Darnell Digital Power Forum, Wednesday

Wednesday September 17, 2008 was the last day of this year’s Digital Power Forum. This session was even more technical than yesterday, which was better than the first day as far as hard facts and equations instead of people wondering when digital power will become mainstream. I mentioned yesterday that I did not attend any of the efficient data-center meetings and it turns out that may have been a mistake as Dave Anderson, a chief technologist at National Semiconductor told me there was a lot of good content at these meeting. Dave, being a power expert, must have felt he could skip the digital power chip-focused sessions in favor of these sessions that represent the market and ecological forces that are driving the chips. I also saw Dave at several of the device seminars that focused on silicon carbide and gallium nitride. The tone of today’s meetings was much more nitty-gritty. I could see why digital power advocates feel analog folks are dinosaurs, because to do compensation and understand the loop you leave the s-plane we learned to love from college onward and go into the z-domain, the operator for sampled-data systems. Your $30k network analyzer is junk, it won’t work with a non-linear loop.

On the other hand, there was also a hint that the digital-loop crowd was still trying to come up with any way to provide a definitive advantage over analog loops. And the realization that if digital power means using digital functions and digital communications between chips, well, we have been doing that for years and it will only get more popular.

The first presentation was by the great folks down at Microchip, the people that gave us PIC microprocessor. Bill Hutchings showed off a system similar to the TI one shown yesterday that did PFC (power factor correction) and then multi-phase buck dc-to-dc converter. It is obvious that digital power is suited to this type of application because PFC is such a slow loop, 120 Hz being the fundamental period, and the fact that you can combine both the ac-dc and the dc-dc control into one chip, helping to offset the higher chip costs of digital power. The Microchip part seems to be a DSP type controller rather than a state-machine, but they provide a GUI for used to help develop their applications. I wonder if Bill realizes the chill that goes down power engineers’ spines when they read a view foil titled “Buck Converter Software Overview”. This part explains a contributed article I reviewed earlier in the year. It was from Intusoft, the SPICE vendor. They were bragging about using SPICE to help develop code for DSPs. The mentioned using the Z-transform. I rejected the article because I just could not seem to get a straight answer as to what this was for, people trying make micro-code at a DSP company or a user trying to develop DSP code. The other thing that really bothered me was that they called it SPICE when SPICE is solving Kirchoff’s current laws with iterative matrix math and z-transforms have nothing to do with that. The PR person mentioned that Intusoft was working with Microchip and now it all makes sense. If you look at the hard-core presentations from Texas Instruments and Infinion, you see that compensation for digital power loops has to be done in the z-domain, where is sometimes called the delay operator and represents the fundamental “chunk” of a sampled data system. So what Intusoft is doing is making a math simulator to translate the real-world time domain needs of a power loop , such as PID control, into the z-domain. This is how you can select all the coefficients you need in order to compensate the digital loop. This demonstrates one of the resistances to digital loops—that we have to use an entirely different and non-intuitive method in order to compensate the power supply. When all the compensation and software is done, the Microchip folks ended up with a pretty nice isolated PFC supply. It took in 36 to 75 volts and put out 200 watts of 12 volts. The converter efficiency was 94% but only for the dc-dc section. The dc-dc section worked at 150kHz, which seems a bit slow to me. This is the headache faced by DSP solutions, since they have to run code to decide what to do every switch cycle, we don’t see any MHz or faster switchers. The benefit to slow switchers is that you have far less loss due to FET switching and gate capacitance. Microchip has always been very customer focused and they have online aids, as well as regional training centers and the aforementioned GUI interface to help develop the loop coefficients is sure to make the transition to digital power easer for the analog power engineer. As I said before, it is really not your concern how they close the loop inside the chip, as long as you feel comfortable with the compensation and development, you will have a general purpose solution to you power problems and the chip can even have a UART so you can talk to it.

The next presentation was form the Primarion group at Infineon. This was a great presentation because it was totally hard-core and without marketing fluff. Half the digital power marketers talk about the cost savings and the other half talk about how cost parity will be reached in 2012, so obviously it is worth tuning them all out and concentrating on the hardware, which is what we engineers like best. This was the presentation that showed the method to devise the compensation coefficients that the controller needed. The Primarion controller is a state-machine type so you are not doing any software, you are just figuring out the filter coefficients so you can define the chip programming at the vendor or your manufacturing floor. This was a wonderful presentation that showed how things in the PID domain (we are all used to) get transferred into the sampled-data domain that digital power parts operate in. The presenter, Philip Cooke was a down-to-earth technical type and a real straight shooter. He gained a lot of respect in my eyes when he told us that you have to build and try your controller before you go into production, something that every analog guy knows from bitter experience. Indeed, a quote from the presentation: “One of the issues of digital controllers that needs to be dealt with are the effects of quantization. Quantization occurs in both amplitude and time—these effects can produce PMW jitter and limit-cycle oscillations in applications.” Then it was pointed out that if you properly compensate the controller and use the procedure established in the paper, you won’t have any of these problems. Hurray, a real technical guy telling us to watch out for. This tells me you can always trust Primarion and Infinion to give the straight scoop on digital control. I am sure if you have an application that was better suited to analog control they would tell you, and you can bet they understand the benefits of digital control as well. A question after the class about when will Primarion have an auto-tune function struck me for a couple reasons. First, since it came from a Primarion customer, it seemed incredibly low-class to put Phil on the spot. Second, it did show that us analog dinosaurs are resistant to all the new math and techniques needed to compensate a digital loop. Finally it showed Phil to be a real class act, since he pointed out that he himself has recommended it and it is something they are working on. I would tell the questioner to go call up Intusoft and they will soon have a simulator that will do the z-transform math as easy as we do SPICE now.

The next presentation was from Anthony Kelly PhD. of Powervation. I guess this was put in to balance the nuts-and-blots presentations we just saw. I was pretty skeptical going in because this outfit had and ad in the conference proceedings and they had a chart that showed 92% efficiency at 0 load. Oh, perpetual motion. As usual, there were no facts from this startup, just how “unique algorithms” were going to provide efficiency, performance and ease-of-use. It looks from the ad that they are doing auto-tuning and diode emulation, techniques done by many of the other digital power companies. It is bigoted of me but there was also the fact that Anthony was a PhD, my experience being that the more education someone has, the less helpful he is in getting me to ship my product. The presentation still leaves me scratching my head. Kelly rightly noted that up to now, digital power has only been able to claim “same as analog” or “as good as analog” and he maintained that there where things about digital that were far better. One thing he mentioned is that unlike analog, digital control has knowledge of the duty cycle. Well, I guess it was 15 years ago Linear Technology was telling us to mix a small bit of the ramp signal into the feedback to prevent sub-harmonic oscillations in converters that have a greater than 50% duty cycle. The presentation then tossed in a little math to impress us, the requisite differential equation and Laplace representation to translate into the frequency domain, contrasted with a difference equation and the Z-representation of a sampled data system, then a very nice mention of the Euler Identity and a really nice chart of how phase and magnetize diverge from reality in a discrete sampled system, something that would seem to show digital control was in trouble by definition. Kelly then introduced the primary technical feature of his presentation, mojo, as in the Austin Powers movie, or for me, something that you can get working after partying all night listening to old blues records. Now that we were in the mojo domain instead of those staid old-fashioned s or z domains, Kelly mentioned that a digital controller and predict the output and that can give digital control a real benefit over analog. He then stated that digital can do dead-time optimization and then went into a demonstration of how digital control can do duty-cycle optimization. The next few slides showed a two-phase converter and asserted that due to part variations one phase can provide 15 amps while the other sinks 5 amps and you get 10 amps and the output. I later asked Chance Dunlop of National Semiconductor if he has ever in his life seen a poly-phase convert with such a mismatch and he pointed out “Not if you design it right.”  I asked the presenter how knowing the duty cycle can give any indication of the currents in the phases—he said you don’t know the currents but the digital control can “balance” them for best results. I will put all this in the snake oil bin until I see a little green circuit board with a chip and connectors on it. This demonstrates how marketing people do disservice to the digital power industry. If you set up straw men and then show that your part is better, that just discredits you to the people that know better, which is about 90% of your audience. I see a similar thing when people talk about the transient response improvement—they always are comparing that to an analog voltage-mode converter rather than a hysteretic converter that can start a switch cycle the nanosecond the output drops. Perhaps multi-phase controllers have made hysteretic control less useful, but digital power advocates have to compare their solutions to the best analog available, not the poorest, especially since they have a price disadvantage.

Next up was a delightful presentation from Nan Shi about a 250 MHz switcher that was completely on a silicon die. The capacitor took up most of the room, with the PFET being the next biggest. The control section was tiny. This is not practical, but a great way to rattle your noodle about what might be coming down the road in power technology. My only complaint was that like most college-age engineers, the report was far too focused on simulations. What I wanted to see was a schematic, a simulation and then extensive real-world testing showing what the die actually did on the bench. I have seen way too many disasters from engineers that maintained, “Oh I simulated it, it has to work right.”

Next up was Zilker Labs, with a real-world paper on diode emulation. Chris Young, the CTO did not have any mojo, just facts and figures and good solid engineering. The diode that people want to emulate is the catch or flyback diode in a buck converter. Synchronous converters replace that diode with a FET. The problem is that the FET can conduct in both directions so that means it is possible to get reverse flow in your inductor and wasted power. By emulating a diode you can turn off the FET in time to prevent power loss. Chris had a great slide that showed that zero current might not be the optimal place to switch the FET to save power. This is one of the benefits of digital control, it allows you to accommodate the ringing and other parasitics in the circuit and then switch the low-side FET at the perfect time to maximize efficiency. Linear Tech has a patent on switching off the FET like a diode would, and I don’t know if these digital techniques would infringe or if they are a clever way around it. Since Chris’s conclusion was that you don’t really switch off like a diode for the absolute best performance this seems to be a novel way independent of what LT has patented.

The last presentation was my favorite of the whole forum. It was form CamSemi, a startup spun out of Cambridge University. Brian Thomas, the business development manager (they used to call this salesman) gave a great presentation about the little flyback regulator CamSemi wants tio use to replace cheap linear supplies in 5 watt and smaller power modules. I have been briefed twice before on CamSemi and did a Voices interview with the CEO that will come out in a month. It is interesting to note that in all these presentations, they never mentioned the words “digital power”. They did not need a marketing buzzword to differentiate their product; they could do it with cost, size and performance. How refreshing. Now that is mojo we all can use. Brian pointed out that people need tighter and tighter regulation in these modules. If you ad an optocoupler that senses the output, that might at 10 cents to the BOM. That is way too much in these low-cost applications. What CamSemi does is add a control winding to the flyback, and then rather than go through a diode and feed back the rectified waveform, they send the ac thought a cap so the signal does not go negative relative to the chip common. Now they have no diode drop that changes with temperature and they understand what the transformer is doing during the entire switching cycle. Then by looking at the tangent to two slopes in this ac signal, the chip knows when to sample to get a representation of the output voltage. Brian claimed that the digital control also lets them ignore ringing at the beginning of the cycle, but us crusty old power junkies know that there have been blanking pins on current-mode converters since Linear Tech made them popular twenty years ago. The digital part of the chip also helps them infer output current since it is directly proportional to the average primary current.. They use the digital domain to do the multiplication and division needed to measure the average primary current. They can also compensate for a long cable you have in the wall wart the chip is used in. By using a resistor externally you can make sure the supply boosts the voltage a little as the current goes up to make up for the loss in your cable. OK, so this was a product pitch and those usually turn me off at conferences, Brian showed how the part meets FCC and CE and Energy Star requirements and pointed out it was in a SOT-23-6 package. The flyback uses a bipolar transistor rather than a FET since they are cheaper. Every since I worked a GM I have been more impressed by low-cost design. It is a lot harder to design a Geo than a Cadillac. When you have money you can just buy your solutions, when you are designing on the cheap you have to use clever design. CamSemi sure seems to have done this. Now, the sad truth is that you don’t buy something because it is clever, you buy it because it works. I am sure the fine folks down at Power Integrations also have low-cost flyback controller that pass EMI and Energy Star. It is just so refreshing to see a startup that is not promising things, but has a real part and a novel architecture and clever IC design. Oh, I didn’t even mention that CamSemi uses the part in a quasi-resonant discontinuous flyback mode. Great stuff, I wish these Brits well, even if King George did slap that tea tax on us 250 years ago.

So to go back to one of the panel questions from the first day—when will digital power become mainstream? Well, CamSemi shows that it is mainstream already. They used a nice mixed-signal chip to do clever things in the toughest low-cost market you can imagine. Those gallium nitride FETs from International Rectifier are not going to effect CamSemi one whit, since they will be too expensive to replace silicon FETs, much less bipolar transistors. There is no real need to make the loop faster, so the chip will be able to provide a solution for a decade. There is the same benefit in the PFC arena—it tends to be a slow loop with plenty of time for a DSP or state machine to do all the sampling and math needed to close the loop. It sure seems like integrating several controllers are a good idea, as Texas Instruments and Microchip demonstrated. As my friend at Summit Micro pointed out, a handheld product that runs off a battery can ill afford the quiescent current inherent in a DSP or even a state-machine based digital power chip. His quiescent current needs to be 100 uA or less. Another thing that seems to threaten digital power is faster switching speeds. Micrel has an 8 MHz switcher for hand-helds. Maxim just did a 4 MHz chip. It seems that everyone is going faster than a MHz. This makes DSP digital power seem awfully tough to do. A state machine can keep up, heck, we have 50 MHz A-D converters, but when you switch CMOS that fast, your power consumption goes way up. This will limit the applications for digital power. Using non-linear digital control loops means you can uses fewer output capacitors and still get transient response, but if you just double the switch frequency you can use a lot fewer output caps as well. As with everything, its an along design matrix with no definite answer, just tendencies that push towards one solution or another.

I had a professor at GMI, James McLaughlin, who in retrospect was one of the best professors I ever had. Among other things, he taught us that you should not change domains if you have to. So if you are trying to control a mechanical thing, don’t switch over to the electrical world, then go into optics then go back into the mechanical. This is why a flyball governor is beautiful, it controls the mechanical speed of a steam engine with a mechanical device. This changing domains is what the digital power people have to face. They are trying to control and analog thing—a power supply circuit. But they change the problem into a sampled-data mode one. Now, in defense of digital power, look what happed to pollution controls in cars. My 1982 Honda had a toaster-oven sized box under the hood full of relays and vacuum dashpots and solenoids and all kinds of mechanical hardware trying to keep the pollution down. But my 1992 Honda gave up on that and just turned all the variables into digital signals, did all the logic in a microprocessor rather than relays, and then used little motors and fuel injection profiles to control pollution. An engine with any kind of pollution control is less reliable since it has so may parts added. Electrical systems are the second least reliable things on a car after hydraulic systems. But if the digital pollution control means that you can use way fewer parts than if you tried to do it mechanically, well then the digital solution is more reliable that the mechanical one.

A similar thing happened in data acquisition. You can look at all these old app notes where analog guys designed all these cleaver circuits to linearize thermocouples. After a while any real-world engineer knew that the smart thing to do was rip all that analog off your board and just digitaze the thermocouple directly, sag offset non-linearity and all into a 2 dollar PIC and then have a look-up table to correct and linearize the output. You need an amplifier and an 8-pin u-C rather than 20 analog parts. It is not only cheaper and smaller, it is more reliable. This integration of functions is where digital power might shine. But the marketing types have to remember, there are tons of people that don’t want to talk to their power chip, they just want the thing to work. For those like that, analog power will be hard to beat.

PS: After the conference was a seminar of PMBus. I heard from a friend that Power-One intends to enforce its sweeping patent on communications with a chip to control the output. A Darnell employee told me this is holding back the entire power supply industry form adopting a standard. So to all those people that call me whacko for insisting that the patent system hinders innovation rather than helps it, I offer this real world example, to set next to your Beta-max and HD-DVD.