You may have seen Resolve, Griffin, and Caleb’s threads already, and having a place to post general impressions, thoughts, and discuss things outside the context of a full review is pretty neat. So here’s mine!
I’ll be posting everything from thoughts on gear to breakdowns of topics and questions that get brought up here and elsewhere, but if there’s something you’d like to ask me here is also a great place to do so.
Excellent. I’m quite keen to get your impressions on the Topping D900/A900 setup. IMHO, I prefer the Topping setup to the Chord Hugo TT2.
I wanted to kick this thread off with a bit of talk on a topic that has come up a few times in the discussion of powerful amplifiers for headphones, that being the use of Class-D amplifiers.
Class-D is something that has become all the more prevalent in the speaker world, with many hugely powerful and excellently performing class-D options becoming available, and in situations like subwoofers, seeing anything other than class-D is almost unheard of! So why don’t we see class-D amplifiers in the headphone space more often? Or in fact how come there are hardly any at all?
Well, if you’ve not watched it already, I’d first recommend checking out my ‘Amp classes explained’ video, as this should help explain in fairly easy terms what differentiates Class A, AB, and D amplifiers.
Class D is unique in that rather than just amplifying what goes in, directly outputting an as-close-to-identical copy of the signal as possible, it actually “creates” a new signal. A stream of very high frequency, high amplitude pulses, which when filtered to leave only the low frequency (<20khz when possible) content, leaves only an amplified version of the signal that was originally fed to the input.
But there is a problem: This means class-D amplifiers need to have an output filter to remove this unwanted high frequency content from the output, and typically, the load/speaker itself actually forms a part of this filter circuit:
This in turn means that class-D amps and their filters need to be designed for a specific load impedance. And a change in impedance of the load will cause the filter to behave differently. We can show this by just doing a frequency sweep through a class-D amp with different load impedances. Here is one done by Amir at ASR of the Fosi Audio V3 with a 4 ohm load and an 8 ohm load:
As you can see, with the 8 ohm load, the frequency response amplifies high frequency content a fair bit, whereas with a 4 ohm load it attenuates it. This implies the amplifier was designed/optimised for likely around a 6 ohm load.
This is generally sensible given as an ‘8 ohm’ speaker will probably have dips in its impedance vs frequency curve to values lower than 8 ohm and so aiming for 8 ohm on the dot would likely not provide the best actual result.
With speakers, you can be pretty certain that almost any speaker is going to fall in the 4-8 ohm range with some exceptions. You have a narrow impedance range to target and so you can design your filter in a way that will work similarly well for the vast majority of speakers. Even if you put a 4 ohm speaker on a class-D amp designed for an 8-ohm load, you might attenuate the uppermost frequencies by a fraction of a dB as shown above, not really a big deal.
Headphones are quite a different story…..
Headphones can range anywhere from 2 - 1000+ ohms, and so designing a class-D output filter with a target impedance in mind is going to leave you with something that only works properly with a small number of headphones no matter what value you target.
Aim for a ~60 ohm impedance for planars? Well now almost any 300-600Ohm dynamic driver headphone is going to be way off. As will any planar that happens to be lower at say 20 Ohm.
Vice versa if you target ~300 ohm dynamic driver headphones, then low impedance planars will not perform correctly.
And with greater impedance differences come greater risks……
Below is a simulation of the frequency response of the class-D output filter circuit at the beginning of this post, with a few load impedance values ranging from 4 ohm to 155 ohm (Modhouse Tungsten)
The following values were used to try to get this filter design to match the behaviour seen by the Fosi amp in the ASR measurement:
L = 11.26uH
Cbtl = 0.703uF
When this hypothetical Class-D amplifier is connected to either a 4 or 8 ohm speaker, it will behave pretty much exactly as seen in ASR’s measurement of the Fosi amplifier. Under 20khz any differences are pretty small and unlikely to be of any concern. And with most ‘8 ohm’ or slightly below speakers it’ll behave pretty much exactly as intended.
But when we swap to higher impedance headphones, suddenly our high frequency content gets MASSIVELY boosted, not just by the couple dB we saw earlier…..
Even a 32 Ohm headphone sees a +2dB boost in audible band content by 20khz, and a roughly 15dB boost in content around 40khz.
If you were to put the 155 ohm modhouse tungsten on this amp, you could see a 30-40dB boost in content around 40khz. Go for a 300 Ohm headphone and it’d get even worse.
The elevation in upper treble is not ideal, but not necessarily the end of the world. However this enormous boost in content just above 20khz IS a problem, and you risk damaging your headphones or possibly ears, maybe even without realising it’s happening.
You might not be able to hear 30-40khz, but if you were playing ~90dB music, and accidentally pumping content up to 130dB or higher through your headphones, you could absolutely damage your hearing or your headphones.
Drivers don’t care what your upper hearing range limit is, they’ll break all the same if you overdrive them. And studies have shown that exposure to ultrasonic content can indeed damage hearing.
The TLDR:
If you use headphones on a class-D amplifier, you could be damaging your headphones and/or your hearing, and because you can’t directly hear the dangerous content, you might not even realise it’s happening until it’s too late.
So, that’s one of the main reasons we don’t often see class-D amplifiers in the headphone space. They can be difficult to make work for the massive range of impedances we see in the headphone market, risky, and personally I’d strongly recommend avoiding them.
That being said, some more modern designs from HypeX/Purifi use a feedback approach that does indeed make them load invariant and avoids the issue shown above. Keeping a flat frequency response up to the cutoff frequency regardless of the load impedance.
I’ve yet to test whether running high impedance headphones on these amps then causes a degradation in other areas of performance or higher distortion, but will do so whenever I next get ahold of one to test.
The other reason however is simply: Class D just isn’t needed for headphones.
Class-D amplifiers have an immense advantage in efficiency compared to Class-A and Class-AB amplifiers. Often comfortably above 90% compared to the ~40% on a REALLY good day of a class-A amplifier. Or 50-75% of typical AB amplifiers.
If you’re building a 3000W subwoofer, you can’t have a class-A amplifier that is dumping 7000-10,000W of heat into the room at all times unless you’re building a sound system that doubles as a sauna heater. And even a class-AB amplifier that gets 75% efficiency would mean that 3000W sub could be pulling roughly 4000W from the wall with around 1000W of that being dissipated as heat. Your subwoofer is going to need some SERIOUS heatsinks and cooling.
A class D amplifier with 90% efficiency would pull 3300W from the wall to output 3000W to the sub, and only 300W of heat would be generated at most. Probably far less during general operation, leaving you able to fit a nice compact amp into your sub, with maybe a small, hardly even noticeable heatsink on the back. Some like the SVS SB-16 ultra even get away with just dissipating what little heat they need to through the backplate itself despite the amp being capable of up to 1500W continuous or 5000W peak!
This is what allows Class D amplifiers to be so compact yet so powerful. Their huge efficiency advantage allows you to put much more powerful amps into smaller spaces with less heat dissipation as you just aren’t burning anywhere near as much energy as wasted heat. Almost all of what power comes into the amp from the wall is going directly to driving the load.
This is a big factor when you’re dumping hundreds of watts into speakers, and ESPECIALLY when pumping thousands of watts into subwoofers.
But with headphones, where you are using at the very most a few watts with even the most demanding headphones on the market, who cares if your 6W amplifier pulls 13W from the wall vs 20-25W? Five to ten watts of extra heat is almost nothing, and even class-A headphone amplifiers can still be pretty small because….well….dissipating twenty or so watts of heat is very easy. Your own body outputs about 100W of heat into the room at rest or 300-400W if exercising.
Class-D solves an important challenge in speakers. But with headphones, we simply don’t need the efficiency benefits of class-D when we are dealing with such relatively tiny amounts of power and heat anyway, meaning there is little reason not to just build class A and AB amps for headphones even before we consider the risks described in the first part of this post.
There are also impedance issues with OTL headphone amps. Whilst they match up well with the Sennheiser HD 600 series, I found that the very low impedance headphones, the sound is not as good with the OTL tube amp as it is with a SS Class A or A/B amp.
This is a bit of a different thing.
Firstly it’s worth noting OTL just describes whether a tube amp has a transformer coupled output or not. It doesn’t determine whether the amplifier circuit itself is Class A or class AB.
Both OTL (output transformerless) and OTC (output transformer coupled) tube amps can be either class A or AB.
OTL amps will typically have high output impedance though, which will then modify the frequency response of a headphone according to its impedance vs frequency. This is different to the class D output filter effect, in fact most class D amps will have extremely low output impedance and so the impedance curve of the speaker/headphone isn’t really much of a factor/concern
Dropped in here to say thanks. Some of the tech is not the easiest to follow, if one is NOT an electrical engineer, but you all here and the folks on ASR, Head-Fi, in each of your own unique ways, is making an impact. Helping us to see through the scams, dodgy specs, and focus on what really matters. Your content is informative, entertaining, and thought provoking.
I’d love to see you release your own products, cos we trust you. Yeah a Golden sound DAC, an affordable dongle DAC, that would be a treat.
And by the way GoldenSound, great to see your physical evolution. Glad for you, as I could also do with similar changes. I can’t be the 1st person who has complemented you about this. Keep well, merry Xmas, best wishes.
I salute all the clarity you have introduced to the world of audio. Amazing. Well done.
I would definitely love to make my own DAC, though honestly, particularly at the lower cost end of things, at this time I’m not too sure how much I’d be able to add. With stuff like the JDS Element IV, Crinacle Protocol Max etc available, there’s some really great choices for not much money and I don’t know I’d be able to improve on them.
I have been working on a streamer, as one area where I have been frustrated with the lack of good, low cost options is the streamer market with pretty much only Wiim being available, I’ll share some updates on that in this thread too. But for DACs, Getting costs down to compete with the existing options whilst providing a product that is as good or better would be quite difficult, and I also wouldn’t want to cause conflict of interest concerns with the products I’m reviewing.
For dongles, honestly grab Crinacle’s one, it’s great!
If I were to do a DAC, it’d probably be a higher end DAC that has more of the features and design aspects I personally think more expensive DACs should really have, that’s where I feel there is stuff I could add or a gap in availability of what I personally would like to see; higher performance upsampling, high quality built in EQ (maybe even room correction), more flexible subwoofer integration features etc
Is the boost in upper frequencies the reason why Class D amps have the cold & sterile sound stigma? Or is it just placebo effect coming from people asuming PWM nature of Class D as digital.
Are these features not able to be present in a less expensive product?
Starting to sound like a DAC/Pre to me, just add volume control and you’re there. Maybe something like a Metrum Adagio with more features in the digital domain?
Some like the EQ are definitely, and are already in various more affordable products. Aspects like the high performance oversampling likely not. The SERCE module used in the Wandla for instance is a few hundred dollar part by itself, as is the FPGA in something like a Chord Mscaler.
Room correction it’d depend how far you’d want to go with it as the number of filter coefficients you have allows you to be more precise and flexible. Stuff like what’s done in MiniDSP (with Dirac or with convolution) for instance is quite limited compared to what you can do in Roon without the same limitations. But doing this on device also then needs more processing power.
But also the issue is just that making “a dac” isn’t too tricky. Making a DAC that does the stuff I want, paying developers for the necessary work, and since I can’t make something with the economies of scale of a Topping/SMSL is going to mean that it’s not going to be able to be <$1000 too easily.
Keeping the streamer I’m working on under $400 without the additional BOM and dev costs of the stuff above has already been really hard and the margin on it is gonna end up tiny.
Yep, exactly the kind of answers I was looking for, especially because it answered the sub question I had (What would these features be?), thanks!
Sorry, you’re saying there are other costs than the BOM? Unbelievable! 
The subwoofer features would be:
I’d be kinda interested in having a ‘high level’ sub output as well for connecting to stuff like REL subs high level inputs. Though would need to look further into whether those do any sort of loading and therefore whether ‘standard’ (although higher voltage) outputs would be fine or whether you do actually need a low output impedance and/or current capable amplifier output.
(To be clear I think sub high level inputs are kinda silly and just create various issues/challenges but hey people like them so might as well offer that feature if it’s doable)
OUTRAGEOUS!
I remember reading (or maybe hearing) somewhere that the “high-level” connection on REL subs has an input impedance of more than 100 k-ohms. If that’s correct, then no appreciable loading.
But what’s the advantage of having that connection on the DAC? Isn’t the point of having the high-level connection that the same signal (amplifier distortions included) goes to both the subs and the main speakers?
I would assume he’s trying to implement a digital crossover when he’s doing digital processing.
It’s the “right” place to do it if you are going to do it.
When I look at the internals of high end DAC’s, the primary design concern seems to be power, and isolation of the digital section from the analog section.
I own two high end DAC’s that sound very different, one is R2R, NOS with an optional correction filter, the other is over sampled DS, and while NOS has a sound, I remain unconvinced that the digital section of a DAC is really what people are “hearing”.
Having said that I think there is a market for something with highly customizable digital processing.
On streamers, I’d be interested in the overall approach, I looked at this at one point, having looked at a few high end options, there is a lot of effort again put into physical layout and dealing with power noise. Most of them though are just low powered ARM boards running Linux and OSS components, the Wii line (which are basically just OK) seem to run some custom software components which is interesting. About the only real exception is at the very high end when you get into effectively custom PC’s, the Antipodes still run linux and OSS, but there was at least one running windows, and custom drivers.
I’ve always thought the right solution here was to go the exact opposite way, you need almost no CPU power to stream data out to an I2S connection, so why run software on an OS that’s not a good real time OS to start with, go minimum processor, and custom software stack. The way UPNP (and Squeeze) works, you literally just end up with an unencrypted HTTP download for the audio data (Quboz works the same way for online streaming). The fact that you need a network connection, does dictate some minimum CPU power, and you probably need at least a SPDIF driver, so there is probably some bare minimum CPU spec but it’s probably not GHz. The core idea being minimize the power draw, and make dealing with the power noise easier.
I would certainly hope that’s the case, but it’s not something I’d looked into previously so wanted to dig a bit deeper.
As to the advantage, well it’s mostly that regardless of whether you’re using high level or regular outputs, the other subwoofer integration features could be a big benefit.
Usually when integrating subs, you position the subs wherever they are ideal in the room, and use a low-pass filter built into the sub itself. On some subs like REL this is fairly basic, usually just a knob on the back. On stuff like the SVS subs it can be a bit more advanced, with the option to adjust filter frequency exactly rather than guessing based on knob position, change cutoff slope to match different speaker designs etc.
But with this approach, you can not only low pass the subs, you can high pass the speakers to achieve a PROPER crossover, as well as adjusting crossover width and linear/minimum phase to taste, plus be able to compensate for the difference in timing if the speakers and subs are at different distances from the listener, and probably most importantly, apply independent room correction to each speaker and sub, not just trying to correct the speaker and sub as a pair.
You can’t do any of this when you’re feeding the subs from the same output as the speakers.
I’m actually doing this in my speaker system currently, using Audiolense for measurement + correction, then implementing the filters in either roon or HQP.
I’m using an RME ADI-2 Pro as an 8ch ‘DAC’, feeding channels 1/2 out from the analog outputs to the subs, and channels 3/4 bitperfect over AES to the Holo May then to the main speakers. That way I’m still using my preferred DAC, but now have full 4ch correction, crossover (not just low-pass), timing offset correction, and the result from that is quite notably better than just using a typical stereo output with the subs built in low pass filter and combined speaker/sub correction
There’s one item I was going to ask you about since you said it makes audible differences:
Surely whatever you’re hearing must be due to the noise shaper, if you go with their recommended settings and use 4x or more oversampling, no? How could you possibly hear if the filter’s slope is god-tier or just-good, once it’s way up at 176k or wherever? Unless I’m missing something, once generous oversampling is used, all those gazillions of taps to get as vertical of a filter slope as possible is just a way to use your CPU as a space heater, right? 
The 176k refers to the way the calculations are done, not to the frequency of the signal (music) that is being created.
In the end it’s the 20-20k that is produced but between the raw sample input and the final output a lot of intermediate calculations take place.
Using higher rates of oversampling doesn’t shift the nyquist frequency higher. The nyquist frequency/bandwidth of what you can reconstruct to is determined by the sample rate of the original file.
Most music is 44.1khz, so you have up to 22.05khz of bandwidth. DACs typically oversample to 8x or 16x rates initially (then often further to 11 or 22Mhz as part of the delta-sigma modulator), but this is just about more accurately reconstructing the <22.05khz stuff, you can’t get back information above that as it was never present in the original file.
Higher coefficient count filters allow you to design filters with a combination of steeper rolloff, deeper stopband attenuation.
This means you can avoid unnecessarily attenuating stuff under 20khz, and/or leaving >22.05khz content remaining.
A typical DAC filter might have responses like this (topping D90):
All of these either attenuate <20khz content slightly, meaning there can be audible treble rolloff, or leave aliased content above the nyquist frequency remaining, meaning they’re not an accurate reconstruction.
NONE of them actually fully adhere to Nyquist theory by leaving all <20khz content untouched but still attenuating fully by 22.05khz.
DACs with better filters like the Ferrum WANDLA or Chord DACs have more processing power put to this task and can have more effective filters, where you can get full, un-attenuated bandwidth right up to 22.05khz, with more instant attenuation of everything above it:
The noise shaper is a separate thing. This can help extend dynamic range within the audible band, though how much is needed is pretty up for debate. Some companies such as Chord design their modulators to be accurate to -300dB or so, though of course analog noise floor is going to be high above that. Though usually modulators are more about how they have somewhat unpredictable knock on effects on the DAC. If you swap modulators on a dCS DAC for example, or try different HQP modulators with a DSD DAC, you’ll see different distortion behaviour at the output of the DAC even though just looking at the audio data itself the THD isn’t changing.
IMO the filter is the bigger effect, which makes sense given as it is directly affecting content in or very close to the audible band.
I did a video on the topic here:
I mean normally it does, when it’s software oversampling and the low-pass is handled by the DAC, but I guess you’re saying PGGB itself doesn’t, as a matter of design choice. Interesting, that makes sense then because you want to have the benefit of that brick wall always at 20-22k (or whatever the file was transporting) and at the same time keep whatever DAC you have from messing with that careful reconstruction. So you push out a sampling rate where the DAC won’t touch anything anywhere near 20k with its own far worse filter. Am I getting warmer?
Because if Nyquist is right, whatever information is captured in the original samples at the original rate is sufficient for perfect reconstruction, and anyway no amount of oversampling can add anything from the analog original to that. So we oversample just to keep our dirty DAC’s hands off of our superb filtering results at 20-22k, right? (Force it to filter at way higher frequency, in case we can’t set it to filterless/NOS.)
