The Dream of Autotuning is a Reality
STMicroelectronics offers since many years, high quality DC-DC converters for high performance CPU power supply. A generation of fully digital devices is already available. They are based on a proprietary architecture called STVCOT, which is a Voltage Controlled Constant-On Time. This family of products has simplified a lot the design and manufacture of the total solution, because it has reduced the number of passive components on board, and because of the ability to program the parts via software, rather than by adding or changing some resistors or capacitors on board.
STMicroelectronics offers since many years, high quality DC-DC converters for high performance CPU power supply. A generation of fully digital devices is already available. They are based on a proprietary architecture called STVCOT, which is a Voltage Controlled Constant-On Time. This family of products has simplified a lot the design and manufacture of the total solution, because it has reduced the number of passive components on board, and because of the ability to program the parts via software, rather than by adding or changing some resistors or capacitors on board.
One challenge that was still left to the designer though, was the compensation network tuning. This is based on a PID filter. Power Designer are usually more familiar with the Analog Compensation Mathematics, so, STMicroelectronics approach, until last year, was to have a Graphic User interface that would accept a pseudo analog compensation, and do the calculations to find the equivalent PID filter values, that the designer could input into the software.
Even this difficulty has been solved in the next generation of high performance DC-DC power converters, and now the dream of Compensation Autotuning is finally a reality.
Figure 1. Block Diagram of STMicroelectronics STVCOT architecture, including the Autotuning Block.
The new devices, targeting Intel VR13 CPUs, are able to self-compensate themselves, thanks to an elegant software algorithm, developed by STMicroelectronics, in collaboration with the Italian University of Udine.
How does this feature work? There is a lot of Matrix and Phasors Mathematics involved, but in this document we will discuss conceptually.
The controller, during regulation, generates a sinusoid signal, Mp, that is fed to the output. The output and the feedback voltages, Mx and My, are read back, as it can be seen on figure 1. The frequency of the sinusoidal signal, will be equal to the desired bandwidth, chosen by the user.
The user will also suggest a desired reference phase margin φm.
So far, it looks like the device behaves like a Bode plot machine. This is partially true. A Bode plot machine would now calculate directly, the ratio of Mx to My, and their phase shift. This requires a lot of hardware and processing power, which could not fit inside a single microchip. The approach of the converter is from now on, indirect.
Figure 2. Details of the Autotuning Block generated signals, and read-back signals.
The device generates the signal Mp, and also two more signals, Vx and Vy, which are functions of the desired phase margin; refer to Figure 2. It can be proven mathematically that, when the overall transfer function has the user desired Bandwidth and Phase Margin, Mx and My are respectively orthogonal to Vx and Vy.
The core of the converter device, will calculate the projection of Mx on Vx, which we call ξx, and the projection of My on Vy, which we call ξy.
Mx and My are a function of the PID filter values, and so are ξx and ξy.
Figure 3. A visual representation of the converging algorithm to find the Optimal Compensation.
As it can be observed in Figure 3, the controller will vary the PID filter values, until ξx and ξy are equal to 0. This will move the compensation to the Target Point, that has the desired Bandwidth and Phase Margin. In Figure 3, the target point is where: the blue line equal to 1, because the target bandwidth divided by the measured bandwidth is equal to 1; and the orange line is -120, because in this particular case that is the difference between desired and measured Phase Margin.
The key point in this process is that convergence of the PID parameters is proven true by the mathematical approach.
Until now, Compensation Tuning was kind of an art, among Power Designers, because of the trial-error-repeat kind of approach of this task. Although the solution exists, it is human unfriendly to bear this process. However, now STMicroelectronics proposes a feature that will relieve the designer from this activity, so that their time can be spend in other aspects of the design, focusing on the overall fit of a design to the customer specifications. It is obvious that Autotuning speeds up the design phase, resulting in a faster time-to-market of an overall project.
It is to be noted that the Autotuning happens during regulation, it lasts few milliseconds, and it can be commanded at any time by software. That means that another advantage of Autotuning is that it can be performed, if desired, periodically on a system on the field, to re-tune the compensation, taking into account component aging, tolerance and temperature variations. In this way the product can function to the best of its performance, all through its life.
We would now like to review a practical example of where self-compensation helps the daily life of the Power Designer.
Figures 4 and 5 report data belonging to a step-down solution, for a High-End CPU. The Power Converter being tested, must be able to respect the Output Voltage specifications, under a defined transient, in this case a load varying in between 25A and 125A, with a slew rate of 625A/us. The controller has a droop programmed, so the Output voltage, the yellow line, should have a rectangular shape following the load, the blue line.
The top half of Figure 4, is what the solution might look like, at the beginning of the tuning phase. Usually the designer inputs an initial PID filter values, and starts testing, to see what parameters need to be changed and how much. The bottom half of figure 4, is what the solution should look like, at the end of the tuning process.
Figure 4. An example of a response to a realistic output current transient, before (top) and after (bottom) the Autotuning procedure.
Up to today, how much time it takes to go from the beginning of the tuning phase, to the end of the tuning phase, depends a lot. Certainly the experience of the Designer contributes to shorter times; but, given a variety of designs, some happen to be tuned quicker, and some other happen to take a long time, for no particular reason, other than the above mentioned concept that Compensation tuning is kind of an art.
The feature of Autotuning resolves this indetermination. In fact, in order to tune this particular rail with the new controller with Autotuning feature, it took the time of a click. The device is given the command of self-tune, and the algorithm find the optimal PID values in milliseconds.
In Figure 5 there are the Bode plots of the measures of Figure 4. Initially the device was configured with not enough Bandwidth (20kHz in the particular example). The Autotune feature will correct the PID filter values to the optimal point (50kHz in this case, making the Phase Margin even more robust).
Figure 5. The Bode plots of the system before (top) and after (bottom) the Autotuning procedure. These plots correspond to the system tested in Figure 4. These numbers are just an example for the purpouse of explaing a concept, and should not be taken as a general reference, as every design might have different specifications.
It is worth adding that every designer can take advantage of the Autotune Property as a tool that brings the solution to a general optimal working point, and then they can change the parameters values to add a specific character to their rail, like a better performance, with an added cost, or a lower but sufficient performance, at the advantage of space.
Conclusions
STMicroelectronics has introduced a family of DC-DC High-End power controller, with the feature of Compensation Autotuning. This feature allows a predictable faster (and easier) time to tune the design, with an obvious faster time to market; it allows a more robust design, as the algorithm automatically can take into account unknown parasitics and board tolerances; it allows a longer final product life, as the Autotuning can be performed upon command, on running systems, to compensate for components aging.
The Autotuning Technique for Digital Costant On-Time Controllers have been developed by: Saggini, Stefano; Loghi, Mirko; Zambetti, Osvaldo; Zafarana, Alessandro; Corradini, Luca.
