FPGA AI Suite: Design Examples User Guide

ID 848957
Date 4/30/2025
Public
Document Table of Contents

22.3.1. [SOC] Streaming Enablement for FPGA AI Suite

In an M2M system, input buffers are provided by the host CPU. However, in a streaming system (S2M), input buffers are created by an external hardware stream. For the FPGA AI Suite IP to process this external stream, several operations must happen in a coordinated way:

  • The raw stream data must pass through a layout transform (either the FPGA AI Suite IP internal layout transform or layout transform in an external IP) to reformat the raw data into an FPGA AI Suite compliant data format
  • The formatted data must be written into system memory at specific locations, known only to the host application and the FPGA AI Suite software library at run time.
  • The FPGA AI Suite IP job queue must be primed at the correct time, in synchronization with the input stream buffers, such that the FPGA AI Suite IP starts an inference immediately upon a new input buffer becoming ready.

Within Platform Designer, a Nios® V based subsystem is added alongside the FPGA AI Suite IP to provide the streaming capabilities. This subsystem highlighted in blue in the block diagram that follows.

In the diagram, the yellow interconnect lines indicate Avalon® streaming interfaces, and the black interconnect lines indicate memory-mapped interfaces.

Figure 20.  Nios® V Streaming Subsystem