4. ArchitectureAccelerator Architecture and AI Datacenters

Primary structure: Survey §2–§6. Core platform classifications follow the survey; additional cases are labeled as corpus extensions or contrasts.

4.1Scope, Terminology, and Evidence Standards for Architectural Comparison8 4.2AI Datacenter Context and Four Categories of Resource Pressure7 4.3Core Taxonomy: Four Classes of AI Accelerator Architecture7 4.4GPU: General-Purpose Parallel Execution and Specialized Matrix Computation6 4.5NPU: Heterogeneous Compute Engines and Shared Local Memory9 4.6Spatial Dataflow I: PE Arrays and Distributed Local Memory8 4.7Spatial Dataflow II: Reconfigurable Architectures6 4.8Spatial Dataflow III: Functional-Slice Streaming6 4.9Compute-in-Memory: SRAM, DRAM, and the Location of Computation8 4.10Compute Engines: Cross-Category Organization and Trade-offs7 4.11Memory Hierarchy: From Hardware Cache to Software Scratchpad7 4.12Programming Models: Understanding Hardware Constraints Through Software Layers8 4.13Packaging, Chiplets, and Logic–Memory Integration8 4.14Host, Remote Memory, and Memory-Centric Extensions7 4.15Physical Deployment Hierarchy and Scale-Up/Scale-Out Domains8 4.16Node-Scale Scale-Up Interconnect6 4.17Rack-Scale Scale-Up Interconnect6 4.18Pod-Scale Scale-Up Interconnect6 4.19Scale-Out, Network Datapaths, and Optical Interconnect7 4.20Topology Performance: Connectivity Is Not Effective Communication Performance7 4.21Collective Semantics: Endpoint Data Transformation7 4.22Collective Algorithms: Logical Communication Scheduling8 4.23Collective–Topology Mapping and Communication Offload7 4.24GPU Generational Evolution: Compute, Data Supply, and Cooperation Scope7 4.25Specialized Accelerator Generations: Workload Positioning and Hardware–Software Co-Design7 4.26Cross-Generation Comparison: Numerical Formats, Structured Sparsity, and Peak Conventions7 4.27Cross-Generation Comparison: Memory Hierarchy and Scale-Up Fabric8 4.28Power Delivery, Cooling, and Practically Operable Capacity9 4.29Future Design Challenges: Generality versus Specialization8 4.30Extensions and Contrast Cases: Photonics, Neuromorphic, and Limited-Disclosure Designs8 4.31Architectural Evaluation and Full-Stack Case Organization8