Mobo designer speaks out

[Groth]

I have seen two phase boards where the VRM controller was capable of three and four phase. The fun will be getting the MOSFET driver that matches the controller IC. And if you can't get MOSFETs and inductors that match the existing ones, you'll have to swap out the ones on the mobo.

Probably easier/cheaper just to buy a good board to begin with.

[jimmyswimmy]

Wow, I am such a narcissist, I googled myself and this was the first result. I'm honored to be quoted like that. And in case you're asking yourself, does this guy have anything better to do than google himself, it's either this or mow the lawn.

Anyway, I thought it would be fun to drop in. And to respond to a couple things I read here... if anybody is interested. This actually got a lot longer than I expected it to. Guess I wasn't as bored at this job as I thought I was.

1. I realized some of the context is gone, so I'll fill in. Up until a few months ago I designed power supplies for a "leading" semiconductor company. I made reference designs to help out manufacturers. Then I got bored (board, hah!) and changed careers.

2. Motherboard designers don't have time to do much of anything, actually. The designers I worked with wanted me to give THEM a finished design that they could just copy and paste onto their board. The most part of their designs involved cost reduction - literally, removing output capacitors until it blue screened and then putting the last one back (the one that broke the camel's back, so to speak). Crazy.

3. As far as the way designs go, a lot of that is due to poor specifications. It's really hard to specify ESL (equivalent series inductance) of a component, because it's hard to test reliably. So you don't really get that data from manufacturers. Which means your design is pretty much plug and play, the hard way.

4. Yeah, changing inductors is a cast-iron bitch. I did it EVERY day and was never happier when I could get my technician to do it for me. The only thing worse is changing thruhole caps -- at least the L's have big holes. It takes forever to clean out those little 30 mil cap holes.

5. Changing switching frequency is easy. Like someone here said, it's almost always just a smd resistor. Could be a cap but you don't see that much. Almost all the parts are Intersil (HIP6xxx) or Analog Devices, can't remember their part numbers. ON Semi makes some, but you don't see them much lately, same with Fairchild. Richtek is the dirty cheap solution, as I recall. But AD and ISL are the big players of late.

6. For the real adventurous types, you can change the droop/AVP/active blah blah blah, the load line/output resistance whatever you like to call it. That's only a two resistor change, typically. But it totally affects the whole dynamic load response, and you want to be pretty careful about that.

7. The real problem with changing the output caps is all because of this: A modern CPU supply is basically a bunch of parallel inductive current sources. When you go from pulling a 100 amps to almost none, all the energy stored in your inductors (0.5*LI^2 for those following along) must go somewhere. The cheezy way of thinking about that is that you have a bunch of caps (C) and a maximum allowable voltage rise (V) so that the 0.5*LI^2 = 0.5*CV^2... but in practice it's nowhere near that simple. Unfortunately. Or fortunately, because it kept me employed for quite awhile. All of those parasitic elements in the capacitor and in the board result in more than the ideal number of caps.

8. Magnetics design is a royal pain. Besides, the magnetics design interacts so strongly with the rest of the system that you're not done until the whole thing works. Say you pick an operating frequency and inductor design, you might need to reduce the L later to get the dynamic performance in spec. Or you might need to increase the L to get the heat down. And inductor core loss is still a near complete mystery to me. When I need to know it, I look it up on the "nomograph" (chart, but that's what they call it for some reason).

9. Power loss in the MOSFETs is easy to calculate. You can come up with a very good estimate of efficiency before even putting one board together. For the high-side MOSFETs power loss in the existing switch is proportional to switching frequency. For the low-side MOSFET the power loss is only related to frequency by the fact that higher frequency reduces the ripple current... and that gets pretty technical. Anyway when you pick MOSFETs you have to really worry about what the drivers can actually drive (the gate of a FET is basically a capacitor) and how fast it can do so.

10. Component temperatures interact with one another in interesting ways. You can model this with some really expensive software that my company wouldn't buy me (nor budget me the time to use it anyway). I spun many a board just to move some parts around or add thermal vias to reduce maximum temperatures. If you have two heat generating parts next to each other that are each 80 degrees and move them closer, they could be 100 degrees. And then you have to worry about the FR-4 board material "glassifying", which doesn't happen until over 100 degrees but that's not a number you want to get very close to.

10. KnightElite-- power electronics is a huge and rewarding ($) field. Learn it and, more importantly, get your HANDS ON some stuff. Do whatever you have to do to get in a lab. Almost all of electronics is not what the book tells you but rather it is figuring out what parasitic element you didn't count on. You can significantly reduce component stresses (say, peak voltage) through good layout techniques, which you will never learn in a book. Hell, most power product manufacturers neglect this stuff in their datasheets. And also, TI has some pretty good app notes describing buck and boost regulators. Starting from square one... they're basic but a pretty good primer. Oh yeah, don't forget to learn your controls stuff. Loop compensation is pretty essential too.

Well, that was a whole lot of fun. It was really nice to see that some people were interested in the work I used to do. I hope this is useful to someone and will be glad to answer questions if anyone is still interested. And now, since it is an option, I will round out this post with as many barfing smileys as I am allowed.



15 barfing smileys. too sweet.

p.s. Oh yeah, the multiphase thing. Most of the time when a system has multiple phases they are switching in an interleaved fashion, that is, only one phase is on at a time. In fact that's 99% of the time. I once saw a board with four separate power stages which all turned on and off together. Crazy, but dirt cheap. Anyway, if you add an extra parallel power stage, you could actually make things worse. One time I built a three phase board and the stages were not sharing correctly. I had a 90 A load (for this VRM board) and one phase was putting out 15 A, another phase was putting out -10 A, and the other phase was putting out 85 A. I ran it for 10 minutes to "warm up" heh heh heh. (ran out of smileys). A couple things from that. First, when I measured temperature, the inductor was over 250 C. Second, I burned the shit out of my hand. Third, I have never actually seen an inductor turn its paint a different color, have the paint actually fall off, and the inductor break into chunks. I was actually pretty proud of that. The pieces went in a bag and I stuck it on my wall of shame.

So the lesson is, even the so-called "experts" can have trouble with current sharing, and the road to hell is paved with good intentions. Your board probably works fine. If you want to make it better, add cooling (preferably passive, second choice airflow, third choice heat pipe and way last is liquid/gas cooling... too much mess if it breaks). Don't add extra output capacitors because you could screw up the control system and make the power supply go unstable (KnightElite -- take those controls classes too!).

Later all.

[jaydee]

If you want to make it better, add cooling (preferably passive, second choice airflow, third choice heat pipe and way last is liquid/gas cooling... too much mess if it breaks).

DOH! Vote for immediate ban! (joke)

Little off topic but being your here I would like to hear your opinion of where you think CPU's are going heat wise and if air cooling is going to be acceptable in the future.

Glad you stopped by!

P.S. Those are drooling smilies not barfing smilies. All good though.

[peepingdan]

This has been a great read for someone with no knowledge in this field. I work at a small networking equipment company and I see all sorts of this stuff every day but I've never thought about the difficulty involved. Enlightening.

[KnightElite]

Thanks for stopping by. I am in fact taking a control systems class right now as well. Yay transfer functions and feedback loops, etc.... .

Very interesting comments though, thanks .

[bigben2k]

Originally Posted by jimmyswimmy

2. Motherboard designers don't have time to do much of anything, actually. The designers I worked with wanted me to give THEM a finished design that they could just copy and paste onto their board. The most part of their designs involved cost reduction - literally, removing output capacitors until it blue screened and then putting the last one back (the one that broke the camel's back, so to speak). Crazy.

I've got a Q: if caps are removed, what could one expect from adding caps? (I know, it seems obvious).

[Brians256]

Hey! Welcome, jimmyswimmy! Since I work in a tangental industry, I get tantalizing glimpses of both sides of the coin. I play with these mobos at home, and I get to work with semiconductor companies and high-end EEs at work. I write software for DC and RF probing of components and on-wafer devices for a company called Cascade Microtech (we produce wafer probers and some RF software that helps calibrate VNAs).

Anyways, why is the ESL so difficult to measure? Is it because the devices vary so much that a representative sample is difficult to achieve? For example, a VNA can measure complex impedance of a device very easily although it is both complex and variant with frequency. The software I'm working on right now has a marker to display inductance (among other things) at a frequency of a smith chart and the ability to show inductance vs. frequency (or vs other things). With a decent modelling program, you should be able to deconstruct a lumped-element model of the device fairly easily and then publish it. No more than a couple days work per device (assuming you have a stable methodology), right?

Originally Posted by bigben2k

I've got a Q: if caps are removed, what could one expect from adding caps? (I know, it seems obvious).

I'm guessing that it depends very much on the characteristics of cap(s) you add. Unless you know the original design, you could add too much capacitance and throw the control system into instability. Or, adding caps with too much ESL could cause problems when the CPU changes current demand. Or it could be just fine. One of the things I'm learning about with the microwave design stuff I'm working on is that capacitance and inductance both affect the relationship of current and voltage of a signal (in opposite directions though). I think you could end up causing resonance because of newly added capacitance interacting with the control system's feedback loop. For those of you who actually know, please forgive my stumbling attempts here. I get paid for writing software, not EE.

My guess is that adding a bit of low ESR capacitance would help out in most implementations, but that adding bulk capacitance could be BadJuju(tm) in edge cases where current demands change rapidly.

As for having the tools to test a design... I probably come closer than most to having the tools. We have scopes and high-bandwidth VNAs here at work. But I don't know how to use the scopes, to be honest. The only reason I know how to use the VNAs is that I'm writing SW to use them and I blackmailed some EEs to teach me a bit about their theory (and I read "Practical Microwaves" by Thomas S. Laverghetta). How would you test a design modification? Would you model it after doing some component analysis or just throw a mod together and set a triggers on the scope to watch for anomalous events like voltage droop/spike?

[greenman100]

awesome comments indeed

so would there be any benefit to adding capacitors in parallel to those existing on say, an NF7-S?

[jimmyswimmy]

Originally Posted by jaydee116

I would like to hear your opinion of where you think CPU's are going heat wise and if air cooling is going to be acceptable in the future.

Well, this is all IMHO, but I haven't gotten the idea from the manufacturers I worked with that they wanted to spend more money on ANYTHING. So at least in the near future, next couple years, I would not imagine that you will see systems shipped with anything more than passive or air cooling equipment included.

If you look at the trends of electrical specifications for most processors over the past several years you will see that they are heading to where that will no longer work. Looking at my 6 month old Dell, which has a 2.8GHz P4, I see that there is some ductwork to vent the processor heat outside of the case, and that tells me that there is concern over the amount of heat generated by the CPU.

I can tell you for certain that motherboard manufacturers hate the fact that they have to spend so much on doing something as "boring" as providing power to the CPU, and the rising percentage of cost on the motherboard that entails. I'd be surprised if they weren't demanding of Intel that they do something about it. But Intel is basically fighting the laws of physics...

So I would say that for the next couple years (IMNSHO, I suppose) you're gonna see more inventive approaches to the same old airflow solution, but if there is not a more thorough change to the way processors are made than you might start to see some active cooling approaches. What that entails... I have no idea.

Hope that's helpful!

[jimmyswimmy]

Originally Posted by bigben2k

I've got a Q: if caps are removed, what could one expect from adding caps? (I know, it seems obvious).

Hi BigBen!

Well, it gets a bit technical. I sort of vacillate between telling people to add extra caps and telling them that it could be bad. The specs require that the system be stable for something like zero to 30,000 uF of caps, I think, which is pretty ridiculous since at zero it might be stable but it will also destroy your CPU. And there is a spec for ESR as well, which I can't remember. Should be available online. Anyway...

If you change the system too much you can, in theory, get it to where it will no longer be stable. There was one controller vendor that depended on the tiny amount of ripple on the output voltage, and so for that one, if you added too much filtering (more caps, essentially) there would be no ripple and it would lose its mind. For the other vendors' parts, depending on the control mode, you could have the same result.

What do I mean by stable? Well, providing the constant output voltage at the programmed load line, within specified limits. To use an idiotic and simple analogy, I am stable for most inputs; I can remain standing if you give me a little shove. But if you give me a huge push, I will go unstable -- first I'll fall over, then I'll get up and try to return the favor. In a similar way, if you push the controller outside its limits, you could receive an unpleasant surprise.

A well-designed control system will have enough extra margin that you could probably get away with a few little changes. So if you're not too worried about toasting that new motherboard and processor, go for it, see what happens, it will "probably" be all right (standard disclaimer: it might not be all right, too; do not taunt super happy funball; always wear your seat belt; product may contain nuts)

KnightElite is learning the basics of this unpleasant specialty right now. Perhaps he can offer a more intelligent analogy. Technically speaking, in a voltage-mode control topology, the ESR zero will alter your phase margin when you use a Type II (2 poles, 1 zero) or Type III (3 poles, 2 zeros) compensation. If you really lower ESR significantly, the ESR zero will move right, which could change the bandwidth of your power system.

Hope that's at least a little useful. Oh yeah, forgot my "drooly" smileys. I'm telling you, those guys are puking.

[jimmyswimmy]

Hi Brians,

Neat response. Now I have to think!

Okay, let's see.... first, off sounds like a great job.

Originally Posted by Brians256

Anyways, why is the ESL so difficult to measure?

Well, it's not impossible to measure by any means. My tech used to do it for me all the time. What used to kill me was that the measurement is so dependent on a good board layout, proper probe placement (say THAT five times fast) and the like. You are basically trying to find V = L * di/dt, where you program di/dt (the current slope) and measure the resulting V to determine L. Since we're talking an L in the nH range, you need either a huuuuuuge di/dt (which is not a great idea in production testing) or the capability to measure small V (back to my point of good layout and probes), not a preferred thing to do in production.

So I'm talking about soldering devices to boards to make this measurement, which means a production sampling program, and those pesky manufacturers never seem to want to do that. It was only a couple years ago that they started specifying a max ESR on datasheets, and that's the kind of information needed to do a proper design.

I used to have a nice HP impedance analyzer, it came in pretty handy a bunch of times. Still couldn't make reproduceable measurements of those nH level inductances though.

Originally Posted by Brians256

One of the things I'm learning about with the microwave design stuff I'm working on is that capacitance and inductance both affect the relationship of current and voltage of a signal (in opposite directions though).

ELI the ICEman, huh? I always liked that one. Voltage (called EMF by the old-timers, thus the 'E') leads current (I... no idea why it's an 'I', actually) in an inductive circuit, and vice-versa for a capacitive circuit. Good stuff. Pretty much a good description too.

Originally Posted by Brians256

How would you test a design modification? Would you model it after doing some component analysis or just throw a mod together and set a triggers on the scope to watch for anomalous events like voltage droop/spike?

So, right on both counts. To find the right design in the first place, you do a paper design. Usually more of an Excel design these days, since it's a bunch of mind-numbingly simple calculations that all depend on one another. So you pick everything and figure out where to start, and then build it.

And once it's built, you tweak it to your heart's content, because the paper design never seems to take into account all of the things you forgot or didn't know. :shrug: And then, time permitting, you go back and model it on some variant of SPICE and see how it handles boundary conditions.

Believe it or not, it's significantly faster to build, tweak and test a board than it is to model it. I find modeling most useful when I need to see either gross functionality of something untested, or edge conditions that I simply cannot obtain in real life (where did I put that 3781.22 uF capacitor, anyway???)

So, the guys who do this for a living have a neat little contraption that plugs in the same socket as the real processor. I had that 478 pin mother for a while, and now they've got some even more fancy stuff out with more pins and a pretty crazy hook up; don't know if those have hit the shelves yet. Anyway, all this thing does is turn on and turn off, real fast, and suck out as much power as possible. So that's what they use. It's a neat little computer-controlled device, and it gets damaged way too easily. I'd tell you that you can build your own, but it's pretty tough work, and the parts are expensive.

Hope that was useful. So hey, try to push those engineers to make the UI easy to use. I have a spectrum analyzer at my new job where the buttons are all the same size and grouped together, wait for it, ALPHABETICALLY. A nice idea, but annoying as hell. I can never remember how to use the damn thing, because I can't even rely on muscle memory to help out. I have to get out the book. It's annoying. Personally I love those Tek scopes with tons of different size and shape buttons and knobs... ah, a whole different discussion.

Hope that's interesting...!

[KnightElite]

I agree about the scope, Tektronix TDS3012A = best scope ever . Best I have used, anyway .

[Starman97]

I hear you on the weirdities of inductors.

Most Mobos use those nice wound core units, they're so forgiving,
they air cool nicely too.

For real fun, try some of the SMT inductors, the Panasonics are
really good, but there are several shell types, and each has different
loss and inductance variation over temp characteristics. But then, I'm
using a 28V Vin and trying to get to CPU core voltage.

The hottest parts in my designs are the inductors, the FETs couple
so well to the PCB with the Power-SO-8 packages that I dont worry
about them any more. Even a 5% duty on the high side isnt bad at 500KHz
amazing new parts out now.

Stabilizing the loop.. UGH
I've gone to current-mode parts from LTC, had such a bad time with the old style parts.

[Brians256]

jimmy, thanks for all your posts. It's definitely fascinating to see some of what really goes on and just how practical it might be for us (as nerdy consumers) to restore some of the safety margins to these designs.