Spatial Dataflow Kernels: PEs, Codelets, and Data-Triggered Tasks

SourcesTenstorrent, Graphcore, Cerebras, Groq and SambaNova layer mappings. Historical interfaces retain their generation scope.

2.18.1Tenstorrent: Boundaries Among TT-NN, TT-Metalium, and TT-LLK#

2.18.2Reader/Compute/Writer and Unpack/Math/Pack#

2.18.3PE-Local Buffers, NoC Data Movement, and Producer–Consumer Synchronization#

2.18.4Graphcore: Vertex/Codelet and BSP Compute–Exchange–Sync#

2.18.5Cerebras CSL: Wavelet, Color, Task, and Data Structure Descriptor#

2.18.6Cooperation Between Local Kernels and Whole-Graph Placement/Routing#

2.18.7Groq and SN40L-Era SambaFlow: Compiler-Owned Compute Paths#

2.18.8Distinguishing Public Kernel Interfaces, Compiler Internals, and Non-Visible Software Layers#