Understanding clock domain crossing issues

Synchronous clock domain crossings

This section describes various types of synchronous clock domain crossings. Clocks which have a known phase and frequency relationship between them are known as synchronous clocks. These are essentially the clocks originating from the same clock-root. A clock crossing between such clocks is known as a synchronous clock domain crossing. It can be divided into several categories based on the phase and frequency relationship of the source and destination clocks as follows:

• Clocks with the same frequency and zero phase difference
• Clocks with the same frequency and constant phase difference
• Clocks with different frequency and variable phase difference
  • Integer multiple clocks
  • Rational multiple clocks

All the above sub categories may not be used in real designs but are being considered here for completeness and better understanding of the subject.

While describing all the above cases, it is assumed that the source clock (C1) and the destination clock (C2) have the same phase and frequency jitter and are balanced with the same specifications of clock latency and skew. It is also assumed that the clocks begin with a zero phase difference between them and the “clock to Q” delay of the flops is 0.

Clocks with the same frequency and zero phase difference

This refers to two identical clocks, as the clocks C1 and C2 have the same frequency and 0 phase difference. Note, that as the clocks C1 and C2 are identical and generated from the same root clock, the data transfer from C1 to C2 is essentially not a clock domain crossing. For all practical purposes, this is the case of a single clock design and is considered here for completeness.

Whenever data is transferred from clock C1 to C2, one complete clock cycle of C1 (or C2) is available for data capture as shown in Figure 9.

Figure 9.          Same frequency, same phase clocks

As long as the combinational logic delay between the source and destination flops is such that the setup and hold time of the circuit can be met, the data will be transferred correctly. The only requirement here is that the design should be STA (static timing analysis) clean. In that case, there will be no problem of metastability, data loss or data incoherency.

Clocks with the same frequency and constant phase difference

These are the clocks having the same time period but a constant phase difference. A typical example is the use of a clock and its inverted clock. Another example is a clock which is phase shifted from its parent clock, for example by T/4 where T is the time period of the clocks.

See Figure 10. Clocks C1 and C2 have the same frequency but are phase shifted and C1 is leading C2 by 3T/4 time units.

Figure 10.        Same frequency, phase shifted clocks

Whenever data is transferred from clock C1 to C2, there is more restriction on the combinational logic delay due to smaller setup/hold margins. If the logic delay at the crossing is such that the setup and hold time requirements can be met, data will be transferred properly and there will be no metastability. In all such cases, there is no need for a synchronizer. The only requirement here is that the design should be STA clean.

Clocks with different frequency and variable phase difference

These are clocks which have a different frequency and a variable phase difference. There can be two sub-categories here, one where the time period of one clock is an integer multiple of the other and a second where the time period of one clock is a non-integer (rational) multiple of the other. In both cases, the phase difference between the active edges of clocks is variable. These two cases are described in detail below.

Integer multiple clocks. In this case, the frequency of one clock is an integer multiple of the other and the phase difference between their active edges is variable. Here the minimum possible phase difference between the active edges of 2 clocks would always be equal to the time period of the fast clock.

For example, see Figure 11. Here clock C1 is 3 times faster than clock C2. Assuming T is the time period of clock C1, the time available for data capture by clock C2 could be T, 2T or 3T depending on which edge of clock C1 the data is launched. Hence, the worst case delay of any path should meet the setup time with respect to the edge with a phase difference of T. The worst case hold check would be made with respect to the edge with 0 phase difference.

Figure 11.        Integer multiple clocks

In all such cases, one complete cycle of the faster clock is always available for data capture, hence it should always be possible to meet the setup and hold requirements. As a result there will be no metastability or data incoherency and a synchronizer is not needed.

However, there can still be a problem of data loss in the case of fast to slow clock crossing. (That is, the source clock is faster than the destination clock.) In order to prevent this, the source data should be held constant for at least one cycle of the destination clock. This can be ensured by using some control circuit, for example, a simple finite state machine (FSM) would work in this case. In the example shown in Figure 11, if the source data is generated once in every 3 cycles of the source clock, there would be no data loss.

For the case of slow to fast crossings, there will anyways be no data loss.

Rational multiple clocks. In this case, the frequency of one clock is a rational or non-integer multiple of the other clock and the phase difference between the active clock edges is variable.

Unlike the situation where one clock is an integer multiple of the other, here the minimum phase difference between the two clocks can be very small-  small enough to cause metastability. Whether or not a metastability problem will occur depends on the value of the rational multiple, and the design technology. Three different cases are being considered here with the help of examples.

In the first case, there is a sufficient phase difference between the active edges of the source and destination clocks such that there will be no metastability.

In the second case, the active clock edges of the two clocks can come very close together, close enough to cause metastability problem. However, in this case the frequency multiple is such that, once the clock edges come close together, there would be sufficient margin in the next cycle to capture data properly without any setup or hold violation.

In the third case, the clock edges of the two clocks can be close enough for many consecutive cycles. This is similar to the behavior of asynchronous clocks except that here the clock-root for both the clocks is the same and hence the phase difference between the clocks can be calculated.

Note that in all the examples given below, some delay values are used and it is assumed that a phase margin of less than or equal to 1.5ns between the clock edges can cause metastability. This is just a placeholder value and in real designs, it would be a function of many things including technology used, flop characteristics, etc.

Example 1

This is the case when the active clock edges of both the clocks will never come very close together, and in all cases there would be a sufficient margin to meet the setup and hold requirements of the circuit.

Consider a clock C from which 2 clocks C1 and C2 are derived with a frequency of divide-by-3 and divide-by-2 respectively with respect to clock C. Here clock C1 is 1.5 times slower than clock C2. As shown in Figure 12, the time period of clock C1 is 15ns and of C2 is 10ns. The least possible phase difference between the two clock edges is 2.5ns which should be sufficient to meet setup and hold time requirements.

Figure 12.        Clock edges never come very close together

However, additional combinational logic should not be added at the crossing due to the very small setup/hold margins. If there is any logic, its delay should meet the setup and hold time requirements. If this condition can be met, there will be no metastability and no synchronizer would be required.

Further, if the crossing is a slow to fast crossing, there will be no data loss. However, in case of a fast to slow clock crossing, there can be data loss. In order to prevent this, the source data needs to be held constant for at least one cycle of the destination clock so that at least one active edge of the destination clock arrives between two consecutive transitions on the source data.

Example 2

In this case, the active clock edges of both the clocks can come very close together intermittently. In other words, the clock edges come close together once and then there would be sufficient margin between the edges for the next few cycles (to capture data properly) before they come close again. Here the word “close” implies close enough to cause metastability.

In Figure 13, clocks C1 and C2 have time periods 10ns and 7ns respectively. Notice, that the minimum phase difference between the two clocks is 0.5ns, which is very small. So, there are chances of metastability and a synchronizer would be required.

Due to metastability, the data may not be captured in the destination domain when the clock edges are very close together. However, in this case, note that once the clock edges come very close together, in the next cycle there is a sufficient margin so that the data can be captured properly by the destination clock. This is shown by signal B2 in Figure 13. While the expected output would be B1, the actual waveform could look like B2, but still there is no data loss in this case. However there can be an issue of data incoherency as described in Section 2.3, Data Incoherency.

Figure 13.        Clock edges come close together intermittently

For a fast to slow crossing, data loss can occur, and in order to prevent this, the source data should be held constant for a minimum of one destination clock cycle. Again, this can be done by the use of a simple FSM.

Example 3

This is the case when the phase difference between the clocks can be very small at times and can remain like that for several cycles. This is very similar to asynchronous clocks except that the variable phase differences will be known and will repeat periodically.

In Figure 14, clocks C1 and C2 have time periods 10ns and 9ns respectively. It can be seen that the active clock edges of both the clocks come very close together for 4 consecutive cycles. In the first two cycles there is a possibility of a setup violation (as the source clock is leading the destination clock) and in the next two cycles there is a possibility of hold violation (as the destination clock is leading the source clock).

Figure 14.        Clock edges are close for consecutive cycles

In this case, there will be an issue of metastability and hence synchronization needs to be done.

Apart from metastability there can be an issue of data loss also, even though it is a slow to fast clock domain crossing. As can be seen from Figure 14, B1 is the expected output if there would have been no metastability. However, the actual output can be B2. Here the data value ‘1’ is lost, because in the first cycle the value ‘1’ is not captured due to setup violation and in the second cycle the new value ‘0’ is incorrectly captured due to hold violation.

In order to prevent data loss, the data needs to be held constant for a minimum of two cycles of the destination clock. This is applicable for both fast to slow as well as slow to fast clock domain crossings. This can be done by controlling the source data generation using a simple FSM. However, the data incoherency issue can still be there.

In such cases, standard techniques like handshake and FIFO are more useful to control data transfer as they will also take care of the data incoherency issue.

Asynchronous Clock Domain Crossings

Clocks which do not have a known phase or frequency relationship between them are known as asynchronous clocks. Whenever there is a clock crossing between two asynchronous clocks, their active edges can arrive very close together leading to metastability. Here the phase difference between the clocks can be variable and unlike synchronous clocks it is unpredictable.

Proper synchronization needs to be done in the destination domain to prevent metastability. Apart from that, there can be problems of data loss and data incoherency (in both fast to slow as well as slow to fast clock crossings). If the source and destination clock frequencies are known, data loss can be prevented by holding the source data constant for two cycles of the destination clock. However, if the circuit is to be designed to be independent of clock frequencies, handshake or FIFO techniques should be used to prevent metastability, data loss and data incoherency.