Friday 1 July 2022

Hi-Res THD(+N) vs. Output Level Measurements (ESS "HyperStream" vs. AKM vs. TI/Burr-Brown). And a bonus R-2R!

Notice last time as I ended off the post, I showed what I think is an interesting "high resolution" graph of THD(+N) vs. Output Level for the Topping D10 Balanced which uses the ESS ES9038Q2M chip. This was spurred on after some discussions on glitches and anomalies one might see due to the "HyperStream" architecture of the ESS chip.

These days, other than the occasional fully multibit or discrete R-2R DACs, the vast majority of what we're using are multibit/multilevel sigma-delta devices. This includes the brands I have listed in the upper graphic; Asahi Kasei Microdevices (AKM), ESS Technology, and Burr-Brown (which was acquired by Texas Instruments in 2000). We'll also talk about the Philips later. ;-)

Today, let's have a look at "high res" THD(+N) vs. Output graphs (XLR output where possible to keep noise as low as possible) comparing different DACs from these companies...


ESS Technologies:

Let's start with some DACs based on ESS' converter technology with their "HyperStream" modulator.

Here is what I published last time (with slight adjustment to distortion units) of the Topping D10 Balanced (ES9038Q2M), examining -25dBFS to 0dBFS using REW's "stepped sine" measurement system at increments of 0.1dB:

Measured values in the legend correspond to 0dBFS.

Below -25dBFS, most high quality DACs will have harmonic distortion low enough to be buried in the noise floor so the pattern isn't as noticeable or interesting. As you can see, we have quite a patterned level of harmonic distortions especially the odd-order 3rd, 5th, 7th, and 9th showing a sinusoidal pattern which gets more compressed closer to 0dBFS. If we were to watch the harmonic distortions in realtime, we would see an undulating increase and decrease in the odd harmonics as output level changes while the even harmonics follow a less predictable pattern at lower levels.

Let's have a look at another ES9038Q2M device to double check - here's the S.M.S.L. DO100 I reviewed recently:

Values in the legend correspond to 0dBFS.

Again, we're seeing a similar pattern. While overall the noise level and THD+N are better at 0dBFS, we're seeing that sinusoidal pattern of odd-order harmonics play out into 0dBFS. As with the Topping D10B, the 3rd harmonic tends to be the strongest with significantly lower 2nd harmonic.

Hmmm, how about a device using the ES9038PRO chip? Here's my Oppo UDP-205:

The Oppo UDP-205, while discontinued since 2018 remains a "beast" of a 4K Ultra-Blu-Ray, SACD, and pretty well everything else including full MQA decoding player (the main exception being HDCD). It has two ES9038Pro chips inside, one for the stereo output, the other for multichannel. Since I have not measured this device for years, I thought it would be fun to have a look at the 1kHz THD+N using the E1DA Cosmos APU + ADC setup first:

THD+N about -119dB - very good. Notice that we see a number of side-bands around the 1kHz signal which are likely power-supply related as they are even multiples of 60Hz, the mains hum frequency here in Canada.

So, what about that hi-res THD(+N) vs. Output Level graph?

Values in the legend correspond to 0dBFS.

Yup, like the ES9038Q2M devices above, the ES9038PRO also has a pattern of fluctuating odd-order harmonics when examined in detail although in the Oppo design, perhaps not as clearly defined below -15dBFS.

Let's look at an ESS chip that's a step down from the ES9038 line, using an earlier version of the "HyperStream" ("HyperStream I" as it were). Here's the S.M.S.L. M100 Mk II powered internally by the ES9018Q2C previously reviewed:

Still one of the cutest looking DACs I think!



While the noise level is higher (to be expected with RCA output), the 2nd harmonic rises substantially with increasing output level, and THD+N is not great at -92dB at 0dBFS, underneath all that, we also see a similar pattern of regular odd-order harmonic rippling with the ES9018's HyperStream I.

Let's now have a look at devices not using the ESS DAC chips...

Asahi Kasei Microdevices (AKM):

Until the AKM factory fire in October 2020, these DAC chips were very popular among hi-fi devices. So while there appears to be some good news on the factory repairs, AKM-based DACs probably will still remain rare or in limited supply through 2022. It is nice to see the recent announcement for the AK4493S and AK4499EX/4191EQ though.

I have a couple of RME devices here to look at based on AKM's 256x oversampling, internal switched capacitor filter, and "OSR Doubler" SDM architecture.

The RME ADI-2 Pro FS R Black Edition is based on the AK4493 "Velvet Sound" chip:

Values in the legend correspond to 0dBFS.

Notice the significant difference compared to the ESS Tech DACs. The AKM DAC has a much more predictable pattern of harmonic distortion with the 2nd order predominating until right around -3dBFS after which the 3rd becomes the highest. Higher order harmonics are nicely suppressed throughout.

Let's double check then with another product. Here is the dual AK4493 Topping DX3 Pro V2 previously examined (note that this is not the current Topping DX3 Pro+ which has switched over to the ES9038Q2M):





While distortion amounts are different from the RME above, and noise overall higher than XLR output, we're still seeing a relatively smooth harmonic pattern across the output levels.

Let's look now at the lower model AK4490 "Verita" chip which is found in the older RME ADI-2 Pro FS:

Values in the legend correspond to 0dBFS.

Very nice -110dB THD+N (XLR out). The "family resemblance" with the newer RME ADI-2 Pro FS R BE is clear with 2nd order predominating until just below -2dBFS then the 3rd harmonic rises above.

Overall, the more predictable "cascade" of harmonics with generally decreased higher order amounts in the AKM I think in principle is very nice; probably more "ideal".

TI / Burr-Brown:

Up next, let's have a look at the TEAC UD-501 again with its "Advanced Segment" (hybrid: likely upper 6 bits multibit and lower bits 3rd-order SDM) TI/BB PCM1795 DAC chip inside:


Values in the legend correspond to 0dBFS.

Like the AKM DACs, this is also a smoother pattern than the ESS. Also like the AKM, depending on the output level, 2nd and 3rd harmonics predominate with a transition taking place around -7.5dBFS with the 3rd harmonic dropping below the 2nd. Higher order harmonics are generally well suppressed.

BONUS! Philips TDA1543 x 4 R-2R NOS DAC:

Okay, thought I'd throw in a bonus "vintage" DAC here for fun ;-).

That's my old "MUSE" Philips TDA1543-based DAC with 4 chips running parallel which I've used over the years to discuss NOS and by extension, R-2R ladder DACs (BTW, nice concise review article on the technical aspects here). I believe there are still DACs in the audiophile world sold based on these old Philips TDA1543 chips which were first released around 1989.

Quite expensive models like this CAD DAC are sold these days using increasing numbers of 8 and 16-chip configurations. Compared to the modern multi-bit SDM technology, the TDA1543 cannot truly be called "high-resolution" or "high-fidelity" depending of course on one's definition of "fidelity" regardless of the hype a company might use. Unless a listener is trained, there is not necessarily close correlation between the fidelity of a device and the subjective "euphonic" quality based on individual preferences.

For my testing, here's what the set-up looks like using the E1DA Cosmos APU and ADC:

Depending on test, source can be the Raspberry Pi or directly from the computer when doing REW stepped sine.

Since this is a purely S/PDIF-input device, I'll using the Topping D10s as a USB → TosLink S/PDIF converter to feed the DAC (optical galvanic isolation). While it's limited to 16-bit conversion, the DAC can accept 96kHz samplerate without complaint (up to 192kHz coaxial I believe). Notice that I'm using a 12V/6A lithium battery pack which is as noise-free as it gets. A good quality glass TosLink cable in the set-up.

Here's the 0dBFS 1kHz 24/96 pure sine wave FFT to give you an idea of the resolution:


That's obviously not very good with THD+N -64dB ("only" 0.06%) at 0dBFS for a DAC. More fancy designs like those 8/16-chip configurations should be able to improve on the noise level to some extent (it better if it costs more than US$50, the approximate price of this old DAC!).

This is a NOS DAC with stair-stepped, squarish waveforms - here's a zoomed-in look at the 1kHz output at -25dBFS:

Those sharp waveforms will show up as excess odd-order harmonics in the FFT. We can see this when we look at that -25dBFS 1kHz tone above:

While noise higher, notice that the -25dBFS 1kHz sine signal actually has better THD+N of around -70dB compared to 0dBFS thanks to much lower overall harmonic levels, especially the 2nd.

While there's still a fair amount of 2nd harmonic, it's obvious the odd-order distortions predominate well beyond the highlighted 9th harmonic label. Folks who express a preference for pure NOS playback of 16/44.1 material - often they will say it sounds more "real", or more "spacious" - presumably like the effect of these odd-order harmonics (and -3dB roll-off around 20kHz).

Since I noticed quite a bit of distortion even at lower levels, here's the "hi-res" THD(+N) vs. Output Level graph from 0dBFS down to -50dBFS (0.2dB steps):

Measured values in the legend correspond to 0dBFS.

While not particularly pretty, this again doesn't have the ESS DAC's sinusoidal pattern of harmonics. You will notice some interesting trends in the distortion data.

* Odd harmonics are higher across the output levels creating the NOS "stair-stepped" morphology. Notice the very significant dissociation between the amount of odd harmonics congregating higher in level and the even harmonics lower (except for the 2nd).

* The 2nd harmonic cuts through the graph and rises persistently from -32dBFS to become the leading amount by about -9dBFS and above.

* Since each bit represents approximately 6dB of output, notice that we see discontinuities in the distortion presumably corresponding with each level of the R-2R ladder as it transitions along the network. For example, in the THD+N tracing, we see this clearly at -12/-18/-24dBFS. With the eye of faith, we can also see smaller irregularities at -30/-36dBFS as well. Distortion levels are already quite high at -6dBFS so anomalies there would be hard to see. Precision of the current source and resistor values are essential for high-resolution performance and even slight imperfections could lead to such irregularities in R-2R DACs; achieving such precision can of course be very expensive. (The goals of technological advancement with modern SDM DACs include better performance and much lower price.)

I think it's interesting looking at old technology like this and comparing it with how things have improved with modern DACs. While the MUSE TDA1543 is inexpensive, I'd be curious how higher-end devices with improved power supplies (note that I'm already running this on battery power) perform based on these old Philips chips. Anyone seen a measurement for the CAD TDA1543 Mk II for example? Regardless of expense, the limitations of the 16-bit TDA1543 I think are clear and there's something to be said about not throwing money into improving something that wasn't great to begin with. Claims about the TDA1543 providing some kind of "natural" sound as if these decades-old budget NOS parts were ever all that great is a bit strange to me.

Speaking of R-2R, recently we also saw how poorly the Cayin RU6 R-2R DAC performed whether in NOS or OS mode.

Summary:

A fistful of DACs!

I think it's quite clear comparing the DACs that the ESS chips' THD(+N) vs. Output distortion pattern is not like the others (although the Philips TDA1543 has its own character as well).

ESS "HyperStream" modulator-based DACs tend to capture a lot of the spotlight as they can achieve high numbers when we look at single measurements like 0dBFS 1kHz THD+N/SINAD. However, over the years, audiophiles have reported measurable anomalies like the "hump" and low-level noise floor "needles" in certain test signals at various amplitudes on the FFT. This variability in low-level harmonic distortion such as the FFT pattern at -2.9dBFS I used for the DSD 1kHz signal can look quite different from typical 0dBFS measurements.

While these anomalies appear to be very low level, the AKM graphs do look more "natural" at least in that given the choice, I would probably want to see uniformly invariant low distortion from a DAC. As with objective testing in general, the only way to truly judge whether something is audible is with actual blinded listening tests; as you're likely aware, these are very hard to come by in the audiophile hobby and that not everyone accepts blinding being important is IMO a problem. In ESS' 2011 presentation on the HyperStream at RMAF (here's the video), Martin Mallison suggest that "when the final artifact is removed, the listener starts to choose the HyperStream modulator (in blind tests) as being better than the conventional DAC" (~32:15 in the video). In 2011, the highest model product was the ES9018. This then suggests that the fluctuating harmonics are inaudible with trained listeners; regardless, it would certainly be interesting to see the results from such blind listening tests and how they were performed (I'm always curious what music was used).

Recently, ESS has announced the forthcoming ES9039PRO which promises better performance although the specs are not that different from the ES9038PRO: 8 channels, up to 132dB dynamic range per channel, and THD+N -122dB mono (same as ES9038PRO). It is said to use a "quad modulator" however, and they call it the "HyperStream IV".

While these specs numbers are impressive, I don't think it's reasonable to suggest or expect that this new DAC would sound much different from the current ES9038 models (with my belief that decent DACs these days are "perceptibly perfect" and all...). However, I wonder if objectively we'll see less harmonic fluctuations in the new chip when put on the test bench. Should be an interesting fight between the new HyperStream IV and AKM's upcoming Velvet Sound/Verita products.

[Yeah, I know there's also the ES9039MPRO. This is the version of the chip with embedded MQA "rendering" (as discussed here with the poor filters). Seriously audiophiles, enough with wasting time and money on this stuff, with basically only TIDAL HiFi+ tier content available. I don't think it's unreasonable of me to call this scheme literally stupid in 2022 and represents a downgrade in performance. IMO, avoid DACs using this chip to speed up the demise of the MQA nonsense and send a message to ESS and companies that still support it to not waste their R&D resources. Vote with your dollars.]

Finally, just a word about how some companies and reviewers keep telling us that "it's not the DAC chip... it's the analogue output circuits, etc... that determine the sound", usually with the intent of differentiating the "house" sound of the companies suggesting one brand might be significantly better than another (I've seen some companies hide the identity of the DAC chip used). Yes, of course that's true in that if the digital interface, or clocking, or power supply, or analogue output stages of the DAC are poor, then there will be sonic compromises. However, do realize that it is the DAC chip (or discrete circuit as the case may be) that is determining the ultimate potential resolution for the conversion. I think it is only reasonable to expect that reputable companies would want to achieve the highest fidelity from the converter chip so the final goal would be similar. No matter how excellent the analogue circuitry might be, how precise the clocking mechanism, the intrinsic properties of the DAC are what they are. As per the ESS graphs above, Topping, S.M.S.L., and Oppo all share in the fundamental property of ESS' HyperStream.

There are many analogies we can draw between DACs and computer components like CPUs these days. When you're sitting in front of a Windows machine working with a moderate to heavy load, you should be able to tell the difference in speed between an i3 and an i7 CPU. So long as the computer manufacturer allows the CPU potential to shine, do we attribute the expected speed mainly to the manufacturer like Dell, Lenovo, or ASUS, or do we recognize that it's mostly Intel's specific CPU model that provided the level of performance? So long as Dell, Lenovo and ASUS' supporting motherboard and components do not hamper the potential speed of the CPU (eg. slow RAM, buggy firmware, unreliable hardware), then in principle, the i3 or i7 will perform very similarly across brands. So too with modern DACs based on high-performance conversion chips; there is a potential level of fidelity which the chip is capable of and manufacturers can optimize that core potential, ensuring reliability, and perhaps adding useful functionality. Devices IMO will sound similar in most regards; and if they don't, we would almost certainly be able to measure the difference these days! ;-)

A fancy brand name can certainly get you the non-utilitarian rewards of a luxury product; just remember to not confound those putative benefits with playback fidelity by simply, truly closing our eyes when listening, dear audiophiles. ;-)

Okay ladies and gents, hope you're all enjoying the music! Happy Canada Day to the Canucks, and a great 4th of July to my American neighbours.

It's summertime in the north... Enjoy!


23 comments:

  1. Hi Archi

    A great report thanks. I noticed this also over the years, that there are big differences in the THD+N versus Level Graph between different DAC Chip technologies, besides differences in Jitter sensitivity and static THD differences. You can also clearly observe those differences between the Topping D90 (AKM) and D90SE (ESS) or RME ADI 2 FS V2 (AKM) and V3 (ESS).

    Also what is reported in a link that you mentioned, that some registers in the ESS chips are not as they are written in the data sheet, so you have to dive deeper into all registers to be sure, how the ESS chip does behave.

    I personally am also very surprised, how different the Topping D90 (AKM) to the D90SE (ESS) does sound, even both have top notch measurements (besides different low pass filters, so you have to select a different number for the low pass filter in the SE version compared to the non SE version, in order to get the same / similar filter settings).

    Have a great Sunday
    Juergen

    ReplyDelete
    Replies
    1. Interesting, thanks for the information Juergen,
      I've heard over the years as well from DAC designers about documentation issues with the ESS chips also - things like errors and undocumented. Looks like as a result there may be different levels of expertise based on experience working with them.

      Hey, in the days ahead would love to hear more when we have more information about those upcoming AKM and ESS chips!

      Have a great summer and safe travels if you're flying around this year...

      Delete
  2. I really liked your review. With your permission, I would like to share it on my social media accounts. I will cite the source. Thank you.

    ReplyDelete
  3. Thanks Archimago for your very interesting insights, which I am studying. For now I have a simple question: if "Measured values in the legend correspond to 0dBFS", why the fundamental value is not 0dB? Thanks again and good luck!

    ReplyDelete
    Replies
    1. Good point MTB ;-).

      The reference to the 0dBFS values is that those dBr numbers are with the cursor at 0dBFS (far right). That fundamental value of say 110.6dB (last MUSE graph) is a dB SPL level which I did not calibrate. You can safely ignore than number...

      Delete
  4. You may want to show these graphs (particularly the Topping D10) to the engineers at ESS. I watched a video on the HyperStream architecture and this may be expected behavior. According to Dustin Forman (VP of Engineering at ESS), HyperStream is supposed keep noise (distortion) roughly constant no matter the output level since this seems to be an artifact that people can detect in blind listening tests.

    See https://www.youtube.com/watch?v=TuEIs2PA1nU

    Skip to about 16:54 for the ESS presentation.

    ReplyDelete
    Replies
    1. Thanks Nobody,
      Interesting interview, haven't looked at these "Pints with Ayre" talks in awhile.

      Good discussion on the evolution of the ESS technology from 1-bit HyperStream to the multilevel, multichannel ES9008 in the mid-2000's, eventually dithered PWM.

      Nice slide set discussing "intersymbol interference" and wanting to achieve constant error. The Dynamic Element Matching might be another technique which could affect that distortion level.

      The high resolution Dynamic Range vs. DC Offset Level graph at 45:00 looks like an interesting test. Probably can manually measure something like that at a lower resolution. No automated stepped sine facility in REW at this time to measure something like this.

      Delete
  5. For the Oppo 205 measurements, did you test with both one DAC chip (the Front L/R choice for the dedicated stereo out) and with two (the downmix stereo choice)?

    ReplyDelete
    Replies
    1. Hi Rob,
      That was just with the rear stereo XLR outputs which is the highest quality audio out from the device...

      Delete
    2. In the setup menu you can choose the output from those jacks. Option 1 is "Front L/R" which uses one of the two DAC chips. Option 2 is "downmix stereo", which uses both DAC's (the dedicated stereo DAC + the multi-channel DAC) for the stereo output. I'm not quite sure about how this is done (one chip per channel? differentially summing the output from each chip as Topping does?), but Oppo designed this to improve the audio from the dedicated stereo outs.

      Delete
    3. Interesting Rob,
      Didn't know the "downmix" stereo option had this effect. I thought that would just fold the MCH to 2.0 and would probably just affect the RCA outputs so that only the R/L used.

      Next time I have this on the table, will have a peek!

      Delete
  6. At my limited level of technical understanding it appears that the AKM DACs perform measurably better than the ESS9038-based units. I currently use a Mytek Manhattan II DAC which I believe uses the ESS9038xxx chips and have been happy with its sound but have to admit not having heard any SOTA DACs.
    My question is whether those better measurements would be expected to translate to audibly "better" sound?

    ReplyDelete
    Replies
    1. ADDENDUM: Please note that I do not use the MQA capability nor do I use the internal preamp.

      Delete
    2. ADDENDUM 2: Anyone have thoughts on sonic benefits if any of a DAC using the AKM chips vs ESS9038 as implemented in the Mytek Manhattan II?

      Delete
    3. How about a $3.78 (yup!) DAC incl cables as your next test subject?

      https://www.aliexpress.com/item/3256803459250713.html?spm=a2g0s.8937460.0.0.f1a32e0eiEzWn2&gatewayAdapt=4itemAdapt

      But would still like RSVP to my previous post/inquiry.

      Delete
    4. Hey BigGuy,
      Depending on the DAC, the AKM and ESS chips take turn with which performs better using standard measurement tests. They're obviously both capable of very high resolution so I would not put any bets on one being able to differentiate two relatively equivalent DACs in a music blind listening test with volume controlled!

      The Manhattan II I'm sure is a fine DAC and assuming they designed it well, should have similar performance to other ES9038PRO's out there. It does offer quite a bit of connectivity options at the back.

      $3.78 eh? Might have to see about getting one of those to try out ;-).

      I bet you it'll measure similar to something like an inexpensive low-power dongle-type DAC like this:
      http://archimago.blogspot.com/2020/04/measurements-soditer-fourth-generation.html

      Delete
    5. Thanks for the comments, Archimago. DAC performance like many things is predicated on implementation so I understand that 2 well designed units could be close in an A-B comparison.

      FWIW, I was able to hear the 4-buck-DAC y'day in a friend's system. I have thoughts but defer until when/if you decide to "blow the ranch"!

      Delete
    6. LOL. Okay. Just purchased this phenomenal DAC for a grand total of CAD$5.28. Hopefully it'll show up in 2022 ;-).

      Sure... Let's "blow the ranch" BigGuy :-)

      Delete
  7. Great article!

    With regard to the specific pattern of ESS (all products, including the ADC which has a DAC inside), you may have noticed my thread over at ASR about the "ESS hump" on the Khadas Tone Board.

    After quite some experimeting and trying different measurement strategies it turned out that only time-domain analysis (with lots of time-domain averaging to see something in the noise) of the distortion residual gave me enough insights on both the exaggerated pattern of the KTB and the Topping D10B which has the same basic pattern, just lower in level, low enough to disappear in the noise when doing the typical THD+N or IMD+N vs level measurements.

    Looking at the time-domain residuals over and over (example: https://www.audiosciencereview.com/forum/index.php?threads/measurements-ess-hump-revisited-khadas-tone-board-v1-3.30136/#post-1057416) and breaking my head about it, at one point it finally occured to me that a *periodic* ripple on top of the sample-to-sample transfer function (see https://www.audiosciencereview.com/forum/index.php?threads/measurements-ess-hump-revisited-khadas-tone-board-v1-3.30136/page-9#post-1083031) is giving exactly the same patterns, structurally. A basic (almost) static transfer function nonlinearity. I even was able the write a small emulator code for the ESS hump of the Khadas Tone Board (with adjustable error level, actually).

    Near zero (below -30dB here) everthing is fine... and at very high levels the selected points along the transfer function vary so much that it creates almost something like dither, just adding noise, notably when testing at higher frequencies. For the same reason, the dithering effect, this distortion pattern only increases the flat bottom line noise floor in 32-tones multitone and thus is very hard to spot without a zero/low-signal reference plot.

    It looks like the underlying mechanism is (at least mainly) RF demodulation in the OpAmp-I/V circuit. But even when taking care to avoid the demodulation the basic pattern still is there, hinting at the glitch energy itself piling up, even in a flawless integrator/filter.

    Key point is that the glitch energy is not constant but appears to vary vs. momentary signal level. The final (thermometer-coded) output DAC of ESS has 64 segments (6-bit) and the modulator (in PCM-mode) changes how many of the segments are used on average and that alters distributed glitch energy. That's at least my current pet theory of things here.

    Note that this would also mean that ESS DACs are pretty much doing what they are exactly claimed not to do in their advertisement: noise floor modulation

    ---------

    With regard to AKM DACs, there I saw non-harmonic dirt components when testing at levels (IIRC) around -30dB (for 4490) or -20dB (4493). Probably much different cause than the ESS ripple, I suspect "limit cycles" in the DWA (data-weighted averaging) section that selects the used output cells for a given code in a pseudo-random fashion, to distribute mismatches. The tones dissappear quickly with the tiniest amount of DC offset and/or other components (even a very small amount of any harmonic, let alone a non-related tone), quite different from the ESS pattern which is rather persistent and where the dirt is strictly harmonic.

    ReplyDelete
    Replies
    1. Nice, impressive work KSTR!
      Great detective work going on. Wondering, do you know if ESS themselves are monitoring the thread or if any of the ESS folks hang out on the forum?

      Since this "ESS hump" has been discussed for a number of years now, I would imagine those folks are well aware and probably have a clear recognition of the issue.

      Delete
  8. This comment has been removed by the author.

    ReplyDelete
  9. It would be interesting to see the results of DAC-ADC loops (AES-118 test, df-metrics, didier@gearspace...) and the same measurements of the SM5847->PCM1704->LTC2380-24 chain (demo boards, for example).

    ReplyDelete