Resurrecting the Circlotron and Other Mid-Centuryisms
Fair warning: this chapter’s gonna be highly technical. The engineers and technology-minded in the audience are probably going to love it. For everyone else, it may be a little hard-going. However, there’s a lot of useful information here that might make some of this “Class A, JFET, circlotron, etc” stuff a little more understandable, so you may want to have a look.
Let’s set the stage first. This is the beginning of 2012. We’re still in the garage, we’re still selling the same 4 basic products: Asgard, Valhalla, Lyr, and Bifrost. Mike has a mini-DAC prototype that I know I need to design a little amp for. But before that amp, I wanted to do a be-all, end-all balanced amp design that was substantially more ambitious than what we’d done before.
“Balanced.” “Substantially more ambitious.” Yep, that’s about as much of a design brief as I had, when I started designing Mjolnir.
In the real world, a design brief can be tens to hundreds of pages, spelling out everything from measurement and power output goals to detailed feature sets, form factor, cosmetics, producability requirements, and company best practices. But that’s a lot like telling an artist, “Sure, do anything you want. However, it has to be done in soapstone in a Bauhaus style, not more than 360 total square inches, with symmetry appropriate to production using not more than a 2-piece silicone mold. And it has to be orange.” And, in our opinion, it’s why you get a lot of stuff that looks a lot like the last product the company made, with nice fit and finish, but no real surprises. No stunning advancements.
And so, in the spirit of exploration that Mike started with the Modi, we set out to design an end-game-worthy balanced amp with really only five words in mind.
Whys, Wherefores, and Design Goals
“Balanced,” someone is probably saying. “Why did Mjolnir have to be balanced? I’ve heard that single-ended amps can be better than balanced, I was told balanced is a scam perpetrated by incompetent engineers, most headphones aren’t balanced, etc.”
Okay. Here’s the deal. Everything has advantages and disadvantages. Everything. The Ferrari you paid $400K for is going to be hard to get into and out of. The clutch packs on the robotic manual transmission will need adjusted every 8,000 miles. The brakes will cost $20,000. The parking lot attendants will hoon it around. People will make jokes about compensating. Blah blah, woof woof. Guess what? If it’s right for you, it’s right for you.
(And, you know what, same goes for Lexus, Tesla, Mercedes, or any other car. All have their advantages and disadvantages. There is no perfect solution.)
So what does this have to do with balanced versus single-ended? It gets down to tradeoffs. Single-ended has some advantages, and balanced has some advantages, too.
- Can have lower noise (only one active stage and ground)
- Simpler to design and implement (usually)
- Easier to connect (most headphones are single-ended)
- Ground path management can make or break the design (ground is not perfect 0V—where do the currents run?)
- High rail voltages required for high power (runs into device limitations)
- Balanced input is problematic (feedback connects to one side)
- 4x the power for the same rail voltages
- Rejection of common-mode noise on the input (if differential)
- Elimination of most ground path problems
- More parts, more complexity (except for maybe circlotrons)
- Gain difference between phases if 2-stage amp is fed balanced or SE input
- Most headphones ain’t balanced, duh
But, you know what? The main reason we went with a balanced design is that, in our experience, balanced designs offer better sonic performance than single-ended designs…as long as it is a purely balanced design.
Yes. A subjective reason.
Aside: Note the “a purely balanced design.” There are tons of amps out there with balanced input and output connectors that aren’t “pure balanced.” They take a balanced input signal and sum it back to single-ended, so they don’t need a 4-gang volume pot. Or they just hang single-ended outputs on a Neutrik 4-pin connector. In our opinion, these aren’t really balanced amplifiers, and performance of a truly balanced system shouldn’t be judged by their capabilities.
Escalating the Headphone Power Wars?
Funny. Some have accused us of firing the first affordable shot in the headphone power wars (Lyr) and escalating it with Mjolnir and Ragnarok. But there never really was a headphone power war at all, at least in our minds. I never had a power output goal for Mjolnir, other than, “About as much as the Lyr would be fine.”
“Well, the fact is, you do a bunch of high-powered amps, and most headphones don’t really need it,” someone says.
Yes. True. And that comes down to the First and Second Laws of Audio, as esposed by John Chen at Grado:
- You can never have too much power.
- See the first law.
Now, I’m being flippant here, because the law should read something more like, “More power is always a good thing, as long as there aren’t any tradeoffs to get that power.”
Yeah. More tradeoffs.
The tradeoffs for more power are usually:
- Higher noise (higher power = higher gain to reach that power output, so higher noise.)
- Greater need for protection (muting, DC detect, automatic shutdown, etc.—although a 100mW op-amp headphone amp can destroy a headphone if it lets go, something with 100,000uF of filter capacitance and 50V rails is gonna be very, very bad news)
- Paralleled output devices (really only in the speaker amp realm here—bottom line, paralleled devices, even matched ones, are never quite the same)
From the start, Mjolnir was going to be a high-power amp, so we were aware of the tradeoffs. #3 didn’t apply, but we paid a lot of attention to #1 and #2. Because, after all, we were in the middle of the New Orthodynamic Revolution, and many of those orthos weren’t that efficient. Also, many great headphones were still 300 and 600 ohm models, which need plenty of voltage to run them. Both mean big rails, and a big amp.
Today, orthodynamics are actually becoming more efficient, so the need for extreme power is abating. The headphone amp power war, which never really existed, will probably seem pretty silly in a few years time.
And yet still…I’ll take the high-powered amp.
Onto the Circlotron
So why was Mjolnir a circlotron from the start? After all, there are plenty of other ways to do a balanced amp.
Well, a big part of it is simply that I have a soft spot for circlotron, or “cross shunt push-pull” amplifier designs. They’re simple, high-performance, and neatly sidestep some of the problems inherent in other amplifier topologies (more on this later.)
Back as Sumo, we made circlotron-style amplifiers, but they were Jim Bongiorno’s designs. I’d never designed a circlotron amp. And yet they kept drawing me back in. First, because the topology is so different than anything else out there. When you first look at it, your natural reaction is “how the hell could that ever work?” Then, when you understand the principle behind it, you think, “wow, that’s really elegant. Why aren’t there more of these?”
And, another big part of the decision was based on the fact that there were no circlotron-style headphone amps on the market. Period. None. Zero. Nada.
“Well, that’s being contrarian,” someone says. Yes it is. But I’m a bit contrarian. I mean, hey, look at the name of the company.
But maybe I should explain a bit more about amplifier topologies, so you can better understand why Mjolnir was always a circlotron.
Some Solid-State Amplifier Topologies. Disclaimer: this is not intended to be an exhaustive summary, so yep, if I missed your favorite topology, sorry.
- JLH. A great example of an early transistor design by John Lindsay Hood. Underscores two facts about early transistors: (1) They were pricey, so it uses only 4 transistors, (2) The PNP versions sucked, so it used only NPN output in a quasicomplementary arrangement. This topology is much more like a tube amplifier than a modern solid-state amplifier—capacitor-coupled at the input and output, using a single voltage rail.
- Lin/Blameless. About 99% of all audio amplifiers today are Lin amps. A Lin topology uses a differential amplifier at the input, a second voltage gain stage, and an output stage that is usually complementary and biased with a Vbe multiplier. This describes virtually every speaker amp on the market. Of course, there are endless variations: complementary, quasicomplementary, symmetrical, buffered, linearized, output-inclusive-compensated, etc…but at the heart it’s a Lin. This topology offers a lot to like, including easy DC-coupling at the input and output, and a convenient terminal to run the negative feedback to.
- CFA/Current Feedback. This relatively new topology dispenses with the Lin’s differential amplifier, which is usually the limitation on a Lin design’s slew rate, and replaces it with a diamond buffer and current-feedback architecture. This topology can provide excellent performance at low parts count, but its advantages and disadvantages are hotly debated on, say, DiyAudio. Typically has better bandwith and slew rate, and worse distortion than a Lin design.
- Supersymmetry. This is a patented Nelson Pass topology that is inherently balanced, and has a fundamental simplicity that is very appealing.
- CSPP/Circlotron. Note that none of the topologies discussed above, except supersymmetry, is inherently balanced. Now we get to the Cross Shunt Push Pull amp, which is a very old topology (from the 1950s, google “Circlotron patent” for more info. First applied in tube amplifiers and named “Circlotron” by Electrovoice, the CSPP topology at first looks like a mistake. Oversimplifying, all of the topologies measured above use the output devices as, well, “valves” that control the flow of current from one or two voltage rails to a single output node. The CSPP uses these devices to “unbalance” the flow of current from two cross-coupled power supplies to two output nodes, one positive and one negative. Thus, it’s an inherently balanced design. It will never be anything other than a balanced design. And you will never get anything out of it except for balanced output (more on that later.)
- Chip/Integrated, or Chip and Buffer. Today, it’s easy to get a headphone amplifier output chip that pretty much does it all. That’s fine, but then you’re beholden to what’s on the chip—it becomes a “black box” topology, which can only be tweaked via power supply and ancillary components. Boring.
So, we went with a relatively old power amp topology because it was inherently balanced—and also very simple. It uses only N-channel devices (or only P-channel devices, if you swing that way), so there’s no worry about the N-channel and P-channel devices being mismatched. It does, however, require a complex power supply—two separate non-ground-referenced power supply rails for each channel. If you look at the CSPP transformer on the Mjolnir (the larger one), you’ll see it has about a billion output pins. That’s why.
And why did we call our version of the circlotron “Crossfet?” Because it’s not really a circlotron, if it’s not tubes. We wanted something that expressed “cross-shunt” and “MOSFET” in a short phrase. It wasn’t because the MOSFETs were mad.
Amplification Devices Disambiguated. And, with that, why don’t we talk about amplifying devices for a bit, because I’m sure that you guys wonder if we engineers just make up silly acronyms like JFET or MOSFET for fun:
- Triode. This is a tube with three elements (hence the “tri”): anode, grid, and cathode. A heater heats the cathode so that it emits electrons, which flow across naturally to the anode due to the overall circuit potential (voltage.) You can control the electron flow by applying a voltage to the grid. Triodes are very, very old (nineteen-teens), and, in some forms (such as the 6SN7, 6DJ8, 417, etc) are the most linear amplifying devices out there. Some tubes can do 0.005% distortion open-loop. So, while many tube amps have higher overall distortion than solid-state, it’s not due to the tubes’ inherent linearity—it has more to do with their current output capability (low) and the use of output transformers (necessitated by their low output current.) Tubes are also interesting in that there is no physical connection between the elements—the grid, anode, and cathode are all in hard vacuum, so a tube can be considered almost a perfect voltage-controlled device.
- Pentode. This is a tube with five elements: anode, control grid, screen grid, suppressor grid, and cathode. These types of tubes have much higher gain than triodes, but are inherently less linear. These were more commonly used as output tubes (driving a transformer) in audio power amps. Some can be run in triode mode for good results.
- Bipolar, or BJT, or just “transistor.” The Bipolar Junction Transistor was the earliest successful solid-state amplification device. Formed of three doped regions of silicon (negative-positive-negative, or positive-negative-positive (NPN and PNP, respectively), it uses current input (not voltage) to control an output current. The gain of a BJT is expressed in current gain, or beta. Betas of 20-500 are common. BJTs are used almost everywhere in discrete amplifier design, and if you avoid common problems (beta droop, nonlinearity, etc), they make great components. Can be made to withstand very high voltages. Just be sure that you obey their current-drive needs.
- JFET. Junction Field Effect Transistor. This takes a hunk of n-doped silicon and uses it to control the current flow through a channel of p-doped silicon. The input terminal is essentially a reverse-biased diode, through which virtually no current flows. However, current flows with 0V bias from the drain to the source, just like it does from anode to cathode in a tube. And JFETs act a lot like a tube—but, to be precise, it acts more like a pentode than a triode. These can be made very low-noise, and are frequently used as inputs on an amps differential stage.
- MOSFET. Metal Oxide Semiconductor Field Effect Transistor. In this case, the control terminal—the gate—is insulated from the drain and source, so it can also be treated as essentially a voltage-input device. With a catch: most MOSFETs have significant input capacitance, and care must be taken that you have enough current capability to drive them at high frequencies. Like a JFET, these devices operate kinda-sorta like a tube, with curves that look like pentodes. Commonly used as output devices in audio power amps, MOSFETS have a reputation for being noisy that takes them out of the front-end and VAS stages (usually.) MOSFETs come in two basic flavors: enhancement-mode (meaning they do jack squat if they don’t have a pretty big bias voltage on them) and depletion mode (which will run current even at no bias, like a JFET or a tube.) MOSFETS can be made very robust and very fast, making them a good choice for output stages (as long as you watch for parasitic oscillation—did I say “fast?”)
- SIT. Static Induction Transistor. A relatively new device that is has some characteristics of both JFETs and MOSFETs—with an interesting twist: the curves these devices produce look a lot like triode curves! Unfortunately, when available, SITs are eye-wateringly expensive, leading to limited use in commercial applications.
- Opamp. The Operational Amplifier is an amalgamation of hundreds or thousands of BJTs, JFETs, MOSFETs, capacitors, resistors, and other devices on a single chip. Available with gains in the tens of millions, these are complete amplifiers on a chip that offer very good performance in terms of traditional measurements. However, if you want to tailor a topology to your own application, or reduce loop gain to ensure constant feedback across the audio band, you’re out of luck.
Okay, so what does this mean for Mjolnir?
Well, let’s leave the tubes out of the equation. Mjolnir was never going to be a tube amp (though we did have a design for a big all-tube amp in a Mjolnir-sized chassis, which we never did anything with, but that’s another story. Mjolnir was going to be solid-state from the start.
So where does that take us? We have BJTs, JFETs, MOSFETs, SITs, and op-amps to play with. We don’t have anything against any of those devices. But for voltage amplification, we tend to like JFETs and BJTs, in that order, and we tend to like MOSFETS and BJTs for output devices, in that order.
Why the hate for BJTs? Well, it’s not really hate. Just caution. Current-driven devices are fine, but they need to have a little extra work to make sure you have enough current to drive them, even when they’re working hard and beta is drooping. And you have to watch their safe operating area and thermal characteristics a bit more.
What we ended up doing in the early Mjolnir design (and we’re talking breadboards here, not PC boards) was trying two different topologies:
- High-voltage JFET front end and MOSFET output with no overall feedback.
- JFET front end, BJT VAS stage, and MOSFET output, with local feedback around the VAS and output stage only.
We focused on these two topologies because both were simple, and both sidestepped the “different gain per phase” problem inherent in balanced amps that are driven single-ended.
What do I mean by this? I mean, if you drive a differential amp with overall feedback with a balanced signal, it produces a balanced output. 1V in, gain of 10 = 10V on either side.
But, if you drive a differential amp with overall feedback with a single-ended signal, it produces an unbalanced output: 1V in, gain of 10 = 10V on one phase, 11V on the other.
Yep. Look it up in an opamp cookbook. You’ll see the different gains per phase and ways to compensate for them.
However, since we wanted to have an amp with both balanced and single-ended input, we wanted to avoid having different gains per phase. That would mean we’d have to switch the feedback resistors (say, with relays) to compensate if a single-ended input was used. No, thanks. I didn’t really want to have 10 relays inside a Mjolnir. This was supposed to be a simple, no-frills, performance-is-everything kinda amp.
What was interesting about those two topologies was how closely they measured. We found that by using 95V rails and a special high-voltage JFET (which I think we own the world stock of), we could get very, very close to the measured performance of the amp with the VAS stage—without any feedback.
This made for a very simple amplifier. The path was set. Mjolnir would be a no-overall-feedback, single-stage amp design.
So Is It Class A?
One of the things we get asked about all the time is “What class amp is it?” It’s a terrible question—not because we hate to answer it, but because manufacturers have mis-applied amplifier classes, especially Class A, to the point where there’s a ton of confusion out there. I won’t repeat my screed about Class A amps a few chapters back, but I think it’ll be useful to go through some common amplifier classes.
Amplifier Classes Explained. While class is in session, why don’t we talk about amplifier classes a bit? This will be fun. Like everything else, this isn’t exhaustive—I won’t be talking Class C or S or T—look ‘em up!
- Class A. Class A amps run full-out all the time. The transistors all conduct, all the time. They never turn off. They’re hot. They’re big. They’re heavy. And they are, by definition, no more than 25% efficient. So if you have a 125W per channel Class A amplifier, it’s going to be sitting there dissipating 1000+ watts at idle. It will get cooler the harder you run it. There are no shortcuts, no excuses, no easy outs. If it’s not hot, big, and heavy, it’s not Class A. Period.
- Class B. This type of amp really isn’t used for audio. This is where the output transistors turn off as soon as they cross zero, because they are completely unbiased. The problem with this: huge crossover distortion, as the transistors turn on and off.
- Class AB. This is Class B, with bias on the output transistors so they run Class A some of the time. Fun fact: BJTs have an optimal bias for linearity, so "cranking up the bias" doesn't necessarily translate into better sound. MOSFETs don't. Crank them! When a Class AB's outputs eventually shut off, the transition is managed much better than in a Class B amp. This is the most popular audio output stage, because it combines high efficiency (up to 75% theoretically) with good performance. Mjolnir is technically a Class AB amp. It runs in Class A up to about 1W, then in Class AB thereafter.
- Class D. Switching or PWM amps. These “digital” amps have gone past the “exploding parakeet” stage (based on a comment from Mike Moffat about seeing one of the first commercial Class D amps, which output so much RF noise that it would probably cause a parakeet to catch on fire), and can provide good measured performance. They can also offer 90-95% efficiency, so your thousand-watt amp can fit in something the size of a cigarette pack. Still, barf.
- Class H. This is Class AB with voltage rail switching. These amps run at lower rail voltages to increase efficiency and reduce heat, then switch to higher rails when output demands. A neat way to get high efficiency without the drawbacks of Class D, but necessarily more complex.
So what does all this pedantic BS have to do with Mjolnir? Think of it as some of the stuff an audio engineer has to hold in his (or her*) head as they work on a new design. How crazy you wanna get? How many chances you want to take? Should it be an all-BJT, Lin topology, Class AB amp, because that’s the best-known and most-documented design option out there, or a whackazoid tube-input, level-shifted, DC-coupled hybrid supersymmetry circlotron with a Class H output stage?
Yeah. You get the picture.
*Too bad there aren’t more female engineers. When I was in school, one of my classmates snarked, “You’ve dated all the girls in engineering.” To which I replied, “Yeah. Both of them.” Which wasn’t far off the mark. Come on, guys can do this. It can’t be that hard. And they teach you wayyyyy more math than you need. Don’t be scared by nonlinear differential equations. You’ll never use them…well, unless you plan on being the Ph.D in residence and presenting papers before the AES. Which is fine. Me, I’d rather blow up…er, I mean build stuff.
Early Adventures with Mjolnir
“Circlotron?” Mike asked doubtfully. “Isn’t that something that only 6 people in the world know how to do, and even then they have to chant incantations and swing dead chickens over their heads to make them work?”
“No, they really aren’t that bad—“ I began.
“Famous last words,” Mike cut me off.
“They’re actually really simple—“
“Except for the keeping it balanced problem, the voodoo transformer, the eight thousand voltage rails, the weird in-the-air outputs, and making sure some idiot doesn’t ground the negative output and blow it up problem, you mean.”
“Well, yeah, but—“
“But you’re gonna do it anyway.”
“I already prototyped it. It works fine.” Which was true. Circlotrons are really dead-simple. They just scare people, because at first glance, they look like a very, very bad mistake that will catch on fire and burn your bench to the ground. In reality, a circlotron using enhancement-mode MOSFETs with no bias and no input will just sit there happily and do absolutely nothing. With a decent function generator, you can program in an offset voltage and two out-of-phase sine waves and run it easily. Really not a big deal. But Mike’s scared of weird analog things, and I’m scared of complicated digital stuff. So there you go.
“Really?” Mike said, doubtfully.
“Really. It took like an hour to build it up.”
Mike sighed. “Ohh-kayy. How are you keeping it balanced?”
“A differential servo.”
Mike’s eyebrows shot up. “A differential servo?”
I nodded and explained. I’d found a cool way to use a servo to compare the difference between the two outputs, and set the bias on one of the paired MOSFETs to match the other. That way, you just had to adjust bias on one side, and the other would follow.
Aside: “Servo” is short-speak for “DC servo.” This is commonly used in amplifier designs to ensure that the DC offset at the output is very low, without having to use coupling capacitors to block DC, or twiddling pots to null DC (and hope it doesn’t drift over time.) DC servos are relatively simple and very powerful, but like most things that are simple and powerful, they demand respect. DC servos are not perfect. They inject some audio back into the servo summing junction, so they’d better be high-quality and well-filtered for best performance (or used at a point that’s not an input, like I was doing in Mjolnir.)
“And the voodoo transformer and eight thousand voltage rails?” Mike asked.
“Already have one. Got prototypes from the transformer guys. And yeah, it’ll have a lot of capacitors in it, so what?”
Mike nodded. “Is this one of those things where the outputs are 40V in the air?”
“Nope, they’re close to 0V.”
“How about the shorting problem?”
I frowned. Because Mike had a point. Both of Mjolnir’s output phases were active. If someone connected them together (say, by using a balanced 4-pin to 3-pin TRS adapter), it would be a very bad day. And I knew that no matter what warnings we put in the owner’s manual about how Mjolnir was a balanced amp and a balanced amp only, and that they should never, ever, ever, not even on a bet try to use an adapter, someone would do just that. Probably about five seconds after the first one hit a customer’s door. Even if we put an electronic flashing sign on the top of the amp saying “NEVER USE ADAPTERS THAT SHORT THE OUTPUTS!” it would still happen. And there would be fireworks and consternation.
“I’ll figure that out,” I told Mike.
“And how about single-ended output?” Mike asked.
“This is an end-game amp. I figured it would just be balanced.”
“And what if these end-game guys have single-ended headphones?”
“Then they’ll have to get balanced cables, or re-terminate them for balanced,” I said. In retrospect, I should have paid more attention to this. Although we tried to work up a single-ended summer for Mjolnir, we were never happy with the performance of the design we had. So it never got single-ended output. And yes, I’ll admit…single-ended output is very useful. But it’ll always be a summed output, if you’re doing a circlotron. There’s simply no easy way to get a single-ended output from it.
And, you know what? Mjolnir really was a fairly simple amp to get working.** Except the protection. Mike called that one right on. A simple output delay wouldn’t protect Mjolnir from shorting adapters. A DC detect circuit wouldn’t do it, either.
In the end, I came up with a complex analog-computer-style circuit that continuously monitors output current and DC offset, and lifts the output if the current goes over a pre-set point, or if DC offset goes higher than a predetermined limit. I think it uses more parts than the original Asgard gain stage.
But it has saved our butts many a time. When we get an email that goes like, “Hey, I plugged in the Mjolnir and it sounded a bit funny, then it went “click,” in the middle of a song,” we immediately ask, “Are you using an adapter to plug your headphones into the jack?” The answer, unsurprisingly, is “yes.”
**Working, yes. Right, not so much. Mjolnir’s first appearance was at the Audeze-sponsored meet in Los Angeles. It was running far too high bias, its protection was only kinda-sorta working, it wasn’t thermally stable, and the servo was being, well, very un-servo-like (we found out later that it was oscillating.) But it ran through the show and didn’t blow anyone’s headphones up. Unfortunately, it sounded very fat and strange. It wasn’t until May that we had a real, production-intent prototype that made us happy to listen to it.
Whew. Big chapter. Lotsa tech talk. But hopefully illuminating.
Next up: Gungnir. More tech incoming. You’ve been warned.