With so many DACs available and more DACs coming out all the time, it's impractical to compare them all and even Rob himself has indicated he doesn't have much practical experience with other DACs but knowing the architecture those DACs employ, he is aware of the constraints those DACs are bound to. As new DACs come out, it is inevitable people will want to know how the latest DACs will perform compared to another DAC. I thought it would be appropriate to present a compilation of Rob's posts on the matter. Rob's responses are in blue.
Here is Rob's response to a question that has been asked of him many times:
"Rob this is just a philosophical / hypothetical question i have thaught of for a couple of month now, and that is if you feel that the DAVE are superior in the digital and analog section from any other high end DACs, or if you had the oppertunity to handpick bits and solutions from other companies, have you then come up with an even better solution if your budjet where in the $100.000 like the MSB Select II or the DCS Vivaldi stack?"
"Clearly if I thought other solutions were better, then I would already be doing it. I have been designing DAC's for over 27 years, and designing with my own DAC technology for 22 years. That's a long time. Also, I am the only DAC designer who has designed silicon chips too, and had a very successful career on that side. Its given me a very valuable insight into the engineering problems of DAC design with silicon, as well as valuable insight into how these devices are developed.
I do get to hear other DAC's at shows, and am interested in the technology. But I have not heard or seen anything that has caught my attention. And just because its 100K does not make it better - too many audiophiles listen with their wallets rather than their ears...
I can say for certain that I do things that I know make a big difference to performance, and no other DAC designer does these things.
I can also say for certain that Dave sets new measurement standards for DAC's at any price - unfortunately only an APX555 is capable of demonstrating that as Dave has better performance than most test gear. But as a listener how it measures is not too important, but its not appropriate for me to comment on sound quality against other products.
But I will say that I have never heard a DAC sound anything like Dave, it is very different. Only you can decide whether it works for you or not."
Regarding the issue of "transparency."
"Designing for high performance audio - when you are interested in making something as truly transparent as possible - is often fraught with uncertainty - you often can't be sure of the right way to do things.
But for the interpolation filter in a DAC there is absolutely no uncertainty - if you want to recover the original analogue signal in the ADC at the point it was sampled perfectly you must use an infinitely oversampled interpolation filter with a sinc impulse response. It is as simple and clear cut as that.
The problem is that you can't have infinite oversample, and you can't have an ideal sinc impulse response as this needs an infinite amount of processing - so you have to optimize things to suit the processing you have, and this is where lots of optimizing with careful listening tests come in. But the point I am making is that the closer you get to the ideal, the more accurate you can make the reconstruction.
There will come a point where if you increase the oversampling rate and increase the tap length and you will hear no sound quality changes - in short its as close as it needs to be to the ideal - but we categorically have not reached that point yet. That's what is so amazing about how sensitive the ear/brain actually is whilst playing back music. And from a measurement POV we are a long way away from an ideal interpolation filter too - you would need about 100,000,000 taps, or an FPGA that has 100,000 dsp cores (Mojo has 44 dsp cores) for it to measure perfectly. Of course Mojo from a measurement interpolation filter POV measures much closer to the ideal than any other non Chord DAC I have ever seen - at any price point."
Regarding the challenges of building something like the DAVE:
"Absolutely - I don't think people realize how much effort goes into developing something like Dave, particularly as every step in the signal path needs to be designed. Dave's first prototype PCB was designed over three years ago, and that was when the decision was made as to which FPGA, based upon price and - much more importantly - availability. FPGA companies have a terrible habit of launching vapor ware, with actual silicon being readily available many years later.
As you say, far more important than the size of the FPGA is knowledge - particularly with the comparatively massive FPGA that is in Dave, and this is something that is on-going. In particular, lots of interesting things are waiting to be discovered with the Davina ADC project."
Regarding how a DAC is supposed to "sound."
"But nobody has heard the perfect DAC (that is one that adds nothing to the sound) so I was prepared to run with it - I certainly was not going to add some noise floor modulation to spice it up a bit. But then I found what was the technical reason for why Hugo had the ability to hear the starting and stopping of notes clearly - and I then improved that ability by designing a WTA filter that ran at 256 FS (that's an FIR WTA output every 88nS). Now this improved the ability to perceive the starting and stopping of notes - and when you can perceive transients properly things sound much faster and sharper and brighter as a consequence.
This actually restored the balance (I am uncomfortable with that term as still nobody knows what the perfect DAC would sound like for sure as perceived balance or neutrality is merely an average of all DAC's) and now it was sounding fast and sharp (when the recording demands) and soft and smooth (ditto).
If you are ever fortunate enough to get a good seat at a classical (unamplified) concert you will know what I mean - real life can sound ultra smooth, and ultra fast and sharp at the same time. I want my audio to have that range of variation - that's my goal anyway."
Regarding the benefits of Pulse Array DACs:
"The key benefit of pulse array - something I have not seen any other DAC technology achieve at all - is an analogue type distortion characteristic. By this I mean, as the signal gets smaller, the distortion gets smaller too. Indeed, I have posted before about Hugo's small signal performance - once you get to below -20 dBFS distortion disappears - no enharmonic, no harmonic distortion, and no noise floor modulation as the signal gets smaller. With Dave, it has even more remarkable performance - a noise floor that is measured at -180dB and is completely unchanged from 2.5v RMS output to no signal at all. And the benefit of an analogue character? Much smoother and more natural sound quality, with much better instrument separation and focus. Of course, some people like the sound of digital hardness - the aggression gets superficially confused with detail resolution - but it quickly tires with listening fatigue, and poor timbre variation, as all instruments sound hard, etched and up front. But if you like that sound, then fine, but its not for me."
And then there's this one that discusses Pulse Array and jitter:
Pulse Array DAC is innately jitter insensitive. What is not readily appreciated is that different DAC architectures have very different sensitivity to clock jitter. DSD is horribly sensitive to jitter, R2R DAC's are very sensitive, but pulse array is innately insensitive. The reason for this is that signal switching activity is completely signal independent - it switches in exactly same way whether its reproducing 0 of fully positive or negative. Because of this, when I get some clock jitter, it only creates a fixed noise. Now one of the really cool things that happens today is that PC resources and simulation tools are so good today, in that I can write a simulation, and add some jitter to the simulation, then measure the results using an FFT. From this, I can see exactly what jitter and only jitter does - and this technique has revealed a few surprises. But what it has done is proven that adding random jitter creates zero signal correlated effects to pulse array - no distortion, noise floor modulation at all - just an insignificant level of unvarying random noise. This does not happen with other DAC architectures, as you will then get significant noise floor modulation, distortion and noise shaper related noise. This is because with the other types of DACs, the switching activity is signal related. So DSD has maximum switching reproducing zero, and no switching at 100% modulation. R2R has no activity for zero, but considerable switching activity when the signal changes.
Regarding R2R DAC architecture:
"The problem with R2R DAC's is they create large amounts of distortion - and R2R fanboys like the sound of distortion.
I don't. I want to hear the sound of un-amplified music exactly as in the original acoustic, not some souped up distorted sound. And Hugo, when it first came out, got me so much closer to the sound of real instruments in a real space that I had not experienced before with any kind of digital equipment.
Each to their own I guess."
Here's another:
"Except the job of a DAC is NOT to reproduce the sampled data perfectly but to reproduce the original bandwidth limited analogue signal that was in the ADC before the signal was sampled. And to do this one must convert from a sampled signal and convert it to a continuous waveform - and that actually implies infinite oversampling, something that a R2R DAC can't do as they are limited to 16FS oversampling due to speed and glitch problems. That's one reason (there are many others too) why Mojo filters to 2048FS and has its DAC run at 104 MHz, unlike any other non Chord DAC's."
And another:
"Another major problem with R2R DAC's is their complete inability to accurately reproduce small signals, as it is impossible to perfectly match the resistors - this is categorically not a problem for my pulse array DAC's as element mismatch creates fixed noise not distortion, as all the elements carry the audio signal (unlike R2R DAC's)."
Here is Rob's response to my post as I discussed the details of resistors used in my TotalDac d1-monobloc:
"I will defer to Rob as always but I have my own contribution to make as I used to own a very good R2R DAC, the TotalDac d1-monobloc. This DAC incorporates 400 Vishay Foil resistors and Vincent Brient, its creator, went out of his way to use the more expensive variety with a very fine tolerance of 0.01%. Bought in bulk, these resistors sell for about $20 a piece and so for this DAC, the resistors alone cost about $8,000 (a reason why R2R DACs are so expensive). Obviously, he felt it was important to pay this premium from a SQ standpoint because the Vishay Foil resistors with a lesser tolerance of 0.05% cost 25% less. He could have gone for the best (0.005% tolerance) but that would have more than doubled the cost to $50 per resistor. Even at that, this suggests no 2 resistors will be 100% the same."
"Just to give you an idea of the scale of the problem - imagine that we wanted a R2R DAC that could resolve to an accuracy of -350 dB - the performance that I get with Dave's noise shapers, and the spec for which was based on listening tests of depth - then if we want to guarantee -350 dB accuracy then that implies a resistor accuracy of 0.000000000000003%. That's why the idea that you need noise shapers to be that accurate is so crazy - and the implication that the brain can detect this level is also crazy. Indeed, I was actually reluctant to talk about this issue, as the numbers are so tiny. But at the end of the day, I stand by my listening tests, and I have been pleased that other listeners has discovered the same thing about Dave in that it portrays depth very unusually.
But you may argue that there is something odd about noise shaping that does not apply to R2R DAC's in regards to depth - I think not, as I have had un-measurable small signal errors that degrade depth, and all R2R DAC's suffer with measurable small signal distortion - but there is another way of looking at the problem. One of the interesting things I got out of the Dave project was reducing high order distortion products and finding out how very audible they were. To match Dave's THD performance you would need resistors of 0.000003% tolerance.
Now whether you need 0.000000000000003% or 0.000003% they are both not possible to achieve. Even if you could match those levels, by hand selection, you could never guarantee matching, as resistors naturally drift, and will always have temperature differences, thus changing the value too. Even the PCB tracks would present a problem - 0.000003% for a 1k ohm resistor is 30 thousandth of an ohm - you can't even etch copper to that accuracy! This of course ignores the switching components which is impossible to match at this kind of level, let alone the problems of getting the timing to be accurate enough. For 0.000003% accuracy, with a 16FS R2R DAC, you would need a timing accuracy of 40 femto S applied to the clock, clock tree, and all of the switching components together. To give you a scale of the problem, the best silicon device I worked on had 4,000 fS accuracy from switching element to switching element, and this used FETs that were very much faster than the FET that has a low enough RDS on for a R2R DAC. Indeed, you would need to place all components to a 6 micron accuracy so they all had the same delay. This of course ignores the fact that 16FS from a timing of transients accuracy is not good enough either.
This gives you a tiny flavor of the challenges that one has in designing R2R DAC's, and why it would be impossible to match Dave's performance with them."
Regarding Delta Sigma / silicon DACs, here is Rob's response to someone claiming ESS chips are more complex and therefore, more capable than the Hugo:
"There is no way the Hugo is a lot more complex than the top end ESS chips. My guess is that it is less complex. Luckily, it is a hard fact that can be proven..ESS could easily do what the core Hugo chip does, but not the other way around."
"Your implication that Hugo has only 2 or 3 times the taps, or is less complex than conventional DAC's is clearly absurd. Conventional silicon ASIC DAC's generally have a single 12MHz DSP core to do the interpolation filter function, and it has been this way for 30 years. Hugo has 16 cores running at 208 MHz - that's like comparing an Intel i7 against an 8086. So why don't conventional silicon DAC's have more taps? It comes down to 2 things; awareness (they think it does not matter) and cost. Cost is the paramount driver in ASIC DAC's, to re-coup the substantial investment cost it must sell in very large volumes - that means not for the high-end (audiophool is the usual term) market. FPGA's now provides orders of magnitude of more functional complexity than ASIC DAC's - but they are an order of magnitude more expensive, and they need considerable investment in design work and expertise.
The second major disadvantage of ASIC silicon DAC's are the inherent problems that silicon has - noise and innate non-linearity. Resistors are non-linear, capacitors are also non-linear, substrate and electromagnetic noise is a major problem. Now discrete DAC's (the FPGA is only the digital part, the DAC element is done with discrete components) do not suffer from these problems, but getting a discrete DAC to work with excellent performance is not easy. It has taken me 30 years to perfect my discrete DAC's, and I am still making improvements."
Are all FPGA DACs the same? If you ever want to get Rob mad, just suggest that they are. Here one poster suggesting that the Nagra HD, another FPGA DAC and the DAVE do things the same:
"Nagra avoided super clocks, used FPGA, special algorithm for time correction.".........remind you of any other DACs?"
"Sorry, but I don't usually comment on other companies products, but the statement that this reminds you of other DAC's (Dave) really raised my blood pressure.
Lets look at each statement in turn:
"Nagra sees the quantization noise of 16‑bit/44.1kHz digital audio, and the methods used to quell that noise"
There is absolutely no problem with quantization noise, if correctly done, as it adds a fixed unvarying noise. Because it is unvarying, it has no consequence on the brains ability to separate instruments, define placement in the sound stage, determine timbre and transients and so has no effect on musicality (that is the ability to enjoy music emotionally).
The scary thing is the statement "methods use to quell that noise". What on earth does that mean? The dither is part of the recording, and applied in the analogue domain, so it is the actual signal you want to reproduce. Nobody quells that noise, as how do you separate it from the intended signal?
"Crude brickwall filters that block out any noise above 22.1kHz can undermine phase above 10kHz"
A brickwall filter is an FIR symmetric filter and these are guaranteed to be linear phase, so this statement is just plain wrong.
"while conventional oversampling and interpolation methods are a cure that Nagra believes is often worse than the disease."
What are they saying here? NOS? That topology creates enormous timing errors, and that is the complete opposite with what I do.
"Nagra instead concentrated on the goals of getting the extraction and converting of data absolutely right, without resorting to ‘cheating’ (oversampling)."
Again completely diametrically opposed to what I say. The job of a DAC is NOT to reproduce the digital data, but to reproduce the analogue bandwidth limited signal at the point it is sampled in the ADC. To perfectly create the analogue signal in the digital domain requires infinite oversampling. This is also the only way to reduce jitter sensitivity, and eliminate noise floor modulation, an effect for which the brain is extremely sensitive to. You absolutely can't refer to oversampling as cheating, as it gets the signal much closer to the original analogue signal. But of course, if your intention is to create distortion and noise, then by all means refer to oversampling as cheating, as it won't allow you to achieve your goal of more distortion and noise floor modulation.
"Add to that a custom time-correction algorithm, in place of the usual demands for atomic clocks at this grade, to keep this DAC temporally precise"
Confusing two separate and independent things. The interpolation filter algorithm (if that's what they are talking about) can not change the requirements for the master clock, which depends principally upon the DAC topology.
"and the result is the removal of that quantization noise up to so far beyond the audio band, its impact is effectively completely eliminated"
The quantization noise of 16‑bit/44.1kHz digital audio? That won't get touched at all by any such process. Are they talking about the quantization noise from the noise shaper? In which case don't use DSD, oversample at high rates, and run the noise shaper with n bit quantizer also at high rates (like my 2048 FS). DSD at 128 FS creates vast amounts of out of band noise with typically -20dB down at 200 kHz. You can never refer to noise at 200 kHz at -20dB as "its impact is effectively completely eliminated."
And yes, my previous posting recently confirmed that it was impossible to properly reproduce depth using DSD, as the noise shaper resolution is inadequate. Dave has noise shapers with a trillion times more resolution than traditional DSD noise shapers.
So no, this DAC bears absolutely no similarity with my work - it may have FPGA's, and not use atomic clocks, but that's the only similarity."
And my favorite:
"As to other companies?" No not really, I am off on my own trek. There is nobody else walking on this path, which is why I am not bothered about the possibility of giving valuable information away. This stuff is complex and takes many years to get right and generally engineers with advanced technical skills (there are very few) generally don't do listening tests, and the people that do listening tests don't have the skills. For example, I have been bleating on about long tap length WTA filters for 15 years, and the rest of the industry has simply ignored it and yet this is a simple concept proven by maths. I am still amazed nobody else has long tap length filters...."