Hi All,
Lars has gone on vacation so it’ll be just me today. It may make for a less entertaining read than the Q&A we did for Audiophilestyle, but I still hope it’ll be worth it.
I see only feedback being employed. Would it make sense to also use feed forward? Why use one or the other (or both)?
The term “feed forward” isn’t nearly as clear-cut in its meaning as “feedback”. I have seen the term used for a variety of quite different circuits, of which I’d say the only potentially useful one is one that uses a second, very fast but low-voltage (still high current though) power stage bridged with the main high-voltage power stage, and which is fed the distortion isolated from the output of the main power stage so it cancels out across the load. What this does is it gives you exactly the same distortion performance that you’d have gotten with a normal feedback circuit of the same order if your main power stage had been as fast as the small one. You can easily see that adding that second power stage, which would have to be class AB on account of bandwidth and which would need its own power rails is not an attractive proposition from a practical perspective.
I’ve also seen the term “feed forward” used to denote a circuit that explicitly measures the error in an emitter follower power stage and then, instead of feeding it into a second power stage, subtracts it from the input of the EF power stage itself. That of course is plain feedback by another name.
Feedback and feed forward topologies fully share that most important unchangeable fact of life “whatever happens, stays happened”, meaning that either can only respond to errors after they’ve occurred. All you can then do is spectrally shape the error so that as little as possible is left in the audio band. Feed forward doesn’t offer a fundamental escape from this.
Confusingly, the kind of method that does get round this is called “forward error correction”. What that means is that you try to calculate the expected distortion just before it happens, instead of waiting for it to happen and then measure it. I would consider that entirely feasible, but it would be complex enough to merit designing a custom IC.
About correlation between sinewaves and music, we see more and more the 32 tones test (32 sines equally spaced on log scale from 20Hz to 20 kHz).
From an instinctive standpoint, it looks like being closer to a musical signal than 1 or 2 sines. Is the instinct not to be trusted in this case?
There’s nothing wrong with your instinct, and the 32 tone test is perfectly sensible. Keith Johnson already used a cluster of 8 tones in the 90’s. He said that for AD/DA this was an excellent indicator of sonic transparency. To my ears his HDCD encoder box is still the gold standard in sonically transparent AD/DA so I’m not about to disagree.
The more complex the test signal, the more likely you are to separate out the different nonlinearities that together make up amplifier distortion. On the other hand, it becomes harder to read the graphs. Once you use more than 2 sine waves the plots become quite hard to interpret. You have to do some calculations to work out what particular combination of tones underlies a particular spur.
In our particular case an attraction of 2 tones is that it allows me to use one output channel of the AP for each tone and then sum them together passively. That way I know for sure that the test signal itself is totally free from IMD regardless of the performance of the AP’s built-in DAC. That still leaves us with the performance of the ADC though, but at least it’s an improvement. With an amp like the 1ET400A it’s rather difficult to do an IMD test that’s not limited by the test kit and that only gets worse with a larger number of tones.
It’s a judgment call. Conceptually you can always break an amp up into a combination of ideal integrators and static nonlinearities. If you want to be able to fully characterise the system (i.e. figure out what parameters to put in a circuit model to be able to simulate the system with an arbitrary signal afterwards) you’ll need at least as many tones as there are nonlinear circuit blocks. That’s not what we’re doing here though. We don’t want to build a model telling us exactly what sonic character the amp will have. We only want to gauge how far it is from not having any. For that purpose you can usually do with fewer tones so long as you’re sure that some pathological combination of circuit and test parameters can’t conspire to produce an unrealistically good readout.
In other words, you need to know something about the circuit you’re testing. For instance, in an optimally biased class AB amplifier the 19+20 kHz test would fail to expose thermal modulation of the bias point, so you’d need to do further testing with a signal that produces significant low-frequency power as well.
Anyone who claims to have found the ultimate test battery that will validate any piece of audio kit (and believe me, I’ve met a few who claimed that there was nothing to amplifiers but a few THD numbers) had better consider the Test Daemon. The Test Daemon is a second Audio Precision that your mortal enemy has programmed first to recognise any test signal and then to set its generator accordingly. That thing, measured as a black box, would pass every test you can do using your first Audio Precision with flying colours but it would fail most amusingly at playing back music.
In Benchmark's literature, they state:
“The important difference between the AHB2 and class-D amplifiers is that the AHB2 does not produce switching noise. Class-D amplifiers are measured with brick-wall AES17 filters that ignore the switching noise above 20 kHz or 40 kHz. Tweeters can demodulate this ultrasonic noise and fold it into the audible band.”
That’s a curiously shaky argument to put forward in favour of a fine amplifier like the AHB2. It doesn’t need to make any excuses. I imagine that it was just a throwaway comment, otherwise it’d qualify as Spreading Uncertainty and Doubt.
I am fully aware that literature exists testing tweeters at blistering levels with signals close to the audio band, but it’s quite a stretch to extrapolate from there to a small 500kHz residual and side bands. Keep in mind that the electrical impedance of a speaker goes up with frequency so the residual (heavily attenuated by the LC filter) amounts to an absolutely tiny current. But hey, should anybody really want to make a point of this claim, the onus is of course on them to provide data showing that the switching residual of a good class D amp adds distortion to the acoustical output that isn’t there when the switching residual isn’t there. If such data were around, it’d probably be online already.
That’s not a very informative answer, so let’s see if we can take the argument anywhere useful. This is about demodulating the sidebands of the switching residual downward. The word “noise” is potentially misleading here, it’s not noise as in “hiss” but as in “signal I didn’t pay for”. The switching residual is actually a highly predictable signal.
Anyhow. Demodulating requires an even order distortion mechanism. One which, as several of you noted, would have to be present even when the voice coil isn’t moving anymore. For the sake of an argument, let’s just imagine that there is something in the tweeter that is able to turn not just the current but also the square of the current into a force. If you designed a tweeter
specifically to be that bad, you could actually do that, but it wouldn’t do much else besides: the exact same nonlinearity would also produce absolutely massive amounts of distortion in the baseband in the regular way.
The hypothesised additional distortion would be the intermodulation between the carrier and its sidebands, and between the sidebands amongst themselves. Spectrally they would therefore appear exactly where regular distortion components are already present, except much lower. These regular distortion products are caused by the same nonlinearity but with a much larger input signal (the actual audio signal instead of the filtered residual). The hypothesised fold-down distortion would be much smaller than regular distortion and indistinguishable from it because it’s in the exact same place.
If we put in a better tweeter, regular and fold down distortion (if any) would fall in tandem. There’s simply no scenario in which fold-down distortion could appear separate from its associated regular base-band distortion and/or with remotely comparable magnitude.
From that thought experiment we can conclude it’s rather unlikely that any measurable or audible mechanism exists that can be traced back to the carrier residual.
As an aside I find the “demodulation” argument also a bit of an own goal when proposed in favour of high-res audio. What they’re saying there is that maybe a bad tweeter might manage to produce distortion components inside the audio band which wouldn’t be there if either the tweeter were better or the bandwidth of the audio signal were limited to 20kHz. That sounds more like an argument
against mindlessly doubling the sampling rate every 3 years and inventing a new format every decade.
I've wondered where timing comes in this wide bandwidth discussion.
We’re well and truly off on a tangent here but anyhow: it comes in nowhere. If you take two signals (say, square waves, or sharp transients) of which one is delayed by 1us compared to the other, and then you band-limit both to 20kHz you'll find that the 1us time difference is still there and perfectly visible on a scope as a horizontal offset between the two filtered signals.
A correctly implemented sampling system, that is one with a LPF before sampling and one after reproduction, is indistinguishable from the band-limiting filters alone, by any means, whether by frequency
or time domain testing (can’t stress that enough). We have the Sampling Theorem to thank for that. So with the sampler in place this small delay remains perfectly intact. In fact an AD/DA can resolve the time of arrival of a transients within a much tinier fraction of a sampling period, limited only by SNR (the signal theoretical pendant of Heisenberg’s Uncertainty Principle). IIRC some listeners can, under careful listening conditions, detect inter-channel time delays on the order of 2us (in the form of a minute left-right shift of the stereo image). This is perfectly within the ability of even 44/16 to reproduce. An ability to discern microsecond scale timing does not mean, and has never meant, that you need 500kHz sampling. Sadly, precisely this misconception is wilfully being repeated over and over by some shysters on the “high res” front.
Do you consider the 1ET400A the lowest distortion module of this technology, the "sweet spot" of the range so to say? Will higher powered modules have a little higher distortion as well, as per previous works, or that is not the case here?
The same laws of physics apply so you should expect somewhat higher distortion from higher powered offerings. If we wanted to get the same low distortion from a higher powered amp, the most cost-effective way forward would be going full bridge so as to keep using the faster, lower voltage FETs. I couldn’t say if and when such a thing is going to turn up on our roadmap.
Also your views on IC versus discrete opamps (measurements and sonics), and opamp rolling to taste in general?
This is about buffers isn’t it? There seems to be quite a thing going on about those. For the record I’d like to state that the buffers that we put on the interface board of the eval kit are deliberately unpretentious. If I’d done a nec plus ultra discrete circuit, we’d then have to explain to potential industrial customers that no, this is not in fact something that is required to make the amp work. But if your aim is to use the module for a separate power amplifier, we’d recommend something better.
The only reason to do a discrete implementation of anything is when it’s not available as an IC. The weak spots in common op amp IC’s are in my view:
1. Higher distortion in noninverting mode, especially when driven with a high impedance when variable input capacitance becomes noticeable.
2. PSRR limited to loop gain by lack of an explicit GND reference node. The OPA1622 is the only exception here.
3. Nonoptimal pinout. I don’t know why magnetic coupling is almost wilfully ignored by IC designers. If any are listening: the + and – supply pins should be DIRECTLY next to each other, and the same goes for the +/- input pins.
What I don’t consider a problem is class AB operation. That only becomes a problem if your supply pins aren’t in adjoining positions…
As soon as someone puts out a chip that fully fixes these outstanding issues I’m done doing discrete op amps.
”Too many to mention” said:
<Long discussion about CMRR, too long to quote>
The module input is designed to have good enough CMRR for signalling inside the box, e.g. between a processing board and the amps. It is not intended for direct connection to the outside world. Of course we could window dress the spec by using even higher precision resistors, but that would not materially change things. You want reliably good CMRR under all real-life conditions which includes realistic amounts of resistance mismatch on the source end. This point has been made at length by Bill Whitlock of Jensen Transformers (his InGenius chips are mentioned in the thread), who was the first to draw attention to the fact that a high common mode input impedance is crucial. The papers and white papers on their web site are well worth the effort of registering a user ID there. The basic idea is also rehashed in my “the G word” article (whose main intent was to introduce the idea of differential component placement and routing for “intra box” communication, on which all my circuits build).
The takeaway point is that you should not rely on the amplifier module for your external CMRR. 60dB will bomb proof you against anything that may happen from one end to the other end of a very noisy chassis, but not against an unknown signal source at the far end of 100 metres of microphone cable.
If you expect your product to see a lot of CM noise, invest in a proper input stage. We can't put one on the module because that would also make it more expensive for anybody else, certainly if we also want that buffer to have distortion and noise specs comparable to the amplifier.
About the famous 22dB difference: the input of the module is literally just a pair of resistors feeding into a virtual short, which also receives feedback from the output though another pair of resistors (one to speaker out, one to speaker gnd). CMRR as tested here is determined by the pairwise matching of these resistors. So you would indeed expect the occasional unit with excellent pair matching by sheer luck, and hence much higher CMRR.
”Too many to mention” said:
It took me a bit to understand why CMRR and PSRR were mentioned in one breath in some posts but of course that must be because, at DC at least, CMRR and PSRR are linked in operational amplifiers. That’s not the case here. CMRR of the amplifier is something that the user can improve at libitum with a modicum of circuitry, whereas PSRR can’t be improved externally unless you want fix things in the supply. Many amplifier companies like to turn a weakness in their design into a sales argument by drawing attention to their big transformer and huge bank of caps, but it seems to me that the better design is the one that doesn’t need such desperate measures. I’d like the PSU performance to affect only the clipped output power – as it does in this review – but not other performance specs.
I currently have an SMPS1200 and two Ncore 400 modules on order to go into a Ghent case. When the Purifi modules become available, is there any way you could include instruction and a wiring harness to connect it to that SMPS1200 if I just want to swap out your modules for my NC400s.
Our connector board is already compatible with the SMPS1200 with only a small flat cable tweak. If you want to fit the modules in an existing case you won’t escape some soldering though, because it’s unlikely that the connector placement on the interface board will fit your case. But it all in all I expect it to be painless even without instruction.
Generally we won’t be supplying explicit documentation about “how to exchange your Hypex modules with Purifi modules” though. My departure from Hypex was taken very sportingly and I feel they treated me more than fairly (this is still an understatement). It’d feel cheap to say the least if I then went out of my way specifically to poach from their existing customer base, other than the ones that are following me of their own accord.
And a wish, if possible make a drop in replacement board to the ICEpower 125ASX2, as it is very popular.
When will I be able to buy a 2.1 DIY kit? Will you sell 2ch boards with integrated power supplies, like the NC252 board? I browsed around a bit and it seems your retailers play in the range "if you have to ask you can't afford it".
Our focus is to be an industry supplier, so although we’re happy to sell unit quantities to DIY customers, we don’t expect to start making products aimed expressly at that segment. The companies you mention aren’t “retailers” but manufacturers that just happen to be designing in our stuff.
Towards the end of the year we hope to have a web shop running, for both the drivers and the amp modules. At that point, pricing will be transparent, at least for small quantities. Till then I try not to go on record with prices other than repeating the line that prices will be “perfectly competitive”.
Other modules will come step by step. Once we have a SMPS design we can start planning the integrated PSU+Amp range. I shouldn’t expect a direct drop-in replacement for other people’s amps though, unless it turned out that the board shape / connector placement / connector pinout of the 125ASX2 happens to be optimal and impossible to improve upon.
(Yes, the 1ET400 does bear a more than passing resemblance to a pre-existing product. Initially I deliberately tried to design away from that for reasons I just mentioned, but it turned out that I would actually have to make technical compromises if I wanted to deviate from what was pretty much the most natural layout flow. All I might have done perhaps was mirror the pinout entirely which would have been a bit silly).
Will the Purifi 1ET400A(s) become available to the DIY market in this exact configuration tested here? Especially regarding the input section.
Yes. That does also mean that we’re not expecting to go out of our way to supply input boards with better input buffers. I’m sure that a healthy cottage industry will spring up of people providing alternative input boards. Means we can keep concentrating on the amp and psu bit.
Do you expect to be marketing any speakers or drivers this year? Are you also considering headphones?
Drivers, yes, fully built speakers no. As with the amps, we supply drivers to people who build speakers. Headphones, no. Lovely subject, but outside our focus for now.
Will the i2c interface documentation be published? Is it utilized for status infos and/or controlling of the amp?
Yes, the full form data sheet, as will be put up as a pdf, contains all the gory details. It provides a complete readout of supply voltages, temperature, clipping and overcurrent indication, and of course control.
Does Purifi amp sound better than NC500?
I’m not sure if this is the right place to start that kind of discussion. But, expecting no difference, I found the difference rather larger than I expected.
What is the Purifi driver dispersion?
Figure 5 on the online data sheet has 0, 30 and 60 deg plots. The dispersion is in line with that of a piston up to about 2.5 or 3k.
What are you looking for in aspiring employees -or do you take any hippie who happens to get past security? Do they have to know what all the technical audio-jargon is or is it enough to guess the hot end of a soldering iron?
Brilliance and diligence in equal measure. We love hippies so long as they have a sense of humour about it. Seriously though, for now we’re set because we’re still in the investment stage and growing staffing levels too quickly is rarely a good recipe.
That's it for now folks, I'll check back in a week!
Cheers,
Bruno (and Lars, from a sunny campsite in Sweden)