The effective throughput of a packet-switched path is reduced by two factors that compound at small packet sizes:
- Header overhead
- Each packet requires one 32-bit header word before the payload, reducing the effective throughput when the packet size N is small.
- Arbitration overhead
-
The stream switch incurs one additional cycle per packet arbitration. On paths that traverse multiple packet-switching stages, this can reach one cycle per packet per switch.
For small packet sizes N, the two overheads multiply. With N=1, the worst case is one 32-bit data word per packet: 1 header + 1 data + 1 arbitration cycle = 3 stream cycles per data word leads to effective throughput ≈ 33% of peak, not the 50% that header overhead alone would imply.
Recommended:
To minimize this impact, AMD recommends using the largest practical packet size for your application.