Evolution of Matrix-Engine Datapaths and Control Granularity

SourcesSurvey §5.2, Figs. 13–14; NVIDIA layer mapping.

1.15.1Volta: Warp-Cooperative Programming Interface and Sub-Warp Machine Execution#

1.15.2Turing: Warp-Level MMA, ldmatrix, and Operand Layout#

1.15.3Ampere: cp.async and Global-to-Shared Data Movement#

1.15.4Hopper: WGMMA, TMA, and Distributed Shared Memory#

1.15.5Blackwell: Tensor Memory and Paired-SM Tensor Execution#

1.15.6Independent Evolution of Operand Movement, Compute Issue, and Result Residency#

1.15.7Boundaries Among Hardware Capability, ISA Exposure, and Kernel Programming Abstractions#