Building a massively parallel op amp power amplifier

Your parallel experiments are showing errors down in the >-110 dB range. Worse or better at that level from a sound card without a good distortion amplifier is at or below the resolution the test can deliver.

Especially if the distortion is already low, a very marginal improvement is easily lost in error bars of a test.

Reading your edit, I think we might be driving at the same point.
 
Your parallel experiments are showing errors down in the >-110 dB range. Worse or better at that level from a sound card without a good distortion amplifier is at or below the resolution the test can deliver.

Especially if the distortion is already low, a very marginal improvement is easily lost in error bars of a test.

Reading your edit, I think we might be driving at the same point.

You only need to read the two or three posts above the one I linked to to find answers to your questions about sensitivity limits, noise and distortion improvements, etc.
 
I designed and built a 16-amp-per-channel stereo 5532 headphone amp inspired by Self's Op-Amplifier.
Charlie, I had a look at your project thread; the quiescent current is indeed something to keep in mind. The quiescent dissipation of the LME49720 is comparable to that of the 5532 ICs, so keeping the supply voltage as low as possible will be important. As I wrote before, my preferred listening level doesn't require heaps of output power.

Instead of doing this work of folly, build a decent class-A amp of 20W a la FirstWatt, or a quiet single ended amp like the MR7-MKIV. These will be a much better use of your time and money.
I like a good class A amp as much as the next guy on this board. I really want to do something akin to this peculiar amp by Self though for this project.:)

Edit: the application note from Analog Devices that jcx linked to in that thread provides some solid info for op amp decoupling in various configurations, for example, when bridging amplifiers.
 
Last edited:
Founder of XSA-Labs
Joined 2012
Paid Member
parrallel and wrap in a multiloop with good audio input op amp

This is a cool looking design with big PC cooler. 200mA is huge - nice.

190037d1285862501-doug-selfs-ne5532-power-amp-thoughts-anyone-amp_front.jpg
 
A sneak peek at the circuit board layout (work in progress -- ignore the color scheme if you will; the preview has been configured to resemble OSH Park boards ;)):

attachment.php


I've also been optimizing some of the resistor values to further reduce the input stage noise without sacrificing distortion performance (in the sim, at least). Plus, a couple of jumpers have been added here and there for testing and verification of isolated parts on the board.
 

Attachments

  • 20171022 PCB layout work in progress.png
    20171022 PCB layout work in progress.png
    455.6 KB · Views: 2,236
You only need to read the two or three posts above the one I linked to to find answers to your questions about sensitivity limits, noise and distortion improvements, etc.

There's a difference between what the software says and your limits of detection. I did not read anywhere that really characterizes your measurement setup, just load conditions etc. Long ffts aren't enough (and may cause other issues). But I'm not trying to be adversarial, just to state that any improvement or not in your setup could easily be lost in the limits of the test. But it's also a low enough error as to be essentially meaningless for audio purposes.

Anyhow, in any case you're results are fine, please don't worry or let it dilute the discussion here.
 
Did you buy the op?

Good news everyone, if you haven't buy op amp, you should consider LME49860 for 2 reasons,

1. it's the same as LME49720 just with better voltage tolerance, 22V
2. I think it's end of product life, so on LME49860MA/NOPB - TEXAS INSTRUMENTS - AUDIO AMPLIFIER, CLASS AB, SOIC-8 | e络盟 台湾 it cost 1.5 dollar, half price compare to LME49720

If you know how to dissipate the heat, it will be a good choice, provides more power, +- 21V output voltage.

Heat is your enemy, quiescent current is around 10mA each, @JA 145C/W. Just the idle temperature will be 42*0.01A*145 = 60 degree, and you have to add signal power dissipate as well, I don't know how to calculate that.

Maybe a copper sheet on top of the SOIC top??
 
OP, given the design and differential input, would it make sense to drop the voltage further and go bridged?
I've considered the option earlier on, but for my specific case the added output power is unnecessary. Of course, you'd have to double the amount of op amps in the output stage on each board, because each amp into the bridged load sees only half the output impedance (e.g., 3Ω instead of 6Ω). So you end up with four times as many op amps per channel.

Good news everyone, if you haven't buy op amp, you should consider LME49860 for 2 reasons,
1. it's the same as LME49720 just with better voltage tolerance, 22V
2. I think it's end of product life, so on LME49860MA/NOPB - TEXAS INSTRUMENTS - AUDIO AMPLIFIER, CLASS AB, SOIC-8 | e络盟 台湾 it cost 1.5 dollar, half price compare to LME49720
Wow, that's a competitive price for that IC. The LME49860 is usually more expensive here than the LME49720 (for example, at Mouser and Farnell). If someone would want to get more power out of such a paralleled op amp output stage without going bridged, this seems like a good way to do it.

Heat is your enemy, quiescent current is around 10mA each, @JA 145C/W. Just the idle temperature will be 42*0.01A*145 = 60 degree, and you have to add signal power dissipate as well, I don't know how to calculate that.
I think you can roughly estimate it like this: suppose you have a 20Vp output signal (around 14Vrms) and you have enough output devices to limit the current each channel sources and sinks to ~20mA, the added dissipation per IC would be 2 (op amp channels per chip) * 14Vrms * 20mA = 480mW added to your 420mW of quiescent dissipation. So without extra cooling measures those ICs are way beyond the specified operating temperature range.

Any thread that says massively parallel should at least recognize this amp.
http://www.firstwatt.com/pdf/art_beast.pdf
Yeah, that's just brutal. :cool: In one of the other threads someone posted another example of an older amp using a massively parallelled output stage using discrete components.
 

PRR

Member
Joined 2003
Paid Member
You took 14Vrms, the _load_ voltage, and accounted it as _chip_ heat.

20mA peak sinewave is like 7mA RMS, under 300mW added. Most of that goes to load.

I don't think it is a real issue.

To check, build one chip with scaled load and beat on it.
 
20mA peak sinewave is like 7mA RMS, under 300mW added. Most of that goes to load.
I don't think it is a real issue.

Thanks for setting me straight here! I had a feeling it might be too pessimistic.

So, most of the heat generated will be the result of the relatively high quiescent current of the LME49720. After crunching the numbers some more, it seems like this IC wasn’t the best pick for a small signal op amp to use in the output stage, especially given how quickly these SOIC packages heat up if I want to maintain some headroom on the supply rails. All in all, something like the OPA1602 might be a better pick here. Its specifications are also impressive across the board, but its Iq is only 2.6mA per amplifier channel (about half of the LME49720).

Next to that, I’m going to add an option to wrap the output stage in the feedback of the buffer op amp after the optional volume control, like this:
attachment.php


If I can’t get it to be sufficiently stable, I can always revert to using local feedback only. In the simulator it looks okay though.
 

Attachments

  • 20171029 Tuned balanced input and output stage feedback loop.png
    20171029 Tuned balanced input and output stage feedback loop.png
    68.9 KB · Views: 2,638
… In the simulator it looks okay though.
I’m not sure what I was thinking when I put that feedback loop together, but it was a mess. Anyway, a few hours of tweaking later, the simulated small signal response looks much better. Turns out a little gain goes a long way when it comes to stability (which needed some attention because the paralleled emitter followers are slower than the buffer op amp). I’ve added a modest voltage divider before the output stage so that the output level stays well within the realms of what the op amps are able to handle. It now looks like this (simplified a bit for clarity’s sake):
attachment.php


The noise at the output has gone up a little because of R2, part of the voltage divider. Its thermal noise dominates the other noise sources here but is of course still negligible. The simulator is currently reporting THD+N numbers that I don’t expect to be even remotely attainable in the real world given my circuit board design experience. ;)

Additionally, the PCB layout is done and I’ve ordered a couple of boards. Hopefully, they’ll arrive within a few weeks so that I can start building the amp and further optimize the component values. Can’t wait!
 

Attachments

  • 20171107 Feedback loop stability improvements.png
    20171107 Feedback loop stability improvements.png
    29.7 KB · Views: 2,752
With parallel discrete output transistors we need to spend time matching them. I wonder if the same is true for parallel opamps? Btw, parallel output devices can never sound as resolving or coherent as a single output device (or opamps). The analogy is the sound of a choir vs a soloist. It's just the limitations of being able to match perfectly.

Data or pure speculation? This dosnt make sense. With the opamps the feedback "matches" them, and even if it didn't, if all outputs are in phase the currents sum together perfectly. If one opamp spits out. Io1=99(f(x))ma and the next one Io2=101(f(x))ma (where (f(x)) is the music signal, what's the total current? 200(f(x))ma. No "choir" going on.
 
Time for some updates

Time for some updates! A few days ago the PCBs finally arrived:
attachment.php


Pitch Black. :cool:

Because most components are surface mount type, I went with ENIG plating so that the pads are nice and smooth. Additionally, 2oz copper on top and bottom helps to minimize track and plane impedance.

I started by soldering and testing some of the op amp arrays:
attachment.php


Here’s some more eye candy with the instrumentation amplifier parts populated:
attachment.php


And right now, I’ve got two finished channels in front of me:
attachment.php


Now on to the important stuff. 1. Does it actually work? (After some debugging, removing a stray metal strand and touching up some solder joints, it does. ;)) And 2. How does it sound?

Both channels perform well beyond my expectations. I’ve already tried them with my regular headphones (HD650) as well as some sensitive in-ears. With the inputs shorted, they’re completely silent. Even with the in-ears there’s not a trace of noise or hum whatsoever.

As for the sound, one word that comes to mind is brutal. It sounds just brilliant with jazz and classical, but the authority with which it reproduces bass-heavy tracks while maintaining perfect separation is just stunning. This is definitely going to be my main amp once I’ve got it neatly packaged.

Right now, the differential inputs are AC-coupled. The Nichicon bipolar caps can be bypassed with some jumpers. However, I didn’t hear or measure any difference when doing so. DC offset is just 0.7mV on channel 1 and 0.4mV on channel 2. The difference between the two can be traced back to offset voltages at the instrumentation amplifier inputs.

Before starting this build, I was a bit worried about keeping temperatures in check. Especially the quiescent current from the LME49720 could prove troublesome here. Switching to the similarly specced OPA1602 seems to have worked out just fine. After some 20 minutes, both boards are about 30°C above ambient at the warmest spot. Of course, this is still with the amp sitting on my desk, out in the open. I reckon that under heavy load, in a closed case, temperatures won’t exceed 70-75°C.

If I do decide to actively cool the boards in the future, they can support a 90mm fan. Also, there’s some exposed copper on the bottom layer which can be thermally connected to the bottom panel of a metal case. For now, it doesn't seem necessary though.

As the photo's show, I may have went a little overboard with the buffer caps at approximately 0.11F per channel, but I reckon it won’t hurt. The Nichicon UHW1E562MHD capacitors used here have an ESR of only 10mΩ.

I’ve been using the QA401 analyzer to measure the performance of both channels as I was building them up. Unfortunately (or fortunately, depending how you want to look at it), the distortion is below the measurement floor of the analyzer itself (about -115dB; see attachment at end of post). Because of this, I’m not able to discern whether putting the paralleled output op amps in the feedback loop of U48 (an OPA1611 buffer) improves performance. For now, I’ve just configured the IC between the instrumentation amplifier and the output stage as a unity gain buffer. I haven’t done any intermodulation distortion measurements yet because they're slightly less straightforward to set up with this analyzer, but I expect to run into the same limitations with those given the op amps used here (OPA1602 in output stage, OPA1611 everywhere else).

All in all, I’ll probably have to build a distortion magnifier if I want to characterize the distortion performance of this amplifier more accurately. I’m leaning towards the one designed by Bob Cordell. His Linear Audio article on the topic mentions a kit that looks good.
 

Attachments

  • 20171121 Blank PCBs for op amp power amplifier.jpg
    20171121 Blank PCBs for op amp power amplifier.jpg
    324.9 KB · Views: 1,743
  • 20171121 Populating an op amp array.jpg
    20171121 Populating an op amp array.jpg
    640.7 KB · Views: 1,796
  • 20171122 Adding the instrumentation amplifier.jpg
    20171122 Adding the instrumentation amplifier.jpg
    282.6 KB · Views: 1,684
  • 20171124 Both channels built.jpg
    20171124 Both channels built.jpg
    620.9 KB · Views: 1,962
  • 20171124 Comparison - Local loopback into QA401 (differential).png
    20171124 Comparison - Local loopback into QA401 (differential).png
    155.7 KB · Views: 391
Congrats. Trying to count...is that 40 op amps per channel? Massively parallel indeed! Would you say that timing is precise and the sound is focused? Wondering if the "choir" effect is possible or not.

Interesting that you can't hear any difference when bypassing those Muse bipolars. I have tried them for coupling and preferred Silmic, KZ, and Panny FR. But if BP sound the same to you as a wire, well, I am realizing that must mean I prefer coloration over accuracy. Yikes. :eek:
 
Congrats. Trying to count...is that 40 op amps per channel?
That's spot on; there are 40 dual-channel op amps per board.

Would you say that timing is precise and the sound is focused? Wondering if the "choir" effect is possible or not.
I'm not sure where xrk's observation earlier in this thread stems from w.r.t. such a choir effect, but there's nothing like it here. Maybe if you managed to introduce large phase shifts varying across the output devices?

Interesting that you can't hear any difference when bypassing those Muse bipolars. I have tried them for coupling and preferred Silmic, KZ, and Panny FR. But if BP sound the same to you as a wire, well, I am realizing that must mean I prefer coloration over accuracy. Yikes. :eek:
Haha, that doesn't have to be the case. I've oversized the input caps to minimize distortion and low-frequency roll-off. If you're using smaller value capacitors, you might also simply have an audible high-pass filter. (Have a look over here for some measurements of the Muse bipolars.)