Soekris DAM1021 R2R DAC Measurements

[rebbiputzmaker]

I did that test and from what I recall, I could hear -90 dB test tone.

Ok cool.

 

[Cosmik]

I do understand the mental imagery side of this: one type of DAC relies on precision components and is influenced by construction quality, and it has a certain appeal in its simplicity and its literal representation of a digital number. It also plays the 'classic', retro card.

The other type of DAC is much less obvious in its operation, is very much software based, can be 'printed' for $3, and doesn't rely on precision of construction to work properly. It is much less 'tangible' and therefore doesn't create such a strong image in the mind.

By knowing a little about each type, and reading vivid descriptions penned by an enthusiast, the sound of each can be (and is) conjured up purely based on imagination. Unfortunately, as far as I can tell, the $3 version does, in fact, work brilliantly despite 'the optics'.

And it feeds into one of my theories about the state of audio: With digital technology and DSP, audio has come perilously close to reaching practical perfection. Everyone involved (including the customers) realises that the only way they can maintain a sense of progress, and continue having fun in sheds and garages, is to go back to previous imperfectible technologies and spend eternity trying to perfect them.

 

[Frank Dernie]

This is not the place to start long discussion about what you can hear or not, just want to point out there is a big difference between measuring using sine tones and playing actual music which are very complex

I have seen this comment quite often on other fora but with no (convincingly scientifically valid for me) explanation, just the implication of "this must be obvious", which to somebody with no technical education it may well seem to be, but everything I learned told me that with a linear system tones were simply additive, so in fact testing with simple tones and testing with a mixture of multiple tones are surely exactly the same???
Sure if the device is not linear there will be intermodulation products but DACs, in general are actually pretty linear, so do you have an explanation as to why tone testing is not valid for music, apart from it being obvious music is more complex? I realise it is not always known that music is a sum of mixed tones and I could be over simplifying/wrong?

 

[Rene]

A few random thoughts...

I've not (or only rarely) used linearity plots like that; other tests show the same thing, just a different plot results. As a way to show relative differences among DACs I like them. Note 0.1 dB ~ 1%, so 1% deviation is not an unreasonable criteria, but only about 0.3 bits if you reference to SNR based upon quantization noise. You could argue for almost anything but a consistent reference is good enough for me and I can always read the graphs for more info or to use my own metric. It's a relative comparison so as long as the basis is the same it works for me.

Wouldn't 3dB be 1/2 LSB? That sounds like a good figure of merit when we're >100dB down from full scale.

It is probably worth noting that correctly wiring pin 1 and in fact just using XLRs (balanced) does not guarantee no ground loops or common-mode problems. And transformers don't always exhibit great CMRR plus have plenty of other issues. Neither active nor passive differential designs are perfect, you're always making trades among compromises. Great and awful examples of either type (active or transformer) exist, natch.

Balanced connections do help, though, even if we only get 40 or 50dB at 50 or 60Hz. About the only place we want or need more is in measurement equipment, like the AP or in medical electronics where we're often measuring signals in the nanovolt range.

The Rane note does not really address transformer vs. active differential circuits, or barely, unless I missed it (possible, quick skim). It does address the issues with certain schemes to convert between balanced (differential) and single-ended (unbalanced?) using transformers or active circuits. In those cases the active circuit implementation is critical and it is true rarely provides decent CMRR. In fact that is one thing I have decried for years, the proliferation of various quasi-differential schemes marketed as having the same benefits as fully differential, which includes a number of consumer and pro audio components that do not implement a truly differential transmitter and receiver but throw in a TRS or XLR jack and call it "balanced". Even something as simple as a resistor on the "cold" leg so the impedance looking in is roughly the same as the "hot" leg is marketed as balanced based on the "balanced" impedance but in practice CMRR is not much better than a single-ended connection (i.e. almost nil, often as low as 6 - 20 dB).

I've worked designing professional audio equipment for ~50 years and have used resistive balancing on outputs on many a product design. There is nothing to fault there. CMRR is dependent on having equal resistance/impedance on both legs of the balanced transmission link, the burden falling on the input side where the higher resistance/impedance lies. It does not depend on both sides of the transmitter having equal signal voltages; that only buys you higher output voltage capability.

 

[Rene]

Sure if the device is not linear there will be intermodulation products but DACs, in general are actually pretty linear, so do you have an explanation as to why tone testing is not valid for music, apart from it being obvious music is more complex? I realise it is not always known that music is a sum of mixed tones and I could be over simplifying/wrong?

Music is not steady-state. Under changing signal conditions the states of sigma-delta modulators are often undefined. Can you offer a repeatable, non steady-state test methodology besides music?

 

[RayDunzl]

Under changing signal conditions the states of sigma-delta modulators are often undefined.

Isn't a sine wave a "changing signal condition"?

 

[Rene]

Isn't a sine wave a "changing signal condition"?

No, it's steady-state. It has to be for the analyzer to analyze it.
(Sorry, Dept. of Redundancy Dept.)

 

[RayDunzl]

No, it's steady-state. It has to be for the analyzer to analyze it.

Isn't a swept sine wave a "changing signal condition"?

 

[Superdad]

Isn't a swept sine wave a "changing signal condition"?

No. No transients there. That is the whole point. Humans recognize sounds as real or not based largely on the attack (it's an evolutionary thing to keep from being eaten by a mountain lion). Looking at the rise time of a square wave will at least tell a little more.

 

[Frank Dernie]

Music is not steady-state. Under changing signal conditions the states of sigma-delta modulators are often undefined. Can you offer a repeatable, non steady-state test methodology besides music?

So you are concerned that a Delta sigma modulators may not be linear in an unknown and unpredictable way if the signal isn't steady state?

 

[DonH56]

Wouldn't 3dB be 1/2 LSB? That sounds like a good figure of merit when we're >100dB down from full scale.

For SNR based upon quantization noise, yes, or we could choose 6 dB for a full bit, or whatever. As I said, for me a relative comparison, as long as I know the basis and the data are there if I want more, is OK. FWIWFM, I would have probably picked half a bit as +/-1/2 lsb is a common figure of merit.

Balanced connections do help, though, even if we only get 40 or 50dB at 50 or 60Hz. About the only place we want or need more is in measurement equipment, like the AP or in medical electronics where we're often measuring signals in the nanovolt range.

Agreed, and I have said many places that I think balanced connections are overplayed in the consumer realm. For breaking a ground loop they are a real help, but most equipment does just fine rejecting other noise even running single-ended. There are always exceptions, natch, and everyone feels that are The One. I am not immune by any means.

For that matter, RF/mW/mmW systems work down to the nV (and sometimes fV) range and virtually all of them are single-ended, so it can be done. I will say I am glad my current job does not require fV measurements; that was a royal PITA. Somehow audio has acquired this mystique of ears being better than any circuit we can design or measure.

I've worked designing professional audio equipment for ~50 years and have used resistive balancing on outputs on many a product design. There is nothing to fault there. CMRR is dependent on having equal resistance/impedance on both legs of the balanced transmission link, the burden falling on the input side where the higher resistance/impedance lies. It does not depend on both sides of the transmitter having equal signal voltages; that only buys you higher output voltage capability.

I have not worked in audio except as a hobbyist (aside from a decade or so as a tech, and some years running sound boards with a little studio work thrown in), so do not have your experience, but am not sure that invalidates my argument. I did not say anything about signal levels; I am not sure how they are relevant to this? CMRR depends upon the circuits at both ends, and I said "transmitter and receiver" in my post. I guess I am not understanding my problem, not an unusual thing. I have measured very poor CMRR in some pro and consumer gear when the link (connection) was not adequately balanced, resistively or otherwise, at both ends but since I don't have those measurements handy I'll let this one drop as IMO.

Music is not steady-state. Under changing signal conditions the states of sigma-delta modulators are often undefined. Can you offer a repeatable, non steady-state test methodology besides music?

This is one of the big problems in device characterization and not just measuring delta-sigma converters. Two-tone IMD, with fairly closely spaced tones, swept over frequency and amplitude provides more info IME. Which I'll reiterate is (mostly) not audio but the physics and basic principles do not change. I'd like to see a lot more multitone testing performed as that is more likely to show some of the squirrely things some circuits (not just delta-sigma converters) can do when stressed. The NPR is another nasty one. The latter few tests are among what we did on RF delta-sigma designs to look for things like noise modulation, tones, and other problems. Delta-sigmas do have their quirks and can do strange things.

Another useful test is to overdrive the input and output to see how it performs. Some circuits behave... poorly. The days of op-amp output inversion are pretty much behind us now but I have seen, and for that matter Amir has measured, some disturbing behavior from some amplifiers when clipped, and some input stages look really ugly when overdriven.

The old TIM test, using a LF square wave with a HF sine wave on top, was also revealing at times.

The problem with most of these tests is having the time and equipment to do them, with time being the limiting factor in some companies, and then explaining them to folk with no technical background. I had the luck of spending a lot of time doing R&D during my career so got to go down more rabbit holes than a lot of my friends in industry. And still get to do a lot of it now though my position is more test than design (but still GHz stuff, not audio).

 

[Cosmik]

Can DS modulators not be exercised and evaluated entirely in software simulations?

 

[DonH56]

Define "entirely". Most of it can be, at least at a basic level, from high-level sims to mixed-mode simulations with behavioral models to transistor level, but it's a big task and there are always things in the real world not in the sims. And despite the speed of modern computers simulation time is still a problem; setting up multiple signals with long record lengths and hundreds of thousands of transistors or more is very time-consuming. Dealing with long time constants, multiple interacting loops, and signals with very low-frequency content often add up to impractical simulation space. In my world running a behavioral model like an IBIS-AMI receiver can take an hour or two to get just a few microseconds of data. And that is to gain a tiny snippet of steady-state'ish information. Running a transistor-level simulation to capture the same time period is not practical (too long to run even if you could get the simulator to converge, another nightmare with large circuits).

Plus the sims are only as good as the models and the test conditions. I sim'd the snot out of the DS converters I was designing for months and still barely scratched the surface. A few seconds in the lab obviates most sims, but in the world of IC design, simulation is required to get as good as you can before before a few million on mask sets. We will spend a year or more simulating everything we can in the design and there can still be problems with the chips, leading to mask spins and so forth...

FWIWFM - Don

 

[Cosmik]

Define "entirely". Most of it can be, at least at a basic level, from high-level sims to mixed-mode simulations with behavioral models to transistor level, but it's a big task and there are always things in the real world not in the sims. And despite the speed of modern computers simulation time is still a problem; setting up multiple signals with long record lengths and hundreds of thousands of transistors or more is very time-consuming. Dealing with long time constants, multiple interacting loops, and signals with very low-frequency content often add up to impractical simulation space. In my world running a behavioral model like an IBIS-AMI receiver can take an hour or two to get just a few microseconds of data. And that is to gain a tiny snippet of steady-state'ish information. Running a transistor-level simulation to capture the same time period is not practical (too long to run even if you could get the simulator to converge, another nightmare with large circuits).

Plus the sims are only as good as the models and the test conditions. I sim'd the snot out of the DS converters I was designing for months and still barely scratched the surface. A few seconds in the lab obviates most sims, but in the world of IC design, simulation is required to get as good as you can before before a few million on mask sets. We will spend a year or more simulating everything we can in the design and there can still be problems with the chips, leading to mask spins and so forth...

FWIWFM - Don

Isn't the question one of 'logic', though, rather than transistor level simulations? I would have thought that it would be possible to simulate at a higher level than transistors in order to establish the principle of operation - in which case it would be very fast. You could simulate all the transients you want; use real music, etc. and establish pretty much what the DAC would be producing without having to resort to measurements or SPICE simulations..?

 

[RayDunzl]

Humans recognize sounds as real or not based largely on the attack (it's an evolutionary thing to keep from being eaten by a mountain lion). Looking at the rise time of a square wave will at least tell a little more.

One day when I was wondering about the "accuracy" of measurements, in the sense of Step response (which would represent the rise time of a square wave) I compared the recording (in-room) with the analysis (REW) and came up with this surprising (to me) congruity.

Top: REW analysis of swept sine test signal played by speakers displayed as calculated Step Response.
Bottom: In-Room Recording 48/24 10Hz square using Audacity:

Thinking that was, if nothing else, interesting, I created and recorded the playback of a single full-scale sample in-room, and compared that to the Impulse Response calculated from the in-room REW swept-sine test signal:

Top, REW swept sine analysis displayed as Impulse Response
Bottom, recorded waveform from speakers at listening position.

About that time I decided that swept-tone analysis tools could do a pretty good job of figuring out and displaying the Impulse and Step response of a system.

So, in my Amateur Capacity, I pretty much eliminated that concern from further thought.

Of course, I suppose this doesn't tell us anything about "real".

I was unable to create what looked like either a leading edge square nor a very nice impulse using the physical sound-making materials I had on hand, judging by the recordings of those sounds I examined.

They did sound "real" and gave interesting waveforms, though.

The recordings sounded pretty "real" though headphones, the double-room effect spoiled the speaker renditions of the played back speaker renditions to some degree.

 

[DonH56]

Isn't the question one of 'logic', though, rather than transistor level simulations? I would have thought that it would be possible to simulate at a higher level than transistors in order to establish the principle of operation - in which case it would be very fast. You could simulate all the transients you want; use real music, etc. and establish pretty much what the DAC would be producing without having to resort to measurements or SPICE simulations..?

The logic you can simulate, yes, but even that is tricky when you have long filters and feedback that makes the results signal-dependent. Defining a good set of signals and adequate range of test cases is non-trivial. But yes, principle of operation is not too bad to simulate, but that is usually just the starting point. The devil is in the details, and even at the logic (e.g. RTL) level the simulations can be very long and output files huge. Simulating a few seconds of music, even at the logic level, can take hours though I do not have much experience with that and at my workplace the projects involve millions of gates so probably more complex than a delta-sigma DAC. But you still have to model the delays and such to ensure timing closure and those little "extras" add a lot of time to the simulations.

That said, the simulations of 1k to 1M points used in the DS DAC simulations I presented in the articles on WBF (and copied here, thanks Amir) only took from a few seconds to a few minutes so for basic analysis it's quick. The architectures were very simple, however, low-order modulators and I did not use long and fancy FIR filters like a real design would use, and that was Mathcad/Matlab so not even at the gate level but one level up.

ADCs are trickier since a lot of analog goes on in the loop and determining stability from just an RTL simulation is tricky.

And of course RTL does not tell you how the analog input and output filters and buffers behave. That is why I asked for your definition of "entirely". In a six-month design effort, I probably spent a week or two on the basic "logic" simulations; the rest of the time was designing the actual circuits and simulating them (and their layouts) with more realistic "real-world" models and parasitics. But, I may be wrong, especially for audio, since the rates are low enough that timing and settling are less a concern? I tend to doubt it; my LF and audio design experience (even my RF data converters had to work down to DC -- IQ systems often require that) says designing an 8- to 14-bit system at 1 to 10+ GS/s and a 200 kS/s, 24-bit converter has many of the same challenges. Bigger devices are needed for lower noise and better matching, long settlers are hard to identify (and often are not simulated, like thermal tails), and things like flicker and popcorn noise are not always adequately modeled (if at all). Quasi-saturation and sub-threshold effects, leakage, the list is endless of things that can and do go wrong at the transistor level.

So at one level you can simulate the logic fairly quickly, but it can take a while to really wring it out, and at the end there is still often a big gap from that to how the circuit performs on the bench (or in a plane, or space...)

But, this has made me think it would be interesting to dig up some of my old work and revisit it to look for some of the things that the logic ought to show... Most of that was proprietary, alas, so I'd have to recreate a bunch of even the basic simulation setups. I'll probably just go flip on the stereo instead since it's Saturday afternoon and I have to work tomorrow.

 

[RayDunzl]

"Real" sounds. Complex? Ummm, yeah...

Recording (top of each), vs in-room playback recording at listening position with AcourateDRC applied.

Playback: Disapproved by Trained Listeners Martin Logans with the flanking you-don't-want'-none-of-these-Dewey cheezewoofers.

"Reflection Free" (except for the top of the couch under the mic) - about 7 milliseconds of sound:

Hand-clap

Striking two Home Depot wooden paint stir sticks (zoomed a little)

Not perfect, doesn't prove anything, looks pretty good though.

Even I'm surprised. The filter applied at the miniDSP is months old...

 

[Cosmik]

The logic you can simulate, yes, but even that is tricky when you have long filters and feedback that makes the results signal-dependent. Defining a good set of signals and adequate range of test cases is non-trivial. But yes, principle of operation is not too bad to simulate, but that is usually just the starting point. The devil is in the details, and even at the logic (e.g. RTL) level the simulations can be very long and output files huge. Simulating a few seconds of music, even at the logic level, can take hours though I do not have much experience with that and at my workplace the projects involve millions of gates so probably more complex than a delta-sigma DAC. But you still have to model the delays and such to ensure timing closure and those little "extras" add a lot of time to the simulations.

That said, the simulations of 1k to 1M points used in the DS DAC simulations I presented in the articles on WBF (and copied here, thanks Amir) only took from a few seconds to a few minutes so for basic analysis it's quick. The architectures were very simple, however, low-order modulators and I did not use long and fancy FIR filters like a real design would use, and that was Mathcad/Matlab so not even at the gate level but one level up.

ADCs are trickier since a lot of analog goes on in the loop and determining stability from just an RTL simulation is tricky.

And of course RTL does not tell you how the analog input and output filters and buffers behave. That is why I asked for your definition of "entirely". In a six-month design effort, I probably spent a week or two on the basic "logic" simulations; the rest of the time was designing the actual circuits and simulating them (and their layouts) with more realistic "real-world" models and parasitics. But, I may be wrong, especially for audio, since the rates are low enough that timing and settling are less a concern? I tend to doubt it; my LF and audio design experience (even my RF data converters had to work down to DC -- IQ systems often require that) says designing an 8- to 14-bit system at 1 to 10+ GS/s and a 200 kS/s, 24-bit converter has many of the same challenges. Bigger devices are needed for lower noise and better matching, long settlers are hard to identify (and often are not simulated, like thermal tails), and things like flicker and popcorn noise are not always adequately modeled (if at all). Quasi-saturation and sub-threshold effects, leakage, the list is endless of things that can and do go wrong at the transistor level.

So at one level you can simulate the logic fairly quickly, but it can take a while to really wring it out, and at the end there is still often a big gap from that to how the circuit performs on the bench (or in a plane, or space...)

But, this has made me think it would be interesting to dig up some of my old work and revisit it to look for some of the things that the logic ought to show... Most of that was proprietary, alas, so I'd have to recreate a bunch of even the basic simulation setups. I'll probably just go flip on the stereo instead since it's Saturday afternoon and I have to work tomorrow.

I'm sure there's all kinds of real world analogue issues to deal with, but the implication earlier from someone was that the actual 'algorithm' is suspect and that steady state signals (even mixtures of tones) wouldn't reveal the problems. My expectation is that real music could be fed into the algorithm and the simulated analogue output (assuming ideal filter) examined for absolute deviations from the near-perfectly reconstructed waveform, calculated with 64 bit arithmetic. If we fed in a whole CD's worth and found that the absolute instantaneous output never deviated more than steady state tests suggested it should, we might feel less helpless about it. If it revealed 'nasties' then the heritage DAC people are right.

I am sure the manufacturers do this kind of thing, routinely, of course.

 

[DonH56]

You can feed a steady-state input into a DS DAC and see tones at the output using Matlab. I have done that and in fact is a few pages in one of the college lectures I gave years back(*). But, the more complex the DAC, the harder it is to find those tones, and the trickier to simulate. The math falls apart and you don't get closed-form solutions anymore so have to start delving in to high-level functions and statistics that make my brain hurt. I be a hairy-knuckled engineer at the end of the day...

Feeding a whole CDs worth would take a long, long time to simulate. What I have done is feed in noise and look for spurs. Of course you have to generate a noise source that does not have patterns, which means very long registers (even in simulation) and lots of data points, so it still takes a long time. Let's say you have 10 seconds at CD rate so 442,000 samples (for one channel) and perhaps 20,000 taps so 8,840,000,000 (8.84e9) points. Square to estimate the number of flops needed in the solver's matrix so roughly 78.1e18 flops. An Intel i7 does around 70 Gflops/s (per Tom's Hardware site, after a quick search) so that will take about 1e9 seconds (310k hours, about 13k days). Plenty of time for a coffee break. Simulators do extensive data reduction, sparse matrix theory and all that jazz I don't claim to really understand, so it would be somewhat faster, but you get the idea... There are special techniques used to get large (single to a few order-of-magnitude) reductions but it would still be a long sim, I think. I'm an analog guy so take this with a block of salt; chances are good we have experts here who can provide a much better estimate.

Much shorter sims can show problems including noise modulation and such but it can still take hours to provide the results for <1 s of real-time data.

(*) Edit edit -- I was going to pull a pix, but the presentation was done in Ami-Pro and apparently MS Word took out the translator so I can't open the file any more. I had forgotten... I actually bought a program that was supposed to do it, including convertering all the figures, but it failed miserably so I gave up. The hard copy, if I still have it, is someplace in the black hole with all the other stuff stored away when e finished our basement years ago.

 

[March Audio]

Now I do question the idea delta sigma DACs are harsh from non-harmonic distortion. What's the basis for that? What do you think is going on? There are certainly ways to detect non-harmonic spurious signals with various test signals.

If this is really the case shouldnt the 1021 sound very harsh due to all the non harmonic spuria noted in in the IMD test?

I think the point that has to be remembered by some is that whilst they are welcome to do some "hand waving" about this and that, they need to providecsome evidence to demonstrate the statements. Its no good just stating "delta sigma sounds harsh and has unknown unmeasureable problems". Nor is saying "just listen".