Advice for Schiit on how to improve Magni 3+ power supply

As has already been said: It's not what you say, it's the way that you say it.

I don't understand a word this guy was saying, but I very easily understand how he was saying it. A gnat's whisker of difference, and it might have been something all you engineers could have joined in with and had fun.

Just amazed by Jason. How many companies would even bother to notice stuff like this, let alone engage with it, give the guy a hearing, and manage not to say f**k off. Am ordering a whole pallet load of kudos for Jason and Schiit.

 

S/he's got a phase offset of 774.5 ms between the two measurements, and you're trying to position your H2 & H3 peaks to match?

 

Close.
The choice of signal generator frequency is independent of the DeltaWave measurements posted. It is one of the tweaks to position signal generator in the center of an FFT bin for a given FFT size and analyzer sample rate. It also zeros the point of discontinuity between the end and beginning of FFT buffers which minimizes measurement artifact side lobes, even without windowing.

 

This post demonstrates why measurements are hard, and will compare with original data posted by kris2014 using the tool he favored for data analysis.

Initial data presented in attempt to demonstrate an observable difference in Magni3+ power supply modifications on amp output noise by kris2014:

Now remember the experimenter claimed something appeared a bit weird in the Magni 3+ square wave response rise times in my technical measurement series. I looked at this display and thought there was something very amiss with this pair of measurements. If one wants to prove something it is necessary to have good measurements. As stated many times here, audio measurements are hard.

Let's begin with the source signal, in this case a sine at 7 KHz and -10 dBFS wavefile. I asked why 16 bits was used for a very good reason. We will examine the spectrum of our source 7 KHz sine at 16, 24 and 32 bits. No hardware is involved; just wavefiles created using Adobe Audition.

At 16 bits the 7KHz sine has numerous additional spurs outside of the desired 7KHz stimulus. Not reference grade material.


At 24 bits the the additional spurs are much lower in amplitude and far more predictable.


A 32 bit source wavefile finally achieves reference quality to be used as a source for the measurement system DAC.

The measurement system setup.
1. Source wavefile sine 7000.488 Hz, -10 dBFS, 44100 Hz, 32 bit
2. DAC output ADI-2 Pro, 0dBFS = +19 dBu
3. Magni3+ volume set for maximum (approximately 1 dB attenuation from input)
4. 300R load
5. ADC input ADI-2 Pro, 0dBFS = +19 dBu
6. DAW: Adobe Audition running 32 bit playback and recording of 12 second clips
7. DeltaWave v1.0.50 used for spectrum analysis
8. DAC measured output 8.89 dBu for -10 dBFS 7000.488 Hz sin source
9. Magni3+ measured output 7.84 dBu for -10 dBFS 7000.488 Hz sin source
10. Deltawave 0 dB is referenced to +19 dBu per ADC reference in item 5 above

Picture of setup:


First a DAC to ADC loopback test to characterize the measurement system performance:

Reasonably clean with a single 2nd harmonic component at approximately -125 dB below reference of +19 dBu.

Expanded y-axis to appreciate ADI-2 Pro performance capability as a measurement interface

y-axis is -10 to -210 dB. ADI-2 Pro noise floor is between -130 and -140 dB relative to +19 dBu reference.
Blue trace is the source wavefile for reference.

Now for the unmodified, stock Magni3+ measurement:

Using the same range as the original experimenter we observe a considerably different residual noise spectrum. Based on this result compared to the original, I'd venture the original measurement (shown first in this post) had many problems. I would not wish to confer any differential power supply improvements had been achieved by the original data until a clean baseline result has been obtained. Back to the lab.

Same unmodified, stock Magni3+ measurement with an expanded y-axis down to -210 dB:

Something a little unusual around 17 KHz, but also below -150 dB from +19 dBu reference or -131 dBu, approximately 218 nV!!!

I hope kris2014 will go back to the lab bench and figure out what is amiss in his measurement setup and try again, even if only for himself. Measurements are hard. Especially when dealing with levels below -100 dBu.

 

Ah, I missed this. Sorry. It's been another very long week. I'm sure it's been a long week for everyone, given everything that's happening.

To contribute: my comments weren't a suggestion to increase the power supply capacitance on Magni 3/3+, it was just to note that a cap multiplier (active stage) after a typical "jellybean" (317/337/78XX/79XX) regulator will reduce power supply wideband noise, and will improve the noise floor of a discrete circuit with low PSRR. For op-amps, it won't make a lot of difference, since their PSRR (power supply rejection ratio) is much higher.

So why not do this in Magni 3/3+? Simple, we'd lose volts in the cap multiplier--as much as 2-3V (total, positive and negative rails) when the amp was driven to its limit into low impedance loads. That would mean we wouldn't be able to rate the amp at 2.4W into 32 ohms (it does about 2.8W at 1% THD, both channels driven, and Magni Heresy does a bit more). So yeah, a numbers game--do you go for higher power or lower noise? Magni 3+'s measurements were dominated by distortion rather than noise, so decreasing power output to get the same numbers didn't make sense. Magni Heresy has huge PSRR, so post-filtering wouldn't matter.

If you want to do something interesting that doesn't throw away volts, DiyAudio has some interesting approaches that use the 317/337s inherent feedback, plus some extra parts. This approach seems to work well in some cases and not in others, we've seen some instability when we played around with it.

Also, whether or or not you'd be able to hear the decrease in noise is another matter--we're talking way way down here, like -120dB level. In circuits in with abysmal PSRR, yes. That's why Asgard 2 (not 3) used a bunch of filtering on the HV rail, because there was effectively 0dB PSRR. It hissed like a stuck toilet without the filtering.

Asgard 3 is much better PSRR but it uses the cap multiplier trick, which reduces the noise floor (and, yeah, to us, sounds better--and yes, I know, I know, this is purely subjective). And you'll see us use similar topologies in a lot of products, up to Ragnarok 2--which, yes, still uses the LM317/337 parts. I've been using them since 1990, and, after flirtations with newer regulators, I'm very comfortable with their performance when used with post-filtering for low PSRR topologies.

What's interesting is how this cap multiplier trick can even improve today's "low noise" regulators. Yes, they are low noise when compared to LM317/337 pairs, but I've gotten measurable--and significant--results using cap multipliers following the TPS7A39, for example. An upcoming product uses that part (impossible to DIY at a 20-pin 3x3mm part with a thermal pad, but very low dropout). Combined with lots of ground hacking and optimization, it kills even the typical -120/130dB 60Hz harmonics you see in our inexpensive products.

You want fun? Do a modern switching regulator, like we use in Hel. 0402 parts and exotic diodes and thermal pads. In this case, paying attention to the layout recommendations of the manufacturer makes total sense (even though they have two different recommended circuits between the datasheet and the demo board docs, LOL). Dave tried to do something with 0805 parts and his typical layout, and it turned out super bizarre. I followed the recommendations (mostly--a couple of tweaks at the output improved the noise, but it's not like they're gonna tell you to use heroic inductors and giant filter caps, because it typically won't matter.)

With regards to datasheets, we have designed products and started shipping with datasheets still at the 0.91 level (names of companies redacted,) and developed one product to completion (that should have been shipped in December/January) based on a detasheet that was later amended to address the performance problems we had with the product (which didn't meet stated manufacturer specs). These are newer devices, yes, and devices that are more complex, but the bottom line was the datasheets were not accurate--and, in both cases, the manufacturers thought it was perfectly fine to give us a data sheet still in beta-level for shipping parts. (And, right now, we're working with a third manufacturer with a similar problem--compounded by the fact we can't get answers that don't change weekly.) I wish it was more cut and dried, but it is not.

And, to go beyond datasheets, holy moly, the stuff I've seen with discrete design, the various interactions and oddities and instabilities--holy moly, sometimes I wonder why I bother. (Because, in the end, it's fun.)

Hopefully that provides some context. If I was testy, please let me apologize.

Also, many many thanks to AtomicBob for his measurements.

 

Rather than testy, I thought you were very patient and professional as always. Far more than I could've managed, though that is admittedly a low bar. Kudos, and be safe.

 

Thanks for the explanation, @schiit ! I wrote that in jest, but what I got was the best explanation for why/how you decided on what you did. Honestly, I couldn't ask for anything better! Cheers!

 

This thread reminds me of Tim the Enchanter warning the knights for the monster they are about to face. So yeah, I present to you...



P.S. I think Michael Bay was inspired by this scene.

P.P.S. After this scene the Killer Rabbit slaughtered the first batch of knights because they ignore Tim's warnings.

P.P.P.S. I thoroughly enjoyed this derailment, just so you know.