2022, Chapter 15
The Search for Unity
If you believe the chatter on the intarwebs, Magni didn’t really measure that well until it’s 7th birthday—that is, the 3+ and Heresy generation.
Heck, even I believed it.
I mean, because of this belief, I worked really really really hard to get Magni 3+ to where it was, performance-wise—and I wasn’t unsuccessful. In terms of performance into low impedance loads, it absolutely smokes the Magni 3, or any other previous Magnis.
But, in the process of developing Magni+, in the middle of doing literally a baker’s dozen variations on the Magni formula, I discovered a lot of things.
One thing was how wrong I was about Magni’s performance through the generations.
Another thing is how one of the best-measuring Magnis went completely unrecognized.
And yet another is how I really like a particular amp topology for Magni—and why it’ll continue into the future, even if it misses a few dB on ultimate THD+N.
And it all began with me picking up an old Magni 1 and running it on the APx555 for laughs—just to see how bad it was. Because it had to be bad. Everyone said it was bad. How could everyone be wrong? Plus, it was designed in a garage. Hell, the one I picked up had been sitting on a shelf in my office for years, was covered with dust and had the serial number 000002 on it.
Except this is what happened:
Actually not terrible! Almost -100dB THD+N. Maybe not the be-all-end-all, but not bad!
And remember: Magni 1 had fixed x5 gain. As in, it only had
high gain. Most amps measure worse at high gain. Especially simple amps like Magni 1.
So I made a quick resistor change to take Magni 1 to x1 gain (low gain). And this happened:
That’s when my jaw hit the floor. Because -108dB THD+N is
better than the Magni 3+!
So…
all along, Magni has been a high-performance amp.
No. Wait. Stop. Read that again.
Magni has always been a high-performance amp.
Despite being conceived in a garage, despite being tweaked by eye, using an oscilloscope and shooting for maximum output swing, despite not having any kind of distortion analyzer except for a decrepit HP stack that looked like it came out of a 1950s science fiction movie, despite being years away from a Stanford Research analyzer and years more away from getting an APx!
That’s when I decided to re-measure all the Magnis, as part of the process of developing a new one.
This is what I learned.
Magnis Thru History
For those of you who want to skim the numbers and come to a quick conclusion, let me jump in front of the train.
I already went into why it’s not possible to compare an old Magni 1 with a new Magni+—due to gain differences. And that carries thru to all Magnis. I’m also simply measuring with a 300 ohm load for these comparisons, which has become a standard load for many headphone measurements. It’s not the more demanding 32 ohm load, which is usually what I’m looking at most carefully.
Magni gains thru the years are as follows:
- Magni 1: 5x
- Magni 2: 2x and 6x
- Magni 2 Uber: 2x and 6x
- Magni 3: 2x and 6x
- Magni 3+/Heresy: 1x and 5x
- Magni+/Heretic: 0.35x, 1x, and 5x
The higher the gain, the higher the noise and distortion. Why? More feedback equals less gain. And more feedback also means better measurements. More on that later.
But, more importantly, this means you can’t just do an apples-to-apples comparison of all those Magnis. If you do, choosing low gain, and running 2V RMS output into 300 ohms you get something like this:
- Magni 1 (5x): -97dB THD+N
- Magni 2 (2x): -108dB THD+N
- Magni 2 Uber (2x): -114dB THD+N
- Magni 3 (2x): -108dB THD+N
- Magni 3+ (1x): -105dB THD+N
- Magni Heresy (1x): -118dB THD+N
- Magni+ (1x): -108dB THD+N
- Magni Heretic (1x): -119dB THD+N
“Wait a sec,” you say. “A 2014 Magni 2 Uber gets damn close to the Heresy and Heretic versions—and it’s better than 3, 3+, and +! And at higher gain! What’s going on?”
Oh, you noticed that.
Biiiiiiiiig surprise. Because that’s from the times when Magnis weren’t supposed to measure well.
So what’s going on? Why did we apparently go backwards? What happens if gains are equalized? Why are the Heretical versions so much better in terms of measurement?
Let’s start by equalizing gains and focusing on the discrete Magnis:
- Magni 1 (1x): -108dB THD+N
- Magni 2 (1x): -108dB THD+N
- Magni 2 Uber (1x): -114dB THD+N
- Magni 3 (1x): -105dB THD+N
- Magni 3+ (1x): -105dB THD+N
- Magni+ (1x): -108dB THD+N
“Uh, shouldn’t these numbers all be better if the gain goes down?” you ask. “You said that more feedback usually gives you better numbers. And Magni 3 actually got worse!”
It’s simple: because it’s complicated.
And no, I am not messing with you. The problem is that a lot of engineering is really complicated, at least when it comes to the fine points, and has many points of contention, where you can go one way or the other, and some engineers will think one way is the best way and others will think the other way is the best way
and the hilarious thing is, they might both be right.
To wit: More feedback, to a point, reduces noise and distortion. But there are other mechanisms that can dominate. In the case of the 1- and 2-series Magnis, they’re about at the limit of their topology and noise floor. In the case of the 3-series Magnis, they’re at the mercy of their diamond input topology, which, in the interest of simplicity (cost), do not use current sources. When you vary the voltage across a transistor by over 1/3 of its rail, you’re gonna see nonlinearity. Welcome to one of the finer points of analog discrete design. +100 points to you if you know how to alleviate this phenomenon. Because it can be done—it just takes more parts. More on this later.
But you aren’t done. “Hey, you dropped out the Heretical versions. What gives?”
What gives is we’re going to concentrate on the discrete Magnis…the interesting Magnis…the ones that build on what Magni always has been—an affordable, discrete headphone amp.
“Wait, you don’t think Heresy and Heretic are interesting?” you ask.
No. Here’s a primer on how to do a great-measuring IC-based amp:
- Pick the best measuring ICs.
- Attach to a decent power supply.
- Update when better measuring parts are available.
And that’s it. Done.
I mean, sure, I tried to make the Hereticals as interesting as possible, with no overall loop feedback (which allows for an insanely low—but completely unnecessary—noise floor when in negative gain), but the bottom line is that their performance is set by the quality of the TI op-amps used in them. And, even though there’s no overall loop feedback, there’s tons of feedback in the front end and buffer stages of the design—like 120-130dB worth.
Does this mean the Hereticals are bad? Not at all. It means,
for me, from my unique perspective as a discrete analog design engineer, they are not very interesting. I am not you, and I am not most engineers. So it may be a truly singular opinion.
Okay, so we’ve now established this chapter is about the discrete Magnis. We’ve gone back through history and saw that there are some, ah,
surprises. And I did mention those 15 different dev versions of Magni I went through—those 15 different versions that were considered
and rejected.
Why are there surprises? And why do a ton of variations and throw them away?
Read on…
A Simple Goal—Or So I Thought
Here’s the deal: two Magnis ain’t ideal. I’d rather have just one. There’s less potential confusion, less variation we have to deal with internally, less stuff to buy—it’s all-around better.
So I figured:
let’s do a discrete Magni with Heresy-like performance. A Magni Unity, so to speak.
Sounds simple, doesn’t it?
Well, the basic idea—doing as good as, or better, than op-amps with a discrete design—is relatively straightforward. Cascaded discrete op-amp designs with hundreds of parts will get you there, no problem.
But when you start adding on constraints like “simple,” “small,” and “at a budget commensurate with a $100-ish product,” oh boy…then things get
tough.
Because, let’s face it, I wanted the new Magni to stay a Magni: Small size. Small price. Great sound.
Aside: I mean, we are already at quixo levels of insanity just doing the Magni as a discrete product. Discrete, small size, and small price don’t naturally go together. But that’s what Magni always has been, and what I’d like it to always be.
In terms of increasing measured performance, I knew I had three choices:
- Make the gain stage more linear. The problem with this approach is that it usually requires more parts. Look at an instrumentation front end compared to a standard diff, and you’ll see what I mean. Even going to an output triple is more parts, and only really comes into play at low impedance, and also has less potential output swing. All of this has to be factored in, when you’re looking at retaining the rated power output of an amp.
- Increase the loop gain, so we could use more feedback. The problem with this approach is that every time I’ve done this, I’ve come away less happy with the sound. There are ways to manage this, of course—local loops, linearization, etc—but in general, the higher the gain the less happy I am. Of course, I may be completely insane or fooling myself, so YMMV.
- Go back to a voltage feedback topology, which measures better. This is something well-known in the amp development community. Go to Diyaudio and do a search on “CFA vs VFA.” I hope you have a ton of time and a big cup of coffee. Oversimplifying: CFA (current feedback, as with Magni 3/3+/+) is usually faster, VFA (voltage feedback, as with Magni 1/2/2Uber) measures better.
“Aha!” some of the techies are saying. “That’s why the earlier Magnis measure better than the later ones! They’re voltage feedback!”
Exactly.
And, after I saw the numbers the Magni 2 Uber put up, I got really excited.
All I’d need to do would be to go back to voltage feedback, and tweak it a bit, and I’d be there! I mean, I had to be able to better optimize a topology in the era of the APx, right? I mean, I was still doing the Magni 2 Uber in our first dusty office, with the old HP analyzer. I had to be able to do better today!
So, one of the first versions of the new Magni was voltage feedback. But it wasn’t just a Magni 2 Uber. It used matched parts, new JFETs, and had multiple options for me to tweak loop gain, topology, and other parameters. By using paired parts, it came in at just a bit more complicated than the current Magni 3+.
I figured,
Hey, this is probably it, we just take this, tweak it, and we have a single Magni! This is gonna be easy.
Except, just like on the chef shows where they have a clip of one of the contestants saying, “I’ve done a million risottos, this can’t fail,” right before a faceplant…this Magni faceplanted.
Not thrown-off-the-show level faceplant, but faceplant in that, no matter what I did, I couldn’t get it to measure significantly better than the old Magni 2 Uber.
That was frustrating! I mean, new matched/paired parts, way more linear, wayyyy better JFETs, wayyyyyyyyyyy more analyzer…and I was essentially back at the Magni 2 Uber. About -115dB THD+N. And that was with the VAS loaded, unloaded, MIC or CDOM compensation, different front end degeneration, operating point tweaks, etc, etc.
Frustrated, I did more variants. In all, there were three different voltage-feedback Magni candidates:
- The originalist. What I just talked about.
- Full complementary. Higher complexity, more parts. Barely fit on a Magni board. Gave us a couple dB better performance. Would have to be a more expensive product due to complexity.
- Wacky VAS. Cascoding and CFP didn’t improve performance significantly, despite the fact that increasing the linearity of the voltage gain stage should have the biggest impact. Also had problems with fitting and cost.
And while doing all these variants, I started remembering what I
didn’t like about voltage feedback topologies. I mean, VFA (voltage feedback amplifier) aficionados will be the first to point out that:
- VFAs measure better than CFAs. As in, THD+N numbers will usually be better. This is mainly due to more loop gain.
- VFAs have greater PSRR. As in, a VFA will typically reject more of the crud on the power supply. So you can get a lower noise floor with a VFA. It’s probably important to note here that much of the improvement in THD+N, when you’re playing at -110dB and lower, can be related to noise floor, so this isn’t an insignificant advantage.
- The transient advantages of CFAs shouldn’t matter. As in, yeah, they might slew at 500V/uS, but you don’t need that for music, so who cares. (But then again, if you’re talking -108dB vs -115dB THD+N, you shouldn’t be able to hear that either, especially on transducers that are -40 to -50dB.)
The first two, yeah, I totally saw that in the measurement and prototyping phase. The third (speed), welllllllllllll…until you’ve compensated a high-loop-gain VFA design, after becoming very very spoiled with CFAs, you haven’t lived. At least if you define “living” as “a descent into a hell of complex and painful decisions, with possible tradeoffs.”
I mean, when the first version oscillated on first power-up, I wasn’t surprised. I’ve done plenty of VFA stuff back in the Sumo days. I know where to tweak, and I can usually get something stable pretty quickly.
Aside: if you ever look at a VFA amp schematic and wonder about all the small-value capacitors strewn around the design, they are usually for compensation—getting the amp stable, so the open loop gain is less than 1 by the time the output phase hits 180 degrees. For all expected loads.
Sounds fun? Well, in a current feedback amp, pick one cap from VAS to summing junction, filter the input, and you’re done. (Usually.)
In a VFA amp, you’re probably looking at least at a Miller cap and one in the feedback. Probably more, though. It wouldn’t be surprising to have a cap and stopper on the front end and caps and stoppers on a degenerated VAS. Getting this right can be pretty painful. And it can result in really weird stuff like asymmetrical slewing, or ringing on one or both sides of a square wave input.
So yeah, compensating the new VFAs was fun. Or maybe “interesting.” And even when it was stable, I had to reacquaint myself with terms like “slew rate,” which I’d never really had to worry about on the current feedback side. The final VFAs never were as fast as the CFA designs, but they were stable, didn’t ring, and had symmetrical slewing.
There were other pains, too, including the use of JFETs in the front end. I’m not ant-JFET, I use them all the time. But they do have their limitations, in that they are:
- Low gain. JFETs don’t have as much gain as BJTs, so a front end, even with degeneration, is less linear.
- Expensive. JFETs cost a lot more to a lot lot more than BJTs.
- Not particularly well matched. In some applications (like Nexus), JFETs have to be matched. In Magni, some were so poorly matched that even in an application where it shouldn’t matter…it mattered.
So why did we use them? Simple. If you aren’t doing a full complementary topology, BJTs can’t be used in opposition, where one sinks the current coming out of the other. Without this inherent cancellation, BJTs will cause current to flow through the volume pot, unless you use a coupling capacitor on the input. Current flowing thru the volume pot manifests as scratching or wooshing noises. Capacitors in the signal path we don’t like. So JFETs it is on non-complementary VFA designs. And on Nexus, where the low-gain disadvantage is relatively moot.
Sounds like a pain, right? Yeah. But that still wouldn’t have killed the VFA idea.
In the end…the thing that killed the VFA designs…we did listening tests between The Originalist, Full Complementary, and Magni 3+.
And we preferred 3+.
Aside: yes, I know. Controversial. How were the tests blinded and matched? Can you really hear a difference below -110dB? Could we be fooling ourselves? Well, here’s the truth: of course we could be fooling ourselves. But, in blind listening, there was a remarkable consensus of opinion about the VFA Magnis—that they sounded a bit “hifi,” as in, boom and sparkle, maybe a bit of suckout thru the mids, a bit flat in presentation. Not horrible by any stretch, but the 3+ sounded more natural.
And so the VFA version was shelved. (But not lost, oh no, there’s a great application for it that’s coming…you’ll see, it’s just not a Magni.)
Current vs Voltage
So I went back to current feedback, or CFA, designs.
CFA aficionados will tell you:
- CFAs are usually simpler. CFA amps can be extremely simple—you can get quite decent results with as little as 4 transistors! Of course, there’s always a tradeoff, like maybe having to have a capacitor in the feedback loops. We don’t go that simple, because we want to be DC throughout.
- CFAs typically have very nice symmetry. Although you can do a non-fully-complementary CFA design (like Jotunheim 2), typically they’re fully complementary when you’re talking about single-ended topologies.
- CFAs are fast and easy to compensate. The “slew rate” in a CFA isn’t limited by the current in the front end and Miller cap size—it’s related to the impedance of the feedback network. This can lead to some truly spectacular speed (as in, bandwidth measured in multiple MHz and 500-1000V/uS slew rates) from a stage that is dead easy to compensate. It can also lead to spectacular confusion and failures when you don’t realize that the impedance of the feedback network also affects distortion performance, and, if the bandwidth is too wide, you’re gonna run into some really bad instability. (Yes, as in, if the feedback resistor value is too high, the THD+N performance will suffer, and if it’s too small, you’re gonna see magic smoke.)
I’d already tweaked the Magni CFA for best performance in its current form, choosing a specific feedback resistor and compensation method for highest THD+N without instability. But that didn’t mean I couldn’t attack it in the same ways I mentioned before—enhance the linearity of the gain stage, or increase the loop gain.
And boy did I attack it. The other ten or so Magnis were all CFAs.
Here are some of the highlights.
- Fancy VASes. For a current feedback amp, all of the gain is really in the VAS stage. This is one reason they don’t perform as well as a VFA—less gain. So making the VAS as linear as possible is key. I tried several versions, including ones with different, more linear transistors, degeneration, compound transistors, local feedback, and more. None of them got us to where we needed to be.
- Hawksford Cascode. I also tried Hawksford’s approach to the VAS…again, performance proved stubbornly intractable.
- Current Sourced Front End. Remember I said something about when you have an input signal that’s 33% of your rail? Yeah. One way to help the input transistors deal with this huge variance is to current-source them. This is what we do on Vidar. Unfortunately, it’s quite a few more parts. This, however, got us a lot closer to where we needed to be!
- Triple Output Stage. Maybe our problem was the load? Nope, it wasn’t. Also we lost swing and suddenly had problems with an unstable CFA. Bottom line, not worth doing.
- CFP Output Stage. Maybe if the output stage used complementary feedback transistor pairs? This would also increase the loop gain. In the end, this was also difficult to stabilize and didn’t get us where we needed to go.
- CFPs Galore. How about a CFP front end and CFP VAS? This is now squarely in the realm of increasing loop gain. Punchline: without the front end current sources, there wasn’t much improvement.
- Alt Simple Topologies. I tried a couple simpler CFA topologies. Performance went down, as you’d expect. Physics still works. Woohoo.
- Local Filtering. Maybe we could get a couple more dB performance if I tamped down the remaining power supply noise? Well, yes, but at the cost of output. It needed enough filtering that we wouldn’t be able to maintain our power rating for the amp.
- Error Correction. If you can’t beat ‘em, join ‘em? I replaced Magni’s servo with a fancy OPA1656 op-amp in an error correction format. Dead easy. Also didn’t sound great. Also would need a lot of resistors switched around when changing gain, so more complex. Also didn’t really do much on low gain. So no.
“Hey, wait a minute, what the heck are you talking about with CFPs and Hawksford and error correction and all that stuff?” you ask. “Is there a way to distill this down to non-engineerese?”
Unfortunately, no.
Let’s disambiguate one thing, though: CFA vs CFP vs current output amplifier.
- CFA is “current feedback amplifier,” as in, “my preferred amplifier topology.” As in, Magni, Asgard, Jotunheim, Ragnarok, Tyr, Vidar, and Aegir are all current-feedback amps. Lyr+ and Vali2++ too, if you squint a bit.
- CFP is “compound feedback pair,” as in “a pair of transistors, connected in such a way that they form a compound, or Sziklai, pair.” This can be all-BJT, all-JFET, all-MOSFET, or a mix to get a particular mix of characteristics, such as high input impedance and better linearity. However, like/like mixes are more apt to cancel distortion and provide higher performance, within the limitations of the N- and P-type device differences. This cancellation is why stacking two JFETs for use as a buffer can provide spectacular performance improvements relative to a simple source follower, but that case is like-like stacking (as in, N-channel and N-channel. In any case, a compound feedback pair is a way to get more linearity and higher loop gain, and also combine characteristics, such as using a JFET/BJT pair for high input impedance and higher gain. It’s a really neat circuit that has very high inherent performance, and it forms the basis of the Freya+’s differential buffer circuit.
- A “current output amplifier” is something that other companies do. It’s an amp with an intentionally very high output impedance. This is thought by some to be beneficial. I’ve never understood the fascination with them. This doesn’t mean they are bad. It’s just we don’t do them.
But, bottom line, one thing you’ll notice is that not much of what I did over all these Magni prototypes had much effect—with the exception of “current sourcing the front end.” So let’s talk about the whys and wherefores of that a bit.
Here’s the deal: the basic Magni architecture is really nice, really simple, and very high performance given its simplicity. But it’s easy to see where things start going whomper-jawed. Measure it at 1V RMS input in low gain, and you’ll get about -110dB THD+N. Measure it at 2V RMS, and performance decreases to about -106dB, with most of the increase being in an equal amount of 2nd and 3rd harmonic distortion.
Why does it do this? It’s because Magni runs on about +/-17V rails, and a 2V RMS input signal is almost 6V peak to peak. With a diamond input, this means that the voltage across the input transistors can vary from 11V to 17V—over 1/3 the total voltage. Transistor characteristics change with voltage across the device, so it’s not surprising you see nonlinearities popping up.
Aside: but we’re still talking about stuff that’s well below audibility.
If you want to get better performance, you’d feed the input transistors with a current source, eliminating much of the potential for variation. And when you do that, you get something that’s much more tolerant of big inputs. Pretty cool, huh?
Aside: but not enough to get us to that razor’s edge of physics, that magical realm where Heretic plays, around -120dB. Again, it shouldn’t be audible, but if you want to go for broke, you need more.
Aside to the aside: question for the class: why is -120dB, plus or minus a couple of decibels, about the limit for audio, when referenced to a couple of volts RMS?
Another aside to the aside: if you want a free performance upgrade, you can sometimes just rate at a higher output voltage. Magni may be -118dB SNR at 1V, but -124dB at 2V! Magic! Heck, it could be -130dB at 4V! And -136dB at 8V! Heck, it could touch -140dB at full output. That’s nearly 24 bits dood! Holy mole!
And another aside, because this is fun: you noticed I changed from THD+N to SNR there, right? But remember, when comparing specs, best to see what they’re referenced to. And if they are weighted. And if they’re in RMS watts. And lots of other things.
A final aside: and maybe you didn’t remember from the front of this giant chapter, but I’ll remind you here: all of this dev work and tweaking is referencing 300 ohm loads, because that has become an easy target. Lots of stuff falls apart at 32 ohms (and lower, these days)—and that’s where all the newer Magnis excel (3+ and up.)
Here’s the thing: to get a discrete Magni to the edge of what’s possible these days (in terms of THD+N performance), it has to become a different device. It needs to get bigger. It needs a lot more parts. And, if you want to get into truly magical numbers, you probably need more loop gain, or an entirely different, cascading discrete op-amp architecture. Both of which we don’t like.
So in the end, you get a Magni+ that simply builds on our well-known CFA architecture. It has a bunch of little tweaks, but it’s not significantly different. That’s why the board reads, “The Archetype.” That’s what it is: a solid, high-performing, great-sounding, affordable discrete headphone amp.
And that’s what it needs to be.
Leaving Las Vegas...Er, I Mean the Constraints of Magni
Now, after reading and digesting all of the above, you might be tempted to ask, “Given a clean sheet, could you get Magni near the best-measuring stuff—but without changing it wholesale?”
Heh heh. I asked myself the same thing. And if you can get a lot closer with current sources, you can get even closer with current sources and local filtering, and you can get even closer with the ultimate cheat, which is
raising the rail voltages.
Using big rails is something we did all the time at Sumo. Hell, our preamps ran on +/-35V rails. And those huge rails gave us a big performance advantage, enabling relatively simple circuits to perform very well. It even allowed me to do a truly bonkers preamp with no overall and no local feedback and great measured performance—in 1992. There was nothing like it at the time.
So yeah, real nice current sources, local filtering, and high voltage rails, and Magni is there.
The problem is, then it’s not a Magni anymore. It would need to be bigger. The power supply would have to be entirely different. It would not be priced at $109.
Fun fact: when I realized that we could do something like this, I thought, “It’d be nice if we could do a super-Magni in a bigger chassis, with all these tricks…” before I realized we already did Magnius. And I don’t know if the world is ready for a single-ended Magnius. And that isn’t really an Asgard either, because Asgard has a particular sound, and it’s a modular amp, and…and heck, I don’t know where that wacky idea would fit, or if it fits anywhere. We’ll see. No promises.
Of course, after reading all this blather, you also might ask, “Well, do the ultimate measurements even matter? What about Magni Piety? Why wasn’t that the new Magni?”
And that’s another great set of questions. Let’s tackle them, and then let’s break down the metrics that we have to factor in when designing any new Magni.
Do the measurements matter? Of course they do. Unfortunately, we don’t know entirely which ones matter or at what level. Focusing on a single number based on steady-state performance at a single frequency and output level is oversimplification. A more broad-based array of numbers that addressed steady state distortion, transient performance, and perceived noise level would be better, but that’s a lot harder to visualize, and transient performance is difficult to nail down, other than looking at square wave performance, and that doesn’t tell the whole tale, especially if you’re talking long-term transient performance such as operational point drift. Aaaand steady-state performance doesn’t have a clear line—is -80dB THD fine, or should it be more? Does it matter with -40dB transducers? Aaaaand even perceived noise is a question mark, because maybe you don’t hear it if the topology of the amp has the potentiometer after the noisy section, and any perceived thing about the amp (how fast the volume ramps, etc) is subjective, aaaannd…and you see how we get in trouble here.
What about Magni Piety? Magni Piety, for those who don’t know, is a Magni version that is produced in limited amounts by Nitsch. A lot of people love the way it sounds, and say it sounds like a tube amp. A lot of people think it should have won our first Magni deep dev dive—from 2016, when we did 3 different Magni versions, one that became Magni 3, one that had a simple Continuity output stage (like Asgard 3), and one that had something I was calling “the programmable output stage,” or Continuum™, which used a whole boatload of parts to allow me to curve the output stage to simulate square-law, or tube-like, devices. We had a number of people listen to all 3, and the results were mixed. Mike liked the “programmable output stage” variant best, I liked the “simple Continuity” variant best, lots of others liked the “one that became Magni 3” best. Given the low power and complexity of the “programmable output stage” version, it was never really in the running. The simple Continuity version also lost due to lower power. And the one that became Magni 3 went through a lot of tweaks before it made it to production, because it had features—most notably an optional DAC card(!)—that never came to pass.
Why’d we end up with Magni 3 rather than the other two 2016 candidates? In retrospect, it’s simple: because it won the power, cost, and measurements metrics. Whenever we’re doing an inexpensive amp, we have to take all three into account—because we know an inexpensive amp will be used by a lot of people in a lot of different ways, and it will be scrutinized by a lot of reviewers in a lot of different ways. It needs to be the most universal amp we can produce. And the simple “3” variant was the one that won: it had the highest power output, lowest cost, and best measurements. And it still sounded very, very good—much better than the Magni 2 and Magni 2 Uber it replaced.
So where do we go from here?
Well, we go forward with a great new Magni+ that is the highest-performing discrete Magni, ever, and the most flexible Magni, ever.
And we go forward with Magni Heretic, when ultimate measured performance is desired. For all my blathering about how op-amp based amps are simple and easy, we put a looooot of work into this one, including the aforementioned unique no-overall-feedback topology, plus the overcurrent and DC protection, plus the ESD protection, etc, etc. Hell, for IEMs, it’s pretty much impossible to beat its noise performance, since the inherent noise of the gain stage is divided as well as source noise.
Aaaand we go forward knowing a whole lot more about what we can do with discrete designs, and with a whole new/old gain stage in our pocket.
Finally, we wait, and watch, and take data, and review how things go, and probably do a few more prototypes…to see if we can, someday, have a single, amazing, powerful, flexible, affordable, great-measuring discrete Magni.
Until then, I hope you enjoy yours, no matter which one you choose!
