AI Architecture & Systems

CIM/PNM Compilation and Memory-Centric Offload

SourcesSurvey §3.2 and §6.2; d-Matrix layer mapping. CXL offload remains a curriculum extension.

3.18.1Identifying and Partitioning Supported Operators, and Host/Accelerator Cooperation#

3.18.2Weight Residency, Bank/Array Layout, and Data Reorganization#

3.18.3Operator Mapping, Tiling, and Cross-Layer Transfer Under Capacity Limits#

3.18.4Compute Coverage, Precision Constraints, and Fallback Paths#

3.18.5The Corsair Aviator Case and the Boundary of Undisclosed Raptor Compilation Details#

3.18.6Research Extension: CXL-Attached PNM and Heterogeneous Offload Compilation#