Measuring DAC Accuracy

Discussion in 'Source Measurements' started by purr1n, Mar 30, 2017.

  1. ultrabike

    ultrabike Measurbator - Admin

    Staff Member Pyrate MZR
    Joined:
    Sep 25, 2015
    Likes Received:
    8,960
    Trophy Points:
    113
    Location:
    Irvine CA
    Yes. That's interesting (-INF). Note also that the MB DACs have 17 levels, and the DS and Hybrid DACs have 18 levels symmetric around -INF.

    Again, very interesting to see the effects of warm up on all these DACs.

    And yes, if I recall correctly, in @atomicbob's previous measurements, the glitch wasn't that much of a problem or as obvious as with these measurements.
     
  2. atomicbob

    atomicbob dScope Yoda

    Pyrate BWC MZR
    Joined:
    Sep 27, 2015
    Likes Received:
    18,652
    Trophy Points:
    113
    Location:
    On planet
    Yes, very interesting. Note the source signal had 17 levels, not 18, with zero included as one of the levels.

    Someday I want to repeat the warm up measurements with a much more controlled data collection. That will have to wait until I have a week I can set aside and plan to multitask carefully. Not easy collecting data regularly on a number of DACs and DDCs over 168 Hrs and assure high quality data obtained. Anyone that has worked with an SRS, AP, dScope, etc. knows how easy it is to have an important parameter change during data collection.
     
  3. ultrabike

    ultrabike Measurbator - Admin

    Staff Member Pyrate MZR
    Joined:
    Sep 25, 2015
    Likes Received:
    8,960
    Trophy Points:
    113
    Location:
    Irvine CA
    Something weird is going on then. Note how the INF/zero level lasts longer in the input signal and the MB outputs, and how it splits evenly for the other DACs.
     
  4. atomicbob

    atomicbob dScope Yoda

    Pyrate BWC MZR
    Joined:
    Sep 27, 2015
    Likes Received:
    18,652
    Trophy Points:
    113
    Location:
    On planet
    Not weird when you think about it. Difference between instrumentation DAC and Comboburrito filter incorporating source samples in the output vs. approximation of D-S DAC and conventional IIR or FIR filters. FOS
     
    Last edited: Apr 4, 2017
  5. ultrabike

    ultrabike Measurbator - Admin

    Staff Member Pyrate MZR
    Joined:
    Sep 25, 2015
    Likes Received:
    8,960
    Trophy Points:
    113
    Location:
    Irvine CA
    I maybe wrong, but I don't think a DS with a conventional FIR or IIR filter can do that.

    That's a non-linear deal there and we are not talking LSBs. So it's not FIRs and IIRs since those are linear (in the distortion sense). Furthermore, a modern audio DS should be able to represent zero.

    Note also that if there is any analog filter on the output of the DAC (which I do hope exists), one may have to forget about incorporating source samples in the output. And then there is noise and settling of the DAC as a function of time/temperature, which will have their hand also in distorting the source samples in the output.
     
    Last edited: Apr 4, 2017
  6. atomicbob

    atomicbob dScope Yoda

    Pyrate BWC MZR
    Joined:
    Sep 27, 2015
    Likes Received:
    18,652
    Trophy Points:
    113
    Location:
    On planet
    A few more 200 Hz -72 dBFS triangle measurements:

    Concero HD bit perfect mode
    20170404 Concero HD triangle 200 Hz  -72 dBFS level acc - 16 bit pad 24 CEP - bit perfect.png

    Concero HD IIR upsampling mode
    20170404 Concero HD triangle 200 Hz  -72 dBFS level acc - 16 bit pad 24 CEP - IIR up smpl.png

    Concero HD Apodizing filter upsampling mode
    20170404 Concero HD triangle 200 Hz  -72 dBFS level acc - 16 bit pad 24 CEP - Apodize up.png
     
  7. ultrabike

    ultrabike Measurbator - Admin

    Staff Member Pyrate MZR
    Joined:
    Sep 25, 2015
    Likes Received:
    8,960
    Trophy Points:
    113
    Location:
    Irvine CA
    Those look more reasonable. Maybe the other DACs do some weird bit mapping or something (which would be surprising). To be honest, I don't know. All I can say is that it doesn't make sense to me that it's the DS or a linear filter in the chain.

    As far as the IIR upsampling mode, the IIR deal affects low frequencies if it's a high pass (DC block). The reconstruction or upsampling filter are low pass, and would affect higher frequencies depending on how close their cut off is to the signal band. It is possible that the IIR used above is phase benign, such as a Bessel or similar.
     
    Last edited: Apr 4, 2017
  8. atomicbob

    atomicbob dScope Yoda

    Pyrate BWC MZR
    Joined:
    Sep 27, 2015
    Likes Received:
    18,652
    Trophy Points:
    113
    Location:
    On planet
    General message. Anyone perusing this particular measurement topic should be aware that this is but one of a multitude of attributes. Do not take this particular experimental assessment as the only guide for ranking DAC performance. One might consider the Concero HD above to provide better performance in this one attribute. I hadn't listened to the Concero HD in several years so I substituted it for the Modi MB in my The Protester system. Listening session lasted less than 5 minutes before I remembered why I hadn't listened to the Concero HD in several years. The edginess had me subconsciously irritated and a sense of tightening in the spine. Modi MB back and all was well once again.
     
    Last edited: Apr 5, 2017
  9. Clemmaster

    Clemmaster Friend

    Pyrate Contributor
    Joined:
    Sep 28, 2015
    Likes Received:
    3,268
    Trophy Points:
    113
    We became addicted to these 0 crossing glitches :p
     
  10. Scott Kramer

    Scott Kramer Friend

    Pyrate
    Joined:
    May 3, 2016
    Likes Received:
    1,446
    Trophy Points:
    93
     
  11. purr1n

    purr1n Desire for betterer is endless.

    Staff Member Pyrate BWC
    Joined:
    Sep 24, 2015
    Likes Received:
    89,772
    Trophy Points:
    113
    Location:
    Padre Island CC TX
    Here are another set of visualizations using the same data. Again, same deal. Codes go from -32 to 0 to 32 and back down. Sampling rate is 96kHz. These are differential error plots (difference from one point to the next point). 100 triangle samples were overlaid with each other. Brighter dots mean higher concentration of values. There is a shift at tips of the triangle waves, presumably because of aliasing due to sampling and change of direction. Graphs are scaled accordingly to min / max of the recorded values for each DAC.

    Laptop Out
    laptop DNL.png

    Modi 2 DS
    Modi 2 DNL.png

    AGD S19
    AGD S19 DNL.png

    Gungnir Multibit
    Gumby DNL.png

    Modi Multibit
    Mimby 2 DNL.png

    Note that the glitch is not there on Gungnir Multibit. The error transition in the Gungnir Multibit "DNL" plot above is at the tips of the triangle waves. Did Schiit fix this a while back, at least for Gungnir Multibit? I recall someone saying on CA in a post that this might have been fixed for Yggdrasil a while ago.

    BTW, based on my experience with prototypes way back, I don't think the glitch makes a difference subjectively. Except for that Mani dude on CA who claims he can hear it, but most likely he just doesn't like the sound of the Yggdrasil. And wouldn't it be funny if Schiit had already fixed the zero crossing glitch as a rolling change 12 months ago, and that Mani dude claims he could still hear it. Hahahaha.


    BTW, my Gungnir Multibit I got a LONG while back.
     
    Last edited: Apr 5, 2017
  12. frenchbat

    frenchbat Almost "Made"

    Joined:
    Sep 26, 2015
    Likes Received:
    1,369
    Trophy Points:
    93
    I would expect stereophile's sample to be a current prod unit, don't you think ?
     
  13. purr1n

    purr1n Desire for betterer is endless.

    Staff Member Pyrate BWC
    Joined:
    Sep 24, 2015
    Likes Received:
    89,772
    Trophy Points:
    113
    Location:
    Padre Island CC TX
    Hey, don't quote me on the Yggdrasil. I never offered any solid assertions on the Yggdrasil, only on the Gungnir Multibit which I have had for a while. In the meantime, I'd refer you to that post I mentioned on Computer Audiophile.

    I have no idea if the zero glitch on the Yggdrasil was fixed or not, and if it was, how long ago it was, or even what constitutes a "current" production model. JA's review was published in January, and who knows how long the Yggdrasil was sitting on his bench, or how long Stereophile's writing, editing, and publishing process takes. I know that Schiit makes rolling changes with their gear, seemingly with every other batch - noting the revision numbers on their PCBs.

    All I can say is that my Gungnir Multibit, which I obtained a long time ago, does not exhibit any zero crossing glitch behavior; and way way too much has been made of this glitch. The glitchy Yggdrasil is still better than a non-glitchy Gungnir Multibit. Seriously, the only reason to fix this glitch is to appease Stereophile and John Atkinson, because we can't even see this glitch in a -90db 16-bit sine wave - only in a 24-bit sine; and the 16-bit original CD of Paul Simon's Graceland sounds way better than the HD Tracks 24-bit 192kHz version.
     
    Last edited: Apr 5, 2017
  14. Clemmaster

    Clemmaster Friend

    Pyrate Contributor
    Joined:
    Sep 28, 2015
    Likes Received:
    3,268
    Trophy Points:
    113
    I remember Baldr was talking about a "de-glitching" circuit, back in the genesis of the Yggdrasil.
    Maybe it's not present on the Bifrost Multibit and Modi Multibit (cost / space contraint)?
    Maybe it varies with DAC chips?
    Maybe I should go back to work?
     
  15. purr1n

    purr1n Desire for betterer is endless.

    Staff Member Pyrate BWC
    Joined:
    Sep 24, 2015
    Likes Received:
    89,772
    Trophy Points:
    113
    Location:
    Padre Island CC TX
    OK, another set of visualizations. This time, I spread out the Y-axis a bit and put vertical markers with each unit signifying 1 LSB of 16 bits. Again, this just looks at the accuracy of the last 6 bits... sort of because of aliasing. 100 triangle waves with 16-bit codes -32 to 32 are processed unless indicated otherwise.

    Here's Modi Multibit. The zero crossing glitch is easily seen. The differential error from one code to the next isn't quite perfect, but still, well within 16 bits. In fact, we could split the 1 LSB lines two times or more, and the DNL plot would still be between these lines. This is more like 18 bits accuracy. Also the Modi Multibit hits almost exactly the same values time and time again with each successive waveform.
    Mimby DNL.png

    Here's Gungnir Multibit, SE output from before. @ultrabike speculated that the fuzzy imprecise stuff might be noise. The good thing is that the grouping of values tends to clump together. Most values under one standard deviation each way look like they would easily fall within 18 bits.
    Gumby SE DNL.png

    Here's one up and down cycle of the Gungnir Multibit SE triangle wave output. Looks like there is a regular low-level high frequency signal, around 19kHz.* Even then, accuracy looks to be about 18 bits. I wonder what this and the above plot would look like if we put the numbers through a low-pass filter?
    Gumby SE DNL 1.png

    *It's actually 24.38kHz.
     
    Last edited: Apr 5, 2017
  16. purr1n

    purr1n Desire for betterer is endless.

    Staff Member Pyrate BWC
    Joined:
    Sep 24, 2015
    Likes Received:
    89,772
    Trophy Points:
    113
    Location:
    Padre Island CC TX
    Now this is interesting. Here are the results for Gungnir Multibit's BALANCED output.
    Gumby BAL DNL.png
    Yeah, I did the recording several times and even went back to SE to check. Holy cow this is good.


    In closing I leave you with the AGD S19. Here we can see the distribution of errors is fairly wide. I am figuring a standard deviation of 0.80 LSB one way. Some individual errors seem to be way off, almost 2 LSB from zero (super faint teal dots). I need to work on the visualizations to normalize the color intensity.
    AGD S19 DNL.png

    Here is a single up down cycle for the AGD S19
    AGD S19 DNL 1.png
     
    Last edited: Apr 5, 2017
  17. SSL

    SSL Friend

    Pyrate
    Joined:
    Nov 12, 2015
    Likes Received:
    1,239
    Trophy Points:
    93
    As I recall, @atomicbob 's measurements of Gungnir Multibit showed a similar discrepancy between SE and BAL output.
     
  18. purr1n

    purr1n Desire for betterer is endless.

    Staff Member Pyrate BWC
    Joined:
    Sep 24, 2015
    Likes Received:
    89,772
    Trophy Points:
    113
    Location:
    Padre Island CC TX
    Just checked the FFT. There was a spike at 24.38kHz, a multiple of 96kHz. The spike isn't there with BAL outputs. The signal isn't even. It comes and goes sporadically. This is why we don't get consistent patterns with each successive triangle wave.

    EDIT: Note correction.
     
    Last edited: Apr 5, 2017
  19. ultrabike

    ultrabike Measurbator - Admin

    Staff Member Pyrate MZR
    Joined:
    Sep 25, 2015
    Likes Received:
    8,960
    Trophy Points:
    113
    Location:
    Irvine CA
    I think usually if the noise in the SE output that is not present in the BAL output, chances are there issue is noise. Usually random, but it could be all kinds of things. It looks pretty good in BAL mode.

    I dunno about the Gungnir Multibit, but I can say that the BAL out of my 2i2 is very clean and nice. Not so the SE output.

    The Gungnir Multibit does look very good in BAL mode.
     
  20. Garns

    Garns Friend

    Pyrate
    Joined:
    Jul 9, 2016
    Likes Received:
    2,483
    Trophy Points:
    93
    Location:
    Sydney, AUS
    What do I need to do measurements like this? Is my 24/96 ADC (Lavry AD-10) good enough or would I need more specialised equipment?
     

Share This Page