Data Communication via Memory Tile - 2024.2 English

AI Engine-ML Kernel and Graph Programming Guide (UG1603)

Document ID
UG1603
Release Date
2024-11-28
Version
2024.2 English

The memory tile of the AI Engine-ML can also be used to communicate between AI Engine-ML tiles. This communication is based on the multi-port DMAs located in these memory tiles. Memory tiles are large memories that can be used to store intermediate results that would not fit in an AI Engine-ML data memory.

A common use case is to have all the data necessary for a complete iteration available in the memory tile. The kernel that will use these data will run multiple times in a single iteration, its input local data memory being filled-in multiple times (ping-pong mechanism) and flushing its output buffer multiple times on another memory tile.

Two different but equivalent ways for read/write access on the memory tile side are available:

  • Read/Write the dataset in multiple tiles, each tile being adapted to the memory module size.
  • Read/Write the entire data set, the associated memory module DMA will automatically switch in between ping and pong buffers.
Figure 1. Communication Using Memory Tile