Models95 6.1 Introduction95 6.2 The Challenges of System-Level CIM Chips to Support Large-Scale NN Models95 6.3 “Digital+CIM” NN Processor STICKER-IM297 6.4 Set-Associate Block-Wise Sparse Zero-Skipping Circuits98 6.5 Ping-Pong CIM and Weight Update Architecture100 6.6 Ping-Pong CIM Macro with Dynamic ADC Precision103 6.7 Chip Measurement Results and Comparison104 6.8 Summary of This Chapter112 References112 7 Summary and Prospect115 7.1 Summary of This Book115 7.2 Prospect of This Book117
以下为对购买帮助不大的评价