2019, Chapter 18:
13 Answers
While I was writing the Full Heresy Mode chapter, I threw in some questions. Seriously dorky engineering questions.
Questions like, “Why would you expect an op-amp design to measure better than a discrete design?” and “Why would adding a driver stage to a discrete design cause the design to have to be re-compensated?”
And, after I did the chapter, I added a few more questions into what I called the “Engineer’s Quiz.”
Like this:
- Why would you expect an op-amp design to measure better than a discrete design?
- Why would you expect both op-amp and discrete designs to measure better in low gain than high gain?
- Why would adding a driver stage to a discrete design affect the distortion performance?
- Why might you have to re-compensate a discrete design when you add a driver stage?
- What other techniques can be used to make a discrete amplifier measure better?
- What is feedforward also known as, and how is it different than feedback? Bonus: write the s-domain equation for feedforward and negative feedback. Bonus bonus: recount the number of times you used s- or jw-domain simplifications as an engineer, and name the mathematical discipline they are simplifications of.
- Which of the current TI audio op-amps claims highest distortion performance?
- What factors will contribute to an engineer not replicating the claimed distortion performance numbers for an op-amp design?
- How large of a difference will jack contact and cable routing make when performing measurements on low-distortion amplifiers?
- What is the typical distortion level of a good loudspeaker transducer? (A range is fine, or multiple ranges for bass/midrange.)
- What is the typical distortion level of a good headphone transducer? (Same notes above.)
- How does transducer distortion compare to amplifier distortion in terms of level and profile?
- What is the typical distortion level of a good recording microphone? Bonus: what is the distortion profile of the chain used to record your favorite music (microphone, preamp, processors, ADC, etc)?
Two brave souls took a shot at answering these questions. Both did pretty darn good (in my opinion, anyway, especially since there are some gotcha questions there.) Both, however, have won a pair of Magni 3+ and Magni Heresy, so they can compare for themselves.
So, please, Ableza and tincanear, please contact me via PM or email (
jason@schiit.com) so I can get you your amps!
(Yes, you read that right. We’re giving away amps because you took the quiz.)
That said, several people also asked me to post up my answers to the quiz. I agreed this was a good idea. And, as I thought about it more, I realized that all the answers deserve some discussion, so let’s turn it into a chapter (and let’s add some bonus questions at the end.)
And yeah, I know, for the less engineering-oriented, maybe this is a huge snoozer. In that case, please accept my apologies. There’s at least the year-end wrapup chapter coming, and (hopefully) Jotunheim R.
But that’s enough blather. On to the answers:
1. Why would you expect an op-amp design to measure better than a discrete design?
One word: feedback. Actually 2 words: more feedback. Or, actually, to be pedantic, higher open loop gain which allows more feedback to be applied to the closed loop circuit.
The more feedback, the better the measurements.
Op-amps have very high open loop gain. An ideal op-amp would have infinite open loop gain. The OPA1662 and OPA1688 used in Magni Heresy are spec’d at typical open-loop gains of 114 and 130dB respectively. The OPA1688 also adds feedforward internally (in addition to high open loop gain.)
This means that Magni Heresy has 100-114dB of feedback in the voltage gain stage, and 130dB in the current gain stage. These are huge numbers. And this is why it produces great measurements.
In comparison, Magni 3+ has much lower open-loop gain, only about 55dB, which means it has only 40-55dB of feedback, depending on the gain switch position. Despite the lower feedback, Magni 3+’s measurements are still very good. This is a testament to the inherent linearity of the stage.
Bonus: tincanear also called out the fact that op-amps have better beta matching, thermal coupling, and offer the ability to use more complex topologies (cascading, etc) which can lower distortion.
2. Why would you expect both op-amp and discrete designs to measure better in low gain than high gain?
See the previous answer: more feedback. And the more feedback, the better the measurements.
So why wouldn’t you go for infinite gain and apply infinite feedback? Well, a couple of reasons:
A. Higher gain means a more complex gain stage, and the more complexity you have, the bigger chance you’re going to end up with instability. Amplifier designers deal with what we call gain-phase plots and phase margin, where if the phase of your output signal goes to 180 degrees with positive gain, you don’t have an amp, you have a power oscillator. Which is no bueno.
B. There is a broad range of anecodotal, subjective experience which suggests that lower feedback is sonically “better.” Now, I know this is venturing into crazytown for those who think that measurements define the totality of audio. But there are companies, like Pass Labs and Ayre, who make “no feedback” part of their point of differentiation. Now, “no feedback” usually translates into “lots of local degeneration to enhance linearity, and lots of attention paid to making a very linear amplifier.” And that approach can sound very good. That’s what we did in the original Asgard and Valhalla. But sometimes adding feedback makes things better—or at least the tradeoffs are more in the direction you want them to go with feedback in place. Feedback certainly makes getting low output impedance much easier in a speaker amplifier, and in ensuring consistency in production. So it’s not as simple as “no feedback is best.” At least not to me.
3. Why would adding a driver stage to a discrete design affect the distortion performance?
It improves distortion by unloading the VAS.
Most discrete designs have the majority of gain in the second stage, the voltage amplifier stage, or VAS. This VAS can then be used to drive the outputs, but its high impedance is loaded by the outputs, which have limited current gain. (Or, in the case of MOSFETs, they have significant input capacitance.) If you add a driver stage, you dramatically increase the current gain of the output stage. If the drivers have a beta of, say, 250, it takes 250x less current from the VAS to drive the output stage.
The lower the load, the better the distortion performance of the VAS.
4. Why might you have to re-compensate a discrete design when you add a driver stage?
If you unload the VAS, its bandwidth increases. You also add a pole for the driver stage. This can affect your phase margin. And instead of an amp, you may end up with a power oscillator, aka “magic smoke generator.”
Which means you may have to change the amplifier’s compensation to get it to be stable again. “Compensation” is the practice of adding small, local or global, frequency-dependent feedback loops to reduce gain at high frequencies, so you don’t end up with too much gain when the phase shift hits 180 degrees. In practice, this usually means adding capacitors at various points in the circuit (sometimes in a single stage, and sometimes in the overall feedback loop.)
Bonus from Jason: but you don’t add capacitors in the overall feedback loop of a current-feedback amplifier! Oh holy moly no! Current feedback amplifiers are unusual in that their bandwidth increases not only with more feedback, but with the impedance of the feedback network. Lower impedance = higher bandwidth. Go too low and you can violate your phase margin, even if the overall amount of feedback looks the same.
5. What other techniques can be used to make a discrete amplifier measure better?
Tincanear got this one spot on. He said: “cascoding (fixed VCE or VDS, or Vplate-cathode), apply local feedback at each gain stage (emitter or source degeneration), grade and sort transistors for higher linearity per expected load-line, feed-forward.”
And yeah, now I know the non-engineers think we’re all just making this schiit up, but he’s serious. I’d just add one thing: complementary feedback pairs, like the Freya+ differential “buffers,” which are no-overall-feedback designs. No, wait, and distortion cancelling with same-type devices, like the JFET buffers in Mjolnir 2. No, wait, and brute-force methods like using huge voltages to make the operational area tiny in terms of overall swing. No, wait, I’ll probably remember a few more…argh.
This would be a heckuva long chapter if I tried to go into what all of these techniques are. Most of them result in higher complexity, may change the poles of the amplifier or add poles you need to deal with, or will have other effects that have to do be dealt with. Sumo’s feedforward was a source of instability that had to be dealt with separately from the amplifier.
But, bottom line, there are lots of techniques that can be used to improve the linearity of a discrete design…it’s just that all will come at the cost of some complexity.
6. What is feedforward also known as, and how is it different than feedback? Bonus: write the s-domain equation for feedforward and negative feedback. Bonus bonus: recount the number of times you used s- or jw-domain simplifications as an engineer, and name the mathematical discipline they are simplifications of.
Feedforward is also known as “error correction.” It is neither new nor a panacea. Sumo used it in the 1980s. I designed amps with feedforward for Sumo in the 1990s.
Nor is it really all that interesting in itself. It is just something that has grown to mythical status, kinda like “FPGA.” The only real difference between it and feedback is that feedback feeds back the entire output signal to the input, reducing gain and increasing measured performance, while feedforward feeds back only the error signal (input minus output), correcting only for errors in the signal.
When used correctly, feedforward has its place. Bob Cordell wrote an excellent paper on the application of error correction to a discrete MOSFET amplifier, which was an extension of the work Malcom Hawksford did in the early 1980s, and the basis for Sumo’s implementation.
Does it sounds like I’m negative on feedforward? No, not at all. It gave Sumo a real advantage in the power amp game, allowing us to get near-zero output impedance, low distortion, and huge current output, without a ton of overall feedback.
Want the equations? Cordell’s paper has them:
http://www.cordellaudio.com/papers/MOSFET_Power_Amp.pdf
As far as when I’ve used s-domain equations, I can recall exactly once: in an argument at an audio show with another engineer, over whether or not our error correction was really feedback or not. Yeah, really stupid stuff.
Hint: for those of you who are thinking of becoming engineers, don’t knife yourself in differential equations, s- and jw-domain simplifications will save your butt. I have never used a single differential equation in my career.
7. Which of the current TI audio op-amps claims highest distortion performance?
Aha. This one is kinda a gotcha.
An astute engineer will ask, “Highest distortion performance into what kind of load, at what kind of output? Are you talking a line driver or receiver? Differential? Or are you asking me to just go to the parametric tables for the SoundPlus product lines and cough up a number?”
Great questions. Because TI offers quite a range of products for a whole lotta needs. If you go to their site and search their SoundPlus products, it’ll give you a whole page of options with the worst of them coming in at 0.00005% THD, and the best coming in at 0.000015%. That is an insane amount of zeros.
But what do you want? It might be tempting to choose the OPA1612, at 0.000015%, but it’s a pricey part…and it isn’t the best at output current. So maybe it’s not the best choice for muscle. And what does the input offset current look like? How will it behave if you put a volume pot in front of it?
Bottom line, choosing an op-amp depends on what you want to do with it. For high current, the OPA1656 is very nice…but so is the OPA1688, which we chose because it’s already an active part (and when we tried the OPA1656, the numbers didn’t change much.) For front end, we chose the OPA1662, again because it’s an active part we already use and like. But we’ve also used the LME49720 for differential applications, and a couple of exotic parts from Analog Devices in Mani.
8. What factors will contribute to an engineer not replicating the claimed distortion performance numbers for an op-amp design?
Again, tincanear nailed it: “power supply bypassing; poor pcb layout, esp grounding and consideration of loop currents, input network impedances interacting with input parasitics.”
Here’s the thing: when you’re playing down in the parts per billion distortion range (which all these op-amps are in, as well as Magni 3+ in its best operational range), you’re dealing with really, really, really tiny signals. And PC boards aren’t perfect. Run two traces next to each other, and they will couple to each other capacitively and inductively. Run them on top or bottom of each other, and it might be better or worse. Parts aren’t perfect, either. Resistors are little inductors and capacitors, capacitors have parasitic R and L components, individual transistors vary, and on and on.
So, it can be fairly easy to have an unintended feedback loop, or unintended compensation, in one or both channels, which will affect performance. Which means you may not hit those published numbers.
The higher the power output, the worse it gets. Amps like Vidar, Aegir, and Ragnarok 2 go through multiple rounds of finger-biting layout and testing to optimize their routing, and that optimization can change distortion performance by 10-20dB.
Similarly, Magni 3+ required a 4-layer board for best performance. And Magni Heresy, despite several revs, still doesn’t quite have the same distortion performance from channel to channel at very high output. We’re talking 0.00015% vs 0.0002% levels here, insanely low. My current guess is that we’re actually seeing some EM coupling to the switch (the worse channel is closer to the switch.)
9. How large of a difference will jack contact and cable routing make when performing measurements on low-distortion amplifiers?
When measuring Magni Heresy and Magni 3+, it was a common occurrence for me to curse and go back to the bench because one channel was bad—I mean, really bad. Like 20dB off. And I’d stare at the PC board, looking for the wrong part, or the bad solder joint, and, after finding nothing, I’d plug it in again…and everything would measure fine.
Then next time it would be bad again.
Intermittent board? No. After a while, I learned to turn the TRS plug in the socket if one channel looked wonky. 99 times out of 100, that fixed it.
So, yeah. Bad contact or bad cables can bork your measurements by 10-20dB. That’s despite using custom cables (Mogami, nothing fancy) and high-quality connectors (Neutrik.) (And no, we don’t hard-wire for measurements, that’s cheating.)
Want more bad measurements? Tangle the wall-wart cable in the signal cables. Boom. Tons of hum. (Still inaudible, but the measurements were off.)
10. What is the typical distortion level of a good loudspeaker transducer? (A range is fine, or multiple ranges for bass/midrange.)
This is another gotcha question, for several reasons.
· What is a “good” loudspeaker transducer?
· Are we talking about a conventional cone speaker, or are we open to exotic designs?
· What frequency range are we talking about?
· Can you even find the measurements?
This whole question was prompted by us arguing internally at Schiit about -106dB vs -114dB THD+N, and me happening on a measurement of distortion for the KEF LS50s. The KEFs claim 0.4% from 170-20Khz at 90dB.
That’s -48dB THD.
-48dB THD? That’s crazy nuts bad, right?
In short, not really.
In fact, it’s actually excellent for a conventional loudspeaker. You might notice they left off the bass frequencies, but that’s because distortion will rise rapidly at lower frequencies. The LS50s might be only -38dB (or worse) in the bass.
And that’s still good!
Because, heck, KEF actually published a spec for THD for their speaker, which means they were proud of it. Lots of companies don’t publish THD for speakers at all.
Now, there are exotic speakers out there (electrostatics and double-sided planars) that can do better, and there are some blindingly expensive conventional designs (like the short-coil, long-gap stuff from ATC) that can do better, but here’s the reality: the THD of the transducer is the dominant THD in any audio system.
(Well, unless you’re using tube amps. Then maybe…)
11. What is the typical distortion level of a good headphone transducer?
Headphone guys now may be thinking, “Well, headphones must be a lot better than speakers, right?”
Um, well, no. Not really. Just like with speaker manufacturers, most headphone manufacturers don’t specify THD.
The ones that do include Audeze, and they spec the same number for much of their LCD series: 0.1% at 100dB, no frequency range spec’d.
This is better than a speaker, yes, but it’s still only -60dB.
Yes. -60dB THD.
And remember…this is a GOOD headphone.
12. How does transducer distortion compare to amplifier distortion in terms of level and profile?
To put it simply: transducer distortion dominates most playback systems.
Even middling solid-state amplifiers will be performing at an order of magnitude better than the best transducers—or better. The best-measuring gear will do two or more orders of magnitude better. Some tube amps may end up being similar to transducer distortion, though.
Not only that, the distortion profile of transducers is different. The more a transducer moves, the more distortion. Which means distortion increases significantly in the bass—which is different than most solid-state amplifiers, which will tend to have distortion that is relatively constant versus frequency.
Fun fact: Rising distortion in the bass is actually similar to a tube amp that uses a transformer-coupled output.
13. What is the typical distortion level of a good recording microphone? Bonus: what is the distortion profile of the chain used to record your favorite music (microphone, preamp, processors, ADC, etc)?
Aha! This is the ultimate gotcha question!
Well, okay, maybe we can answer the first part—what the distortion of a good recording microphone. But I’m having a heckuva time finding many specs.
Ableza says it’s about 1% at 1kHz. If that’s true, then it’s about -40dB.
Anybody else have more data?
But the ultimate gotcha? What’s the distortion profile of the chain used to record your favorite music? Ah, you gotta be kidding me, right? I mean, unless you were the recording engineer…no, wait, nevermind, even if you were the recording engineer, would you have the entire chain characterized? Or is it more than one chain, if you’re talking about a studio album? Or something that didn’t ever exist at all, as with electronic music?
Bottom line, there’s a lot of complexity in audio, and it doesn’t always lend itself to easy explanations. I hope you enjoyed looking into some of this complexity in more detail!