AXI4-Stream Support - 1.0 English

Versal Adaptive SoC Programmable Network on Chip and Integrated Memory Controller 1.0 LogiCORE IP Product Guide (PG313)

Document ID
Release Date
1.0 English

An NMU can be configured to support the AXI4-Stream protocol. Please refer to the AMBA AXI4-Stream Protocol Specification (ARM IHI 0051A) for rules and description of the protocol. In AXI4-Stream mode the NMU can act as a slave to carry one or more uni-directional stream connections with the following constraints:

  • Source data width can be 128-512 bits.
  • Source and destination data widths must match each other.
  • The TKEEP signal is supported and can be used to send partial transfers.
  • The TSTROBE signal is not supported.
  • The TUSER signal is not supported.
  • The TID signal is supported and can be used to disambiguate source streams. The TID field is 6 bits wide.
  • The TDEST field is 10 bits wide. TDEST can be used for local routing after egress from the NoC, or to select among multiple NSUs in M:N streaming mode.
  • Stream packets are not ECC protected in the NoC. If data protection is required it must be embedded in the data stream.
  • The stream source does not have complete knowledge of the data widths of all the slaves it communicates with. The packing rules assume the worst case (512 bits) to ensure correct operation under all conditions. If the NoC destination data width is >128 bits, the stream beats must be pre-arranged to be able to pack together. Stream packets should be sent such that the source data beats can be packed together and form an integer number of destination stream packet beats. If the number of beats with the same TID/TDEST is not 64-byte aligned, the last beat can have TLAST=1, but this might result in inefficient data packing.

The NoC will pack stream transactions together but has no functionality to reorder based on TID around width conversion points. This leads to a requirement that the sender must group transactions with the same TID to take into account width conversions. The transactions can either be grouped together as n beats or a group can be terminated with TLAST, leading to less efficient data packing.

The minimum size n that the sender must adhere to depends on the configured data width at the NMU and NSU and takes into account the 128-bit NoC transport in between.