Read and Write Bandwidth - 1.1 English - PG313

Versal Adaptive SoC Programmable Network on Chip and Integrated Memory Controller 1.1 LogiCORE IP Product Guide (PG313)

Document ID
PG313
Release Date
2025-12-17
Version
1.1 English

The traffic class chosen primarily impacts latency, while the bandwidth requirements ensure enough physical resources are allocated given the bandwidths specified. You can over-constrain the bandwidth requirements but you should never under-constrain them.

The user-provided read and write bandwidth numbers express the raw AXI-level transfer bandwidth. However, when designing a system, it is important to be aware of the overheads imposed by the NoC protocol. A NoC lane typically transfers 16,000 MB/s. However, in the downstream direction, overhead in the form of read and write request header flits are also part of the bandwidth (see Packetization Overhead), as are write response flits in the upstream direction. These can each consume an additional 6% to 100% of the requested bandwidth.

Transaction overhead in the NoC is dependent on burst length. To minimize this fixed overhead on bursts, choose a burst length such that:

Burst_Length x Data_Width_in_Bytes = multiple of Chop_Size

The maximum Chop_Size is 256 bytes, unless the data is targeted at an interleaved DDRMC with a smaller interleave granularity.

As an example, for a chop size of 256 bytes, and a burst size of 256 bytes, the NoC adds one write header flit per 16 data flits (a 6% overhead). By contrast, very small bursts of 16 bytes add one write header flit for each data flit (a 100% overhead).

Important: Because read requests are overhead on the downstream path, and write responses are overhead on the upstream path, the amount of read traffic impacts the maximum achievable write bandwidth, and vice versa. (See Table 1.)