AI Architecture & Systems

Cross-Generation Comparison: Numerical Formats, Structured Sparsity, and Peak Conventions

SourcesSurvey §5.1, Fig. 10, Fig. 12 and Table 3. This chapter compares hardware support; quantization methods remain in Algorithms.

4.26.1IEEE, AI-Optimized, Block-Scaled, and Integer Formats#

4.26.2Separating Input Precision, Accumulation Precision, and Output Precision#

4.26.3Shared Scales and Representation Overhead in Microscaling#

4.26.42:4 and Variable M:N Structured Sparsity#

4.26.5Compressed Weights, Non-Zero Metadata, and Regular Datapaths#

4.26.6Supported Formats, Supported Operations, and Operator Mapping Conditions#

4.26.7Dense Peak, Sparse Peak, and End-to-End Effective Performance#