Limiting Synchronous Clock Domain Crossing Paths - 2023.2 English

UltraFast Design Methodology Guide for FPGAs and SoCs (UG949)

Document ID
Release Date
2023.2 English

Timing paths between synchronous clocks driven by separate clock buffers exhibit higher skew, because the common node is located before the clock buffers. That is, the common node is farther from the leaf clock pins, resulting in higher pessimism in the timing analysis. The clock skew is even worse for timing paths between unbalanced clock trees due the delay difference between the source and destination clock paths. Although positive skew helps with meeting setup time, it hurts hold time closure, and vice versa.

In the following figure, three clocks have several intra and inter clock paths. The common node of the two clocks driven by the MMCM is located at the output of the MMCM (red markers). The common node of the paths between the MMCM input clock and MMCM output clocks is located on the net before the MMCM (blue marker). For the paths between the MMCM input clock and MMCM output clocks, the clock skew can be especially high depending on the clkin_buf BUFGCE location and the MMCM compensation mode.

Figure 1. Synchronous CDC Paths with Common Nodes on Input and Output of a MMCM

AMD recommends limiting the number of synchronous clock domain crossing paths even when clock skew is acceptable. Also, when skew is abnormally high and cannot be reduced, AMD recommends treating these paths as asynchronous by implementing asynchronous clock domain crossing circuitry and adding timing exceptions.