SoCs are becoming more complex these days. A lot of functionality is being added to chips and data is frequently transferred from one clock domain to another. Hence, clock domain crossing verification has become one of the major verification challenges in deep submicron designs.
A clock domain crossing occurs whenever data is transferred from a flop driven by one clock to a flop driven by another clock.
1. Clock domain crossing.
In Figure 1, signal A is launched by the C1 clock domain and needs to be captured properly by the C2 clock domain. Depending on the relationship between the two clocks, there could be different types of problems in transferring data from the source clock to the destination clock. Along with that, the solutions to those problems can also be different.
Traditional methods like simulation and static timing analysis alone are not sufficient to verify that the data is transferred consistently and reliably across clock domains. Hence, new verification methodologies are required, but before devising a new methodology it is important to understand the issues related to clock domain crossings properly. Different types of clock domain crossings are discussed here along with the possible issues encountered in each one of them and their solutions. A new verification methodology is then proposed which will ensure that data is transferred correctly across clock domains.
In all the subsequent sections, the signal names shown in Figure 1 are directly used. For example, C1 and C2 imply the source and destination clocks respectively. Similarly A and B are used as source and destination flop outputs respectively. Also, the source and destination flops are assumed to be positive edge triggered.
Clock Domain Crossing Issues
This section describes three main issues which can possibly occur whenever there is a clock domain crossing. The solutions for those issues are also described.
Problem. If the transition on signal A happens very close to the active edge of clock C2, it could lead to setup or hold violation at the destination flop "FB". As a result, the output signal B may oscillate for an indefinite amount of time. Thus the output is unstable and may or may not settle down to some stable value before the next clock edge of C2 arrives. This phenomenon is known as metastability and the flop "FB" is said to have entered a metastable state.
Metastability in turn can have the following consequences from a design perspective:
- If the unstable data is fed to several other places in the design, it may lead to a high current flow and even chip burnout in the worst case.
- Different fan-out cones may read different values of the signal, and may cause the design to enter into an unknown functional state, leading to functional issues in the design.
- The destination domain output may settle down to the new value or may return to the old value. However, the propagation delay could be high leading to timing issues.
For example, see Figure 2
. If the input signal A transitions very close to the posedge of clock C2, the output of the destination flop can be metastable. As a result it can be unstable and may finally settle to 1 or 0 as depicted by signals B1 and B2.
2. Metastability has consequences.
Solution. Metastability problems can be avoided by adding special structures known as synchronizers in the destination domain. The synchronizers allow sufficient time for the oscillations to settle down and ensure that a stable output is obtained in the destination domain. A commonly used synchronizer is a multi-flop synchronizer as shown in Figure 3.
3. Multi-flop synchronization.
This structure is mainly used for single and multi-bit control signals and single bit data signals in the design. Other types of synchronization schemes are required for multi-bit data signals such as MUX recirculation, handshake, and FIFO.
B. Data Loss
Problem. Whenever a new source data is generated, it may not be captured by the destination domain in the very first cycle of the destination clock because of metastability. As long as each transition on the source signal is captured in the destination domain, data is not lost. In order to ensure this, the source data should remain stable for some minimum time, so that the setup and hold time requirements are met with respect to at least one active edge of destination clock.
If the active clock edges of C1 and C2 arrive close together, the first clock edge of C2, which comes after the transition on source data A, is not able to capture it. The data finally gets captured by the second edge of clock C2 (Figure 4).
However, if there is sufficient time between the transition on data A and the active edge of clock C2, the data is captured in the destination domain in the first cycle of C2.
4. Effect of metastability on data capture.
Hence, there may not be a cycle by cycle correspondence between the source and destination domain data. Whatever the case, it is important that each transition on the source data should get captured in the destination domain.
For example: Assume that the source clock C1 is twice as fast as the destination clock C2 and there is no phase difference between the two clocks. Further assume that the input data sequence "A" generated on the positive edge of clock C1 is "00110011". The data B captured on the positive edge of clock C2 will be "0101". Here, since all the transitions on signal A are captured by B, the data is not lost. This is depicted in Figure 5.
5. No data is lost in this case.
However, if the input sequence is "00101111", then the output in the destination domain will be "0011". Here the third data value in the input sequence which is "1" is lost as shown in Figure 6.
6. Data is lost in this case.
Solution. In order to prevent data loss, the data should be held constant in the source domain long enough to be properly captured in the destination domain. In other words, after every transition on source data, at least one destination clock edge should arrive where there is no setup or hold violation so that the source data is captured properly in the destination domain. There are several techniques to ensure this.
For example, a finite state machine (FSM) can be used to generate source data at a rate, such that it is stable for at least 1 complete cycle of the destination clock. This can be generally useful for synchronous clocks when their frequencies are known. For asynchronous clock domain crossings, techniques like handshake and FIFO are more suitable.
C. Data Incoherency
Problem. As seen in the previous section whenever new data is generated in the source clock domain, it may take 1 or more destination clock cycles to capture it, depending on the arrival time of active clock edges. Consider a case where multiple signals are being transferred from one clock domain to another and each signal is synchronized separately using a multi-flop synchronizer. If all the signals are changing simultaneously and the source and destination clock edges arrive close together, some of the signals may get captured in the destination domain in the first clock cycle while some others may be captured in the second clock cycle by virtue of metastability. This may result in an invalid combination of values on the signals at the destination side. Data coherency is said to have been lost in such a case.
If these signals are together controlling some function of the design, then this invalid state may lead to functional errors.
For example: Assume that "00" and "11" are two valid values for a signal X[0:1] generated by clock C1. As shown in Figure 7, initially there is a transition from 1->0 on both the bits of X. Both the transitions get captured by clock C2 in the first cycle itself. Hence the signal Y[0:1] becomes "00".
7. Data coherency is lost in this case.
Next, there is a transition from 0->1 on both the bits of signal X. Here the rising edge of clock C2 comes close to the transition on signal X. While the transition on X is captured in the first clock cycle, the transition on X gets captured in second clock cycle of C2. This results in an intermediate value of "10" on Y[0:1] which is an invalid state. Data coherency is lost in this case.
Solution. In the above example, the problem results because all the bits are not changing to a new state in the same cycle of destination clock. If all the bits either retain their original value or change to the new value in the same cycle, then the design either remains in the original state or goes to a correct new state.
Now, if the circuit is designed in such a way that while changing the design from one state to another, only one bit change is required, then either that bit would change to a new value or would retain the original value. Since all the other bits have the same value in both the states, the complete bus will either change to the new value or retain the original value in this case.
This in turn implies that if the bus is Gray-encoded, the problem would get resolved and an invalid state would never be obtained.
However, this is applicable only for control busses as it may not be possible to Gray-encode the data busses. In such cases, other techniques like handshake, FIFO and MUX recirculation can be used to generate a common control logic to transfer data correctly.
The MUX recirculation technique is shown in Figure 8.
8. MUX recirculation technique.
Click here for a larger version
Here, a control signal EN, generated in the source domain is synchronized in the destination domain using a multi-flop synchronizer. The synchronized control signal EN_Sync drives the select pin of the muxes, thereby controlling the data transfer for all bits of the bus A. In this way, individual bits of the bus are not synchronized separately, and hence there is no data incoherency. However, it is important to ensure that when the control signal is active, the source domain data A[0:1] should be held constant.