Phase- and Delay-Locked Loop Clock Control in Digital Systems

Synchronous sequential systems rely on globally synchronized clocks. With the increase in clock rates, low-skew clock distributions are becoming increasingly critical to achieving design speed objectives. High-speed circuits may also require clocks with programmable duty cycle and delay. For all these applications, we must employ a comprehensive clock management on a chip. The Phase- and Delay-Locked Loops are used to achieve low clock skew distributions. The principles of frequency synthesis, by which the clock rates can be multiplied and divided, are outlined, together with its applications. We then focus on the correct design of the PLL/DLL building blocks. These include the voltage-controlled oscillators (VCOs) that employ delay elements, phase and frequency detectors, and, the loop filters that make the control feasible. The study concludes with the outline of patented solutions to scalable, robust, and technology-independent solutions for future clock managers.

Figure 1:  Two closed-loop clock distributions. (a) PLL and (b) DLL

The active skew compensation can be achieved by using either PLLs or DLLs. Both compare the input and feedback clocks, and guarantee that they are aligned. The difference between the two is in the use of the internal delay line.

In DLLs, the delay line inserts the controlled delay between the input and output clock. In PLLs, the delay line is used as a ring oscillator that is realized by closing the loop and guaranteeing that the inverted output of the delay line is fed back. Hence, while DLLs only delay the incoming clock signals, the PLLs actually generate a new clock signal in such a way that the delay in the clock distribution is completely eliminated.

In addition to aligning clock signals, PLLs can synthesize frequencies. To achieve clock division, it is sufficient to place a divider in the incoming clock path; in other words, PLLs are not necessary to perform divisions. Multiplying clock rates, on the other hand, generally cannot be done without using the PLL closed-loop control. Since PLLs adjust the internally generated clock rate to that of the input clock, you can only achieve the stable-state solution if the generated clock is multiplied.

Delay Elements
You can realize the delay line used in PLLs or DLLs used for clock distribution purposes in two ways. Both realizations use CMOS inverters connected in series. The delay in each inverter is controlled by the input control voltage.

Current-Starved Inverter Lines vs. Capacitor Loaded Lines
The current-starved delay line consists of a series of inverters. The delay through each inverter is controlled by the amount of current that is allowed to pass. For that purpose, voltage-controlled resistors are inserted in inverters, between the NMOS and PMOS transistors. One such implementation realizes the resistors with a complementary transistor pair whose gate voltages are self-biased by complementary current mirrors. Based on the input control voltage applied to the NMOS transistor gates, the complementary PMOS gate voltages come from the gate voltage of the PMOS transistor in the input mirror.

The capacitor-loaded delay line adds capacitors to the output of each inverter. The capacitance discharge is then dependent on the control voltage. One way of achieving this goal employs a serially connected resistor, similar to the previous case.

The delay range in the VCO depends on the minimum and maximum delays in each element, and on the number of delay elements. The VCO control voltage controls the delay in each delay element. The exact relation depends on whether the delay line is current-starved or capacitor-loaded. In the former case, the transfer function is monotonically decreasing, while in the latter it increases monotonically. Semiconductor-process variations can slightly alter transfer characteristics—in general, there will be more spread from the capacitor-loaded elements, which implies that the current-starved implementations are better.

Another point worth noticing is that you can extend the clock-rate range of a VCO by changing the delay-line length—the number of elements by adding taps to each inverter. Practical realization of such VCOs is simple and requires the addition of a single multiplexer that chooses the delay line length.

Jitter Reduction Through Noise Reduction
To reduce the clock jitter induced by external noise, the delay line should possess good noise rejection properties. In a simple model, the noise is added to the power supply and the delay through an inverter is then expressed in terms of the power supply voltage. Using the same approach, you can express the sensitivity of the complete delay line, including the delay-control elements.

To protect the circuit from the delay variations that cause clock jitter, reduce noise through the power supply. Either passive or active methods can reduce the power supply noise. The passive methods employ filters on the power supply lines. The active methods achieve better protection by regulating the power supply. On-chip power regulators are usually of linear design. Linear regulators include a feedback control loop that keeps the voltage constant. In practice, such regulators significantly reduce the power supply noise, typically between 30 and 80 dB. However, standard operational-amplifier feedback-loop stabilization is critical to achieving correct linear operation.

In situations where you cannot effectively reduce the noise, or where additional tolerance is needed, use differential delay elements. The simplest such implementation consists of two parallel inverter chains. The elements in one chain carry the signals of the polarity opposite to the corresponding elements in the second. The delay control circuitry is replicated as well. To save the space, you can merge both inverter chains in a delay line that includes the delay control circuits as well.

Phase Comparators
Phase comparators are the second major elements of PLLs/DLLs. Their role is to detect the phase or frequency difference between the two compared clocks. The phase information is used to generate the VCO control signal.

Several implementations are possible. The simplest implementation consists of a single XOR gate that generates the Boolean difference between the two clocks. The difficulty with this implementation is in the exact interpretation of the comparison.

Sequential implementation consists of a single flip-flop. Here, one input (B) is used as a clock signal. The other input (A) is then latched into a flip-flop at every positive edge. If the edge of the first input arrives first, zero is latched. Otherwise the flip-flop latches the value of logic one. The output value then uniquely determines if the delay line should be sped up or slowed down. This value is regenerated at each positive clock edge of the first input.

Best suited are detectors that compare both the phase and the frequency of the two inputs. These comparators are referred to as the phase-frequency detectors. They consist of two flip-flops clocked by different clock inputs. The first flip-flop to be clocked produces the logic-one output, and resets the other flip-flop. When the clock edge of the second clock arrives, the second flip-flop shortly outputs the logic one, and both flip-flops are reset. The overall effect is that, between the positive edges of the two inputs, either the signal for speeding up or slowing down the VCO is active.

You can represent the detector as a three-state finite-state machine. The signal states that the VCO should speed up, slow down, or perform as before. The examples of the detector executions illustrate the operation of the phase-frequency detector. When the feedback signal lags the input clock, the VCO should speed up, otherwise it should slow down. In the locked state, when both edges coincide, both detector outputs become active for a short period of time. The durations of the two outputs are the same to achieve low jitter.

PLL Stability
Proper filters are used to stabilize the control of the PLL. This is the stability problem similar to the problems in any closed-loop control systems. In the case of PLL control, the state signals are the phase difference, the VCO control voltage and the output frequency. Exact analyses are possible by various simplified models. The transfer function includes the filter components and you get stability analysis using the usual closed-loop stabilization techniques. The required filter can be implemented as passive or active. The exact form of the filter depends on the implementation of each element of a PLL.

The most common form of a PLL uses a charge pump, consisting of a capacitor and the switches that either charge or discharge the capacitor. The charge pump PLL consists of a phase-frequency detector, a charge pump, and a delay line VCO. The two outputs of the phase frequency comparators control the charge accumulated on a capacitor, and consequence the VCO control voltage. Relatively simple passive filters are sufficient to achieve loop stability. Simple rules of thumb have been used in filter design since the early days of charge pump PLLs.

DLL implementations can achieve stability by simpler means, as no internal feedback loop exists. However, many difficulties can mar the practical implementations. Since DLLs employ the controlled delay line, their correct operation requires that the delay is exactly one clock period. It is possible to have situations in which zero periods, two periods, or any even-integer multiple of the periods are possible. In general, the recognition of such false locks is difficult, and relatively complex digital circuitry might resolve the false locking problem. Some applications of DLLs might require the turbo mode in which the lock is acquired quickly, sometimes in one clock cycle. A class of DLLs uses the bang-bang controllers that quickly and accurately achieve the locking state.

Systems where clock rates change during chip operation may require dynamic clock management. The recent introduction of Transmeta processors demonstrates the low-power benefits of such systems.

For a number of reasons, dynamic clock control will not work properly just with ordinary clock managers. First, clock division and multiplication requires the use of clock dividers—current dividers cannot do this operation. Second, the rest of the clock manager, including the VCO and loop filters might not work properly during the dynamic clock scaling.

Clock dividers, when used dynamically, introduce a series of problems. Under some conditions, the forbidden states of the divider are introduced, causing undesired cycle lengths, irregular clocks, and oscillations at clock rates that are too high for the rest of the sequential system. In addition, even if the divider will eventually work properly, there might be transient states during which the clock outputs are unsafe. Remember that under no circumstances can the clock signals exhibit glitches. Our patent-pending dynamic clock divider protects against all these effects. Similarly, our patent-pending VCO design is optimized for dynamic clock management applications, their scalability to high bandwidths, low voltage, and their easy use in programmable systems.