Saturday, 9 March 2019

MEASUREMENTS: A look at the audio "ultra high-end" - ultrasonics! (And changes at Stereophile announced.)

In this blog installment, let's look at the "ultra high-end" of sound. Of course, I'm not talking about the audiophile "high-end" marketing term which is meaningless (beyond just another phrase for "expensive"). Rather, let's look at the frequency high-end, especially all the stuff our DACs can produce in the ultrasonic range!

If you've ever wondered, ultrasound devices used in medical imaging typically function at around 2MHz on the low end up to about 15MHz. However, the term "ultrasound" simply refers to wave
compression and rarefaction outside of the hearing range which by convention are those above 20kHz or so.

While the presence of ultrasonic content coming out of our DACs is not a surprise, what might be unclear or debatable is whether there is much of it and whether this then affects the "sound" of one's system. Rather than get bogged down in opinions, let's first have a look at what's in the ultrasonic frequencies coming out of DACs... Starting from facts, we can then perhaps come up with opinions.

Remember that over the last few months, I have been using the RME ADI-2 Pro FS ADC as a measurement device. This device is capable of operating at 768kHz although up to now I have rarely used it at the highest rate. So why not push the system up to that limit and examine the spectrum up to the full 384kHz bandwidth to see some ultrasound components? While we're on this topic, let's also see if we can examine the frequencies beyond using an oscilloscope.

I. A look at the Digital Filter Composite (DFC) at 768kHz sampling rate...

As you know, over the years, I have used Juergen Reis' idea of overlapping FFT tracings to look at digital filter performance (a variant of the description in Stereophile back in 2013). In one graph, we can see the frequency response of the DAC, whether it overloads with 0dBFS peaks, examine intermodulation distortion products, the presence of imaging distortions and even amount of harmonic content.

For the most part, I've been running these DFC measurements at 192kHz; a great samplerate to make sure that filters operating on 44.1 data are working well and that there are no anomalies up to 96kHz. IMO, this is good enough for a very reasonable evaluation of the quality of a DAC's filter; with almost 100kHz of bandwidth to check for anomalies, it's hard I think to argue that more is needed.

However, given that in the last while I've been looking at "perfectionist" upsampling techniques like HQPlayer which seeks to squeeze out the last drop of computational accuracy, let's go all the way and have a look at whether there's anything up there in the ultrasonic frequencies to be concerned about.

From a procedural perspective, instead of using the one-step WaveSpectra software which allows me to watch the FFT in real-time and take snapshots, in order to construct the graphs at 768kHz required that I split this into a 2-step process. First, I recorded the waveforms in 768kHz using Sound It! 8 Pro from the RME ADI-2 Pro FS. Then I would import the audio into Adobe Audition CS6 to run the FFT (65536 points, Blackmann-Harris windowing used).

And here is what I get with the SMSL iDEA DAC, plotting the components of the DFC at 768kHz samplerate (Nyquist out at 384kHz):

Certainly not bad for a <US$100 USB DAC! Notice for the wideband white noise with 0dBFS signal peaks, there is a bit of intersample overloading which goes away when we attenuate the noise (by -4dBFS). The 19 and 20kHz high amplitude tones look good although we see a number of higher harmonics present with the last cluster just minimally poking out over the modulator noise floor of the ADC up above 170kHz (remember that the ADC is a sigma-delta device, hence the modulator noise).

What if instead of using the standard filter in the iDEA's ESS SABRE9018Q2C DAC, we use a very high quality resampling algorithm in HQPlayer? Let's try "poly-sinc-xtr", upsampled to the full 768kHz limit of the DAC:

As you can see, the main difference is that the 0dBFS wideband white noise overloading is no longer an issue. Also, the "brick wall" at 22kHz is steeper using HQPlayer's resampler, demonstrating a tighter transition band. Notice that managing the overload is accomplished with a reduction in the signal level. We see this with a reduction in 19 & 20kHz signal peaks as well as the 0dBFS white noise level compared to the -4dBFS signal. We're looking at about a 3dB attenuation of these strong signals.

As for the 19 & 20kHz signal's higher order harmonics, they are attenuated a bit as a result of the primary signal being lowered. While software upsampling and filtering can improve things like stop-band rejection, prevent intersample overloading, and providing a steeper filter cut-off, the hardware limitations like harmonic distortion are inherent in the device.

Suppose we now compare the above with the much more expensive Oppo UDP-205 DAC / player. Remember that the chipset here is the "flagship" ESS ES9038Pro SABRE DAC. Here's what the 768kHz "DFC" looks like with the standard "linear phase fast" filter (my preferred setting for the Oppo):

That looks very good already! Remember that this is not a surprise given my findings last year. I recorded off the XLR outputs for the lowest noise level. What's also very evident compared to the SMSL iDEA is the extremely low harmonic distortion which is one of the ES9038Pro's claim to fame.

Now then, what happens when we run the UDP-205 output through HQPlayer using the same settings as above with the iDEA?

Clear improvements, but not huge... Notice that while it is certainly cleaner than through the standard DAC filter, the difference is not as marked as with the SMSL iDEA. The Oppo's built-in filter already did not experience intersample overload. But the even sharper "poly-sinc-xtr" filter did remove the small image just beyond 22.05kHz.

The bottom line is that if one's DAC already has a good digital filter in place, as expected, doing the job in software will not result in much improvement.

From what we see here, there's certainly not a huge amount of ultrasonic content to be found with these ESS DACs. As far as I can tell, there's really nothing above -85dBFS beyond 22.05kHz with the Oppo/ES9038Pro when playing 44.1kHz audio. For the iDEA/ES9018Q2C, while the 2nd and 3rd order harmonics were significantly more prominent compared to the Oppo/ES9038Pro, there's still very little high frequency noise/distortion above 40kHz.

II. Filtering the 384kHz samplerate signal...

I know this next topic is rather ridiculous since I'd be surprised if anyone cares about filtering of a 384kHz samplerate signal. DAC chips are low-power devices and have limited computational ability. We know that some DACs will not implement digital filtering at high samplerates and may just implement an analogue filter to roll off the high end. However, for those DACs that can play 768kHz, we should be able to use HQPlayer and demonstrate the application of a filter even out at almost 200kHz!

We can do this with the Oppo UDP-205:

Notice that the Oppo does implement filtering around 192kHz, but it's much "weaker" than using a more capable computer with HQPlayer.

We can do the same process using the SMSL iDEA:

What we can say about the iDEA is that the frequency response is rolling off already quite a bit by 192kHz. I didn't check if this is because they're using a very weak digital filter. HQPlayer is obviously effective as you can see. Again, whether anyone cares is another matter :-).

III. MHz range ultrasonic noise?!

Going beyond 384kHz bandwidth requires a different tool than the RME ADI-2 Pro FS. I don't have anything fancy for this but I do have a Rigol DS1104Z digital oscilloscope that can give me a low-resolution FFT to look above 384kHz and into the MHz frequencies. Remember that the majority of these less expensive digital storage oscilloscopes provide only 8-bit resolution and are meant to be used on multi-MHz waveforms, not kilohertz audio. As a result, we can't expect detailed graphs, but it'll give us an idea if there are large amounts of noise out in the MHz range.

[As an aside, to those who might have a Rigol DS1000-series scope, there is a Python-based project whereby higher quality FFTs can be derived (software on Github). I'll give it a look when I have more time.]

On screen: A 100kHz signal off the SMSL iDEA...
The oscilloscope is able to measure up to 100MHz which is beyond "ultrasound" of course. To see if we can correlate the oscilloscope's FFT with what I see using the RME ADC, here's what happens when we send a 0dBFS 24/384 150kHz signal to the SMSL iDEA:

Hmmm, ugly. It looks like the high amplitude, high frequency signal is too much for the little DAC/headphone amp to handle; note the grossly elevated noise level across the spectrum and "skirting" of the primary signal. Furthermore, there's also a strong image of the primary tone out at 234kHz. Knowing this, let's cross-correlate this with the oscilloscope FFT to make sure the two peaks can be seen easily there as well:

Yup. We see the 2 frequencies. Each horizontal division is 200kHz and we can also see the 427kHz image a little further out.

Can this anomaly be fixed with software upsampling like HQPlayer, JRiver, etc? "Yes we can!" - as in Part II above, assuming you think adding a high quality digital filter is worth it for 384kHz samplerate material, it can be done with good results...

Beauty. Excess noise improved and strong 234kHz image filtered out; demonstrated on both the RME tracing and Rigol FFT.

In comparison, when we put the same 0dBFS 150kHz (24/384) test signal through the Oppo UDP-205 with standard "linear phase fast" filter setting, we don't see any issues with the noise floor nor is that 234kHz image present. In fact, it looks excellent! No need for HQPlayer here:

We see only a single 150kHz primary peak as expected on the Rigol FFT. There's possibly something out at ~1.3MHz however but clearly at much lower level than the primary signal.

Moving along, let's push this even further to levels that the RME ADC cannot capture. How about a -3dBFS 300kHz 24/768 sine signal? Here's the SMSL iDEA:

We see imaging of the 300kHz primary signal out at 468kHz. This suggests that the iDEA doesn't use a digital filter with 768kHz audio data, or if implemented, a weak filter.

And how about the Oppo UDP-205?

Looks good with just the single 300kHz tone. Nothing terrible all the way to 2.4MHz... Perhaps a bit of noise around 1.2-1.4MHz, but again this does not appear to be of high amplitude.

For "fun", here's a wideband 24/384 white noise signal using both DACs:

Despite the low resolution of the FFT, as expected based on Part II, we see that there's more high frequency roll-off with the iDEA. While the resolution obviously isn't good enough to catch small anomalies in the noise floor, there does not appear to be strong noise all the way up to 2.4MHz. If anything, with the Oppo, we continue to see a "little something" around 1.3MHz.

Before I dismantled the set-up, I thought I'd just throw out one last test. Here's the Oppo using a track from "Archimago's Insane and Unreal Esoteric Tests" collection - a 24/384 signal with 140 & 170kHz -6dBFS sine waves:

Even with these high amplitude and very high frequency signals measured at the analogue output, I see nothing all that scary using the high-resolution RME ADC, nor am I seeing anything of concern with the oscilloscope scanning out to 1.5MHz.

IV. A quick look at the TEAC UD-501 (TI/BB PCM1795 DAC)...

For years, since 2013, I used the TEAC UD-501 as my reference DAC. Although I don't use it as much these days, I figure it would be good to put the TI/Burr Brown PCM1795 DAC chip to the test as well using the 768kHz DFC and my oscilloscope.

While there's some intersample overload with that wideband white noise with 0dBFS peaks, this does not look bad although we have some -70dBFS distortion out beyond 330kHz. The DAC chip implements 8X over-sampling of the 44.1kHz signal to 352.8kHz and what we're seeing are images out at 332.8+333.8kHz and 371.8+372.8kHz.

One "nice" feature with the TEAC UD-501 is that we can turn off the digital filter from the DAC menu. This makes the DAC into a "Non-OverSampling" (NOS) device. Here's what the DFC looks like with the TEAC in NOS mode:

Remember this distorted "cascade" when you listen to devices that are NOS at 44.1/48kHz without appropriate upsampling applied! In fact, using the TEAC, it's a much cleaner looking pattern than one would see with old NOS parts like the multibit Philips TDA1543.

If we look at just the 19 & 20kHz sine waves in the oscilloscope with filtering turned off (NOS), while the resolution isn't good enough to see individual frequencies (like the separation between 19 and 20 kHz), we again will see the cascade of ultrasonic distortions:

Isn't that "pretty"? Not for an audiophile who desires the cleanest, highest fidelity signal! :-)

The maximum samplerate for the TEAC is 384kHz. At this high samplerate, there is no digital reconstruction filter applied and one would expect therefore that something like my 140 + 170kHz signal (@ 384kHz as per Part III) will show imaging distortions as well:

Yup. Notice that even with the relatively low resolution of the Rigol FFT, we can see the images out to the 524+554kHz and 598+628kHz pairs. Otherwise, not much else of concern out to 1.5MHz.

V. Ultrasonics and DSD...

I suspect in the back of your mind, you must be thinking - "What about DSD!?" After all, what kind of article talking about ultrasonic frequencies for the audiophile neglects DSD, right?

That is indeed an important question which we must not avoid regardless of one's opinion on the value of having music encoded as DSD. The simple fact is that with a 1-bit digitization scheme, quantization error for each sample is large, +/-0.5 for complex signals. This translates to a large amount of noise in the signal which needs to be addressed if we're aiming for clean, high-fidelity playback.

Over the years we've seen this noise manifest every time I measure DSD output, most noticeable with DSD64 and DSD128. In order to improve the noise floor within the audible frequencies, aggressive high order noise-shaping is used to shift the noise to the ultrasonic range and DACs will typically then implement an analogue filter to roll-off high levels from there.

For comparison then, here are the "DFC" graphs of DSD64/2.8MHz, 128/5.6MHz, 256/11.3MHz, and 512/22.6MHz from the Oppo UDP-205 using HQPlayer 3.25 to perform the PCM-to-SDM realtime conversion with an i7-3770K CPU + nVidia GTX 1080 GPU using "poly-sinc-xtr-2s" resampling (the "poly-sinc-xtr" setting is extremely CPU/GPU hungry especially at DSD512). I chose to use the "DSD7" 7th order SDM setting as a demonstration of the kind of aggressive noise shaping often used to improve dynamic range below 20kHz.

There are a couple of unexpected noise spikes in the "Digital silence" plot at DSD64. Not sure where that's coming from. Otherwise, nothing unexpected here.

Notice the ultrasonic noise pattern - how quickly and how high it "takes off" beyond 25kHz with DSD64 especially!

As you can see, noise is not as severe with higher samplerate DSD and is similar to PCM by the time we hit DSD512. Remember what I've said many times before... The higher the DSD samplerate, the more the objective characteristics look like PCM; in the case of these DFC diagrams, the true noise floor is below the ADC's ultrasonic modulator noise level by the time we hit DSD512. If you like the sound of DSD, it's quite possible that the noise is experienced as euphonic/preferable, hence the high noise DSD64/128 patterns could sound "better", with "more air", appear "more three-dimensional", "more analogue", etc. with some sound systems.

VI. In Conclusion...

With the <US$100 SMSL iDEA, the >US$1000 Oppo UDP-205, both using ESS DACs, and adding the TEAC UD-501 from 2013 with its TI/Burr Brown DAC in the mix, I think we can say based on the data from the RME ADC with 384kHz bandwidth that there's really not much in the "high end" at all during PCM playback when using appropriate digital filters. Sure, likely there's ultrasonic stuff below the limits of the RME's AKM AK5574 ADC modulator noise floor, but we'd be looking for signals below about -85dBFS at frequencies significantly above 100kHz!

Furthermore, using the oscilloscope's FFT function, though the results are of much lower resolution, we are not seeing large amounts of ultrasonic noise in the MHz range.

In contrast, we can certainly see significant amounts of ultrasonic content when we turn off the digital filtering (ie. NOS DACs) and with DSD playback, especially DSD64 and DSD128. So, for audiophiles, if the extremely high ultrasonic imaging found with NOS playback and persistent DSD64 ultrasonic noise are not considered generally objectionable, that's in fact evidence that ultrasonic content is not a big deal and do not severely affect sound quality.

By the way, remember that the SMSL iDEA is a tiny, USB-powered DAC. In these tests, I'm connecting the DAC to my Intel i7-3770K desktop CPU currently overclocked to 4.0GHz. Inside that machine is also the nVidia GTX1080 GPU, 16GB of DDR3 RAM, an SSD and Firecuda hard drive, connected to the home ethernet network with no special power conditioning. Notice that despite what should be a "noisy" computer system, that neither the 768kHz "DFC" nor the oscilloscope readings show terrible noise.

As discussed in late 2018, I'm far from impressed by audiophiles who seem to believe that there's all kinds of noise, jitter and whatnot affecting the sound output from otherwise competent DACs connected to computers through USB. Anxiety over things that appear to have no basis in reality amounts to perpetuation of unsubstantiated fear, uncertainty, and doubt. I hope magazines and on-line sources seek to find balance and evidence when these kinds of ideas are presented.

Inevitably, we must ask ourselves then, is there evidence here of anything to worry about when it comes to ultrasonic noise? Is there any ultrasonic noise that likely will be affecting the "sound" of our DACs? From what I've found and what I hear, I have no concerns about the SMSL iDEAOppo UDP-205, or TEAC UD-501.

As for software upsampling (eg. HQPlayer, Roon, JRiver...), yes, they work and can improve DAC output where the built-in digital filter may not be great. Software upsampling is also particularly useful for NOS DACs (as discussed before). Remember this the next time you hear an audiophile claim that their Audio Note "1x oversampling" DAC somehow sounds remarkable. I believe this says more about the individual's preferences and other sound system components than actual high fidelity performance from the DAC itself. Remember though that if you have a good DAC already with good filtering (like the Oppo with its "linear phase fast" setting), do not expect remarkable sonic revelations using software upsampling. At best, differences will be subtle.

Historically, there was a time when audiophiles and magazines hyped up the importance of ultrasonic frequency reproduction. Remember back in the late 1990's and early 2000's with the advent of SACD and DVD-A, the idea of "super tweeters" was in fashion? Back then, I remember reading magazine articles tout the ability of these ultrasonic transducers for improving soundstage and "air" (I discussed this back in 2017). Consider the possibility that these "super tweeters" could actually make sound quality worse if not implemented properly by adding yet more drivers to the speaker design.

These days, the need for "super tweeters" and claims that extended high-frequency response being beneficial are quite rare. In fact, if we look at the in-room frequency response in speaker measurements, typically there's roll-off into 20kHz ("house curves" with high frequency roll-off are often subjectively preferred). Likewise, headphones, including "hi-fi sounding" devices like the Sennheiser HD-800 typically have more than -10dB dips by 20kHz compared to the amplitude at 1kHz. Furthermore, little recorded music actually contain large amounts of >20kHz content (remember, studio microphones also have frequency limitations). This, combined with the fact that most of us adults likely at best can perceive frequencies up to 16-18kHz argues strongly against needing to pay much attention to ultrasonic content. As I noted above, if audiophiles aren't complaining about the ultrasonic distortion/noise of NOS DACs and DSD64, what's there to worry about? Therefore, technically, likely the only time we need to be concerned is with unstable amplifiers unable to handle excessive high frequency content.

To end off... These are probably some of the most insane/extreme measurements I've published in this blog. However, I'm sure I'm not the only person curious about the presence of ultrasonic frequencies. Remember, as per the post a couple weeks back, test tones like these are obviously not like "real music", but they certainly can at least give us some idea of the technical limits of equipment and whether the gear still performs in an expected fashion.

In any event, while I'm sure we will continue to discuss and measure using 384 and 768kHz sample rates, I hope this is the last time we need to discuss frequencies in the MHz range from audio DACs. :-)


I see the announcement this week that John Atkinson will be stepping down from being the Editor of Stereophile. This is certainly a significant change and one that is inevitable. Nobody stays around forever...

Despite the numerous complaints against audiophile magazines/press and the false ideas they may perpetuate, I do appreciate John Atkinson's Stereophile in maintaining some objective perspective for how reviews should be presented incorporating both artistic sensibilities and scientific evaluation. This is far and away superior to many alternatives out there both in print and online (no need to name names - rather obvious). Congrats to Mr. Atkinson on the years of work and the legacy this will certainly leave.

It's good to hear that "major editorial changes are not envisioned". Hopefully this implies that Stereophile can maintain the balance between subjective and objective perspectives in their reviews. In my opinion, the objective measurements in the magazine are the portions that endure and make articles worthy of repeated reading over the years. The objective results speak of true performance. Stereophile has documented the advancement of technology over the decades, and helped educate hobbyists seeking to understand the pursuit of fidelity as practiced over the generations. This is the stuff that transcends ephemeral, subjectively idiosyncratic, toss-away fluff that fades like the fads of this world or even just the next news/product cycle.

At a time in history where the boundary between fact and fiction has been blurred sometimes beyond recognition, where it seems "my truth" has frequently overtaken "the truth" even in common sense, rather unequivocal matters, it would be rather unfortunate at least for North American audiophiles to lose the perspective Stereophile has brought to our discussions.

Hopefully, as Mr. Austin takes over the helm of the publication, he will not forget the science of his earlier years (talked about in the announcement article linked above). In the words of the inimitable physicist Richard P. Feynman:
"I would rather have questions that can't be answered than answers that can't be questioned."
"Religion is a culture of faith; science is a culture of doubt."
I'm certainly not calling for the magazine to turn into some kind of serious scientific journal; rather, let's talk about fun hardware and great music while maintaining a decent level of reality-testing and willingness to examine unconventional beliefs using scientific principles.

Here's to hopefully a helpful, truthful, honest, intelligent, reasonable, educational while also creative, entertaining, and ultimately successful tenure for Mr. Austin. Here's to wishing that Stereophile remains a publication of significance in building the culture of the hobbyists it seeks to serve.

Until next time, I hope you're all enjoying the music!


  1. Thank you very much. Very nice. Even a cheap SMSL dongle with "very limited processing power" has nothing to show beyond 352.8k. I am glad I don't need to force myself to listen to music with >200W TDP CPU/GPU while worrying about noise slipping through unbalanced interconnect to my platinum ears! A waste of energy IMO.

    The DSD tests showed similar results as your previous ones: highest <20kHz noise floor with DSD256, but I am a bit surprised that the DSD512 one has more spurs below 20kHz.

    Also a good laugh about the 140/170kHz IMD tests. Sometimes I also do ridiculous experiments like resampling to and from prime numbers :-)

    With these kinds of measurement wars one just need to pick some questionable boutique DACs to show the "benefits" of extreme external upsampling and spreading FUD.

    1. FYI:

    2. Interesting... LPS = audiophoolery, as expected?

    3. Wow, nice work with the Topping (AK4493-based DAC) and AP gear!

      Yeah, sometimes I need to do these kinds of experiments to make sure I address my own internal neurosis when people tell me stuff I'm supposed to "hear" ;-).

      It's fun nonetheless looking for answers to calm my nerves and hopefully the collective nerves in the audiophile community!

      Yeah verifonix, this linear power supply stuff had always IMO been suspect from the start. Remember when Benchmark put out their post and video back in 2016:

      Over the years, various people have explored this including some of my postings here. Even if there are small differences, like Bennett's lower 400-500kHz LPS Wyrd4Sound vs. switching supply graph, I have not seen evidence of problems when using otherwise good DACs.

    4. I can run upsampling to DSD512 using 35W TDP i5-7600T.

      Smallest I've used for DSD256 has been 5W TDP Atom.

      When I want to run digital room correction at 2.8 MHz sampling rate (DSD64) for SACD rips, then sure I will need much more processing power!

      Or running 16 million tap filter...

      So depends on what you want.

  2. I may be going out on limb here, or just prove myself an utter fool, but I believe there’s a huge misunderstanding of the digital model here.

    The sample rate frequency, be it 192 or 768, has nothing to do with the audible range of the sound curve. It’s merely the sample rate. To wit: it denominates how often the original analog wave has been sampled in order to create its digital profile. The more data points, the smoother the digital curve. (Traditionally, we depict the timing along the X-axis and its granularity is the sample rate). To what extend a smoother curve is beneficial to the listening experience is a subject to debate, although empirical evidence suggests its far below 192.

    As for the sonic properties of the sampled sound wave, we associate with each data point a voltage value, where the number of values and maximum value is confided by the bit depth: 2^16, 2^24 and 2^32. We traditionally depict that information along the Y-axis. If there’s any ultrasonic information recorded in the process, it will be preserved here, in a voltage form. And it has nothing to do with the sample rate frequencies.

    I believe the confusion stems from the term “frequency.” The y-axis presents the sonic frequency of each data point. The granularity of x-axis is defined by the frequency of the sampling process. Unfortunately, the two got conflated, causing havoc in our understanding of the digital model.

    1. To address the comment by Jitter Free:

      "I may be going out on limb here, or just prove myself an utter fool, but I believe there’s a huge misunderstanding of the digital model here."
      If I may be so bold to suggest, this would be more than a slight misunderstanding on your part. Perhaps a retreat to the safer confines of the trunk might be a more prudent choice.

      "The sample rate frequency, be it 192 or 768, has nothing to do with the audible range of the sound curve. It’s merely the sample rate. To wit: it denominates how often the original analog wave has been sampled in order to create its digital profile. The more data points, the smoother the digital curve."
      It has everything to do with the audible range, because ANY audible frequency in up to a value of ν/2 (ν=sampling frequency) can be FULLY and FAITHFULLY captured (in other words f < ν/2)Hence, the higher the sampling frequency, the wider the frequency range that may be captured. This is a fundamental THEOREM, mathematically proven, and not "just a theory" as trotted out by the misinformed in tropes that really ought not to surface these days. The entire scientific field of signal processing rests on this and has done so for close to 90 years and will continue to do so, despite the linguistic machinations of the Bob Stuarts in the world.

      "As for the sonic properties of the sampled sound wave, we associate with each data point a voltage value, where the number of values and maximum value is confided by the bit depth: 2^16, 2^24 and 2^32. We traditionally depict that information along the Y-axis. If there’s any ultrasonic information recorded in the process, it will be preserved here, in a voltage form. And it has nothing to do with the sample rate frequencies."
      Sadly not. The minimum bit-depth required is stipulated by the dynamic range of the analog signal that we seek to faithfully capture. To repurpose your language, it has nothing to do with ultrasonic information.

      "I believe the confusion stems from the term “frequency.” The y-axis presents the sonic frequency of each data point. The granularity of x-axis is defined by the frequency of the sampling process."
      I believe the confusion arises from a lack of knowledge and understanding of the sampling theorem. This in itself is not a big deal; the big deal is a result of a lack of desire to understand the simplicity and elegance of the sampling theorem. Statements such as "The y-axis presents the sonic frequency of each data point" merely illustrate this issue in glaring splendour.

      "Unfortunately, the two got conflated, causing havoc in our understanding of the digital model."
      Unfortunately there is little conflation here to write home about. By appropriating the tautology of the eminently quotable Theresa May, I think it is fair to suggest that the only thing causing havoc in your understanding of the digital model is your understanding of the digital model. With all the due respect.

    2. Despite your demeaning and epileptic tone, I thank you for pointing out that the highest frequency we can capture is half the sample rate used.

      “It has everything to do with the audible range, because ANY audible frequency in up to a value of ν/2 (ν=sampling frequency) can be FULLY and FAITHFULLY captured (in other words f < ν/2)”

      Still, here you let your froth take over:
      “Sadly not. The minimum bit-depth required is stipulated by the dynamic range of the analog signal that we seek to faithfully capture. To repurpose your language, it has nothing to do with ultrasonic information.”

      If it’s not stored as a voltage value, where and how is it stored?

      Two more points of worth:
      The highest ultrasonic frequencies in nature are in 80khz. Thus, the only reason you’d use sample rates over 192 is to produce a more continuous curve.

      Many sound engineers clip all frequencies over 23khz…How about that.

    3. Jitter Free, not sure where you get your information, but you are REALLY off-base here. You don't understand sampling theory, you don't understand bit depth, you don't understand sound frequencies and make untrue claims about what exists "in nature", etc.

      I encourage you to keep reading and learning, but also to refrain from posting non-factual material until you have a bit more understanding, if only to keep from misleading others who might stumble across your posts via a search, etc.

    4. Archimago you are my hero, thanks for all the work you do; the best audio blog and great for de-bunking myths. I have yet to persuade my friend though that the UK £500 interconnect they bought to link his CD transport and DAC is not making any difference. Any tips; or is it better to let him believe it is?

      Jitter Free as per previous comments it would be worth taking some time to do a bit of research. If interested this is the best video I know that covers a basic introduction and tackles head on the 'The more data points, the smoother the digital cure' myth as attractive it may seem it is simply wrong; video here or originally here; original article from which it came here

    5. There are things in life that may seem intuitive and the idea of "the more data points, the smoother the digital signal" is one of them. I suspect to some extent we have to blame those silly stair-step curves people like Sony used to suggest that "hi-res" is better because the more steps there are, the "smoother" it looks and hence "better".

      Maybe another source of the confusion comes with audiophiles themselves, associating the "infinite" smoothness of analogue with samplerate or maybe manufacturers purposely turning to NOS and literally choosing to keep the waveforms "stair-stepped" (and suggesting this is a good thing!) contributed.

      The bottom line is this: A properly reconstructed digital signal faithfully and completely allows us to reproduce all the frequencies below fs/2. Make sure your DAC's filter is good, and we're all set. This might not seem intuitive at first glance but once one spends time reading and thinking about this, the beauty and elegance cannot be denied.

      As an anecdote of something similar, the other day one of my kids while in science class learned that objects of different mass are affected by gravity in the same way. Heavy cannon balls fall the same speed and accelerate at the same rate as a cannon ball half the diameter! Who would have thought until Galileo dropped some objects off the Leaning Tower of Pisa as the story goes in the late 1500's...

  3. I always enjoy your blog posts. These are the kind of questions I like to see answered, and the kind of experiments I'd be doing myself if I had the equipment and (frankly) the testing background ;)

    Thank you!

  4. Hi Archi

    Again, a great post. Thanks.


    1. Thanks for the note Juergen! A please chatting over the years :-).

  5. These are probably some of the most insane/extreme measurements I've published in this blog.

    ROTFLMAO. Exactly. You read my mind... both fun and funny. I also like to put myself in the mind of the engineers that designed these products coming across some rando on the interwebz using them waaaaaaaaay outside their design criteria: "LOL. He did wut!?! Hey, wow, they still work. Who knew?"

    1. Hey Allan...

      Yeah, it's gotta be done :-). However, I suspect the engineers who designed these things must have testing the devices at these specs to check on the design.

      Imagine if a company releases a DAC advertising PCM768 and DSD512 but it ended up failing and sending extreme levels of noise... That would suck if audiophiles start a class action lawsuit because mega-power amps started to oscillate and catch fire. Then again, maybe the amp manufacturer needs to be sued as well :-).

  6. Hi @archimago and @miska

    Looking at the Oppo 205 results, fed via HQPlayer poly-sinc-xtr filter, DSD512 vs PCM 768Khz:

    which of the DFC charts looks best? I'm not asking about audible differences, but technically best objective data?

    Does the Oppo 205 perform technically better fed PCM768 or DSD512?

    Thanks again for the great review

    1. Hi Unknown,
      Yeah, that's a hard question because as you suggest, it's not about audibility. If I were to perfectly match the output level, I do not believe I can hear a difference at all.

      Therefore it comes down to the objective data and what I see suggests that the PCM768 results are probably cleaner with better dynamic range and lower overall noise.

      The other benefit of PCM is that it practically requires much less processor power. DSD512 with "poly-sinc-xtr" overwhelms my i7-3770K + nVidia GTX 1080 GPU combination! Furthermore, even if I could run DSD512/"poly-sinc-xtr" all day long on my computer, that means more energy use, more heat, potentially more fan noise, and perhaps even more electrical noise (not necessarily resulting in a sonic difference from a DAC, but there could be microscopic effects).

      Remember that since I also like using DSP for room correction, DSD512/"poly-sinc-xtr" would require quite a "He-Man"-level computer just to be glitch-free when one tags on even more processing!

    2. It doesn't take that much with poly-sinc-xtr-2s, or even with poly-sinc-ext2 to DSD512 with room correction. Since room correction is run at the source rate, upsampling doesn't increase load from that perspective. And you can offload the room correction (convolution engine) in HQPlayer to the GPU while keeping the upsampling on CPU. So for RedBook sources, room correction doesn't really have any notable impact on load. Only when you start doing room correction for DSD sources it gets heavy.

      My current machine for playing to Holo Spring DAC with poly-sinc-ext2 (and poly-sinc-xtr-2s) at DSD256/DSD512 with room correction is passive cooled (totally fanless) i5-7600T (35W TDP model) and it works fine. No GPU. It cannot do room correction for DSD sources though, but I don't have so many of such.

    3. Thanks for the note Miska,
      Yeah, "poly-sinc-xtr-2s" and "poly-sinc-ext2" works well for me also. Sounds and measures great...

      So long as I stay away from "poly-sinc-xtr", no problem.

  7. Please do seek out correct information before doubling down. There are some fundamental misconceptions that make it very difficult to have a fact-based discussion.

    "If it’s not stored as a voltage value, where and how is it stored?"
    What is associated with bit-depth is, as mentioned, not ultrasonic information. Bit depth has nothing to do with frequency response/bandwidth; fundamentally, bit-depth determines the granularity with which the amplitude of the sample is stored. Bit depth places limits on S:N and dynamic range (oversampling, dithering, noise-shaping, etc notwithstanding). It is analogous to colour depth in digital imaging, if that helps.

    "The highest ultrasonic frequencies in nature are in 80khz. Thus, the only reason you’d use sample rates over 192 is to produce a more continuous curve."
    Please do yourself a favour and relinquish the notion of "continuous curves" — this is not how the math works.

    The reason for higher sampling rates has to do with the conversion process from D-A, specifically for moving quantization noise out of the audio band into the ultrasonic from where they may be filtered out by the audio DAC; however with decent up/oversampling available now, that is not as valid a reason as it once was. Another reason for higher sampling rates is marketing to the folks (and this includes studio execs) who think that higher numbers naturally give moar.

    I confess that I, too, get some irrational comfort in having a 24/192 studio master knowing fully well that a 24/96 or sometimes even a 16/48 downsample would be a more than adequate medium.

    "Many sound engineers clip all frequencies over 23khz…How about that."
    Many microphones and analog tape recorders cut out at 16kHz.* I’m not sure what this non sequitur is in aid of.

    Instead of considering responses as being epileptic (although I admit they may be somewhat elliptic at times (•◡•) ) or frothy (I personally prefer the term "bubbly" — it conjures up happy images), and attempting to defend the indefensible, it might be helpful to pave over any gaps in understanding. It is it is far easier to grasp the Universe as it really is than to persist in delusions, no matter how comforting and reassuring they are, once dogma is replaced with doubt. I regret the misconstrual of tone, and certainly did not intend for the message to be demeaning. At the end of it all, we are all here to learn more and be a tad wiser, both of which are enjoyable rewards unto themselves, imho.

    *check out the 24-192 PonoMusic remasters of early Neil Young albums; there is pretty much no musical content above 18kHz, IIRC. My playback files are downsampled 24/48 but the backups on the backup server are the original 24/192 because, although I consider myself to be a rational audiophile, I am an irrational human. ¯\_(ツ)_/¯

  8. Nice set of results!

    To me it looks like DSD256 performs cleanest in this case... But due to how ESS Sabre works with DSD (it is not direct conversion) some of the ESS Sabre "wandering noise humps" cannot be fixed with DSD either.

    From the Rigol plots, all the DACs seem to put out some ultrasonic dirt.

    1. Would be really interesting to see measurements of the RME ADI-2 with DSD256 versus PCM768, using the same HQPlayer filter, since the RME feature 'Direct' DSD conversion.

      Is there a way for @Archimago to do this with his RME DAC/ADC? Or would he need a 2nd unit?

      Otherwise @Miska, do you have similar plots of the RME ADI-2 fed DSD256 versus PCM768, via HQPlayer?

    2. I've posted results already at I cannot post any pictures here.

      Running the ADI-2 at 44.1k leaves first image around 352.8k at around -50 dB level (about 8-bit accuracy). Which is maybe below the Rigol resolution shown here. Running it at 705.6k drops the image around that frequency to around -80 dB. Still not perfect reconstruction (only to about 14 bit accuracy), but a lot better.

      Running it at DSD256 using poly-sinc-xtr and ASDM5 gives noise hump peak at around -85 and absolutely no images visible (of course because they are all below -240 dB which is about 40 bit accuracy).

    3. Digital filter in ESS Sabre used in Oppo has 64 taps and to fit the filter in that they need to limit the stop-band attenuation of images which in ESS Sabre is -120 dB, equivalent to about 20-bit accuracy for it's computed results.

    4. @Miska

      Regard: "But due to how ESS Sabre works with DSD (it is not direct conversion) some of the ESS Sabre "wandering noise humps" cannot be fixed with DSD either."

      I wonder why do the manufacturers of the two most popular DSD A-to-D's (Mytek and Merging?) prefer to use ESS for their D-to-A converters?

      They must prefer it for technical performance, over other DAC chips? Even those that do 'direct' conversion from 1 bit SDM to analogue?

    5. I don't know why or how they choose things. At that point when they picked up the chips, AKM probably didn't yet have the new chips with DSD256/DSD512 capabilities. The TI PCM1795 works nice when running at DSD, up to DSD256, but it is not so nice with PCM inputs. So you need to pick your poison if you go with COTS DAC chips at the time you select the chip.

      The older series they used didn't have it when synchronously clocked (like Mytek Stereo192-DSD DAC). The newer series seems to have it. Probably you cannot anymore buy the older ones though.

    6. Thanks for the posts Miska on Audiophile Style.

      I'm certainly curious myself whether the AKM way of dealing with DSD being more direct with the "volume bypass" feature results in a significant difference compared to measurements of the ESS. Alas, while I can't use the RME ADI-2 Pro to measure its own DSD output, I can use my old Focusrite Forte to see.

      It'll at least show major discrepancies and noise differences...

  9. "FFT (65536 points, Blackmann-Harris windowing used)."

    I've always wondered what the 'correct' FFT settings should be when plotting audio data. What drove your choices here, if I may ask?

    1. Good question Steven!

      The main reason I included that information was so that others can try their hand at reproducing the settings used; not necessarily that I put too much thought into it :-).

      There are a number of resources out there to discuss various idiosyncrasies like "side-lobe rejection", changes in the size of the peak, and the varying abilities of windowing to affect adjacent "bins" which I'll link below. Various compromises between the functions and basically it ends I think with giving the options a try. Blackmann-Harris is similar to the commonly available Hanning and Hamming variants; a good general option I've seen used commonly...