Over the past few months I have reviewed several DACs, all of which are DSD capable. Not all of them, however, are what I would consider a 'native' DSD DAC.
'Native' DSD DACs have some differences in their various implementations, but common to ALL of them is digital to analog conversion via an analog filter at the end of the signal chain. I think this rather simple fact is where a lot of people get hung up; however, it is this fact that makes 'native' DSD conversion what it is; an on/off square wave bitstream that reveals the musical signal when band limited by an analog filter.
Unfortunately, this is where what should be an easy conversion to analog becomes somewhat difficult. The DSD ultrasonic noise monster rears its ugly head. The best analog designs will follow the conversion filter with a secondary filter to better manage the noise, because the initial analog filter is virtually always of the 'moving average' type, and is designed to be at its best in the time domain, which is exactly what DSD is - a time-splicing format. Making a filter with lots of taps, or using unequally weighted taps/switches are but a couple of the means used to improve the frequency domain performance, however a balance must be found between time and frequency domain, lest the DSD superior time domain performance be lost. Compromise too much for the sake of frequency domain, and the lines between DSD and PCM really begin to blur.
Finally, one must deal with how any remaining ultrasonic noise affects the 'in-band' audio. This is no trivial matter, because even after the best filtering choices, there can be and there is enough ultrasonic noise left to cause artifacts in-band, exactly where we don't want them! These ultrasonics can and will cause intermodulation distortion, idle tones, and linearity problems.
It is no wonder that the Schiit Audio company in its earliest days made a DSD 'only' DAC because of these various issues. The analog output stage must be optimized for DSD, and here we are with DACs today that use a single analog stage that must somehow effectively passthrough both PCM and DSD formats! Schiit's valid concerns aside, there are indeed many DACs today with a single analog output stage that are very effective passing true 'native' DSD and 'standard' PCM. However, the very issues mentioned by Schiit are exactly why some DSD compatible DAC's do NOT have a 'DSD direct' mode. These 'non-native' DSD converters can optimize the DSD input, add in digital filtering, volume controls, and can re-modulate with their highly linear, high resolution multi-bit Delta Sigma outputs where all is then ready-made for their optimized 'one size fits all' analog output stage. This solves many problems and adds back conveniences that do not exist with 'native' DSD. And indeed does sound quite good if not outstanding in the end. But it isn't 'native' DSD, although many will argue that point. In reality what these DAC types do is actually easier than 'native' DSD strictly defined. It is therefore ironic that the format that can be converted more simply than any other, in the end can become so very difficult.
So why bother with the crusade for the simplest 'native' conversion?
Because: when done well, in my opinion no other other method of DSD conversion sounds better.
Furthermore, if you have been bit by the bug of software conversion to DSD via Roon or HQPlayer, a 'true, fully native' DSD DAC is a MUST. One is especially lucky if he or she has available to them a build of the open hardware Signalyst DAC made of completely discrete parts, that will compete anyday with such commercial discrete converters as those made by DCS.
The DACs I have had in the Euphonic Review lab which convert 'natively' are all iFi DACs with the Burr-Brown DSD1793 chipset, the S.M.S.L D300 DAC with the latest ROHM BD34301EKV chipset, the RME ADI-2 PRO FS R Black Edition AD/DA with the AKM4493 'pre-fire' chipset, and the Signalyst DSD only converter with proprietary chipless converter, which will be looked at at a later date in its own dedicated thread.
All of these DACs use a certain variation on the same theme. The 1-bit signal at the input is completely unadulterated until the conversion filter. These conversion filters are 'moving average' filters, with the slight exception that many of them will use unequally weighted taps/switches that allow for change in frequency cutoff while maintaining the same bitstream speed. A great example of this is the DSD1793 used in the iFi converters. Although iFi in recent years chose to software lock-in a single DSD filter, changed only by bitstream speed, the DSD1793 as implemented in their legacy products can change frequency cutoff regardless of bitstream speed. I have no direct knowledge of how the remainder of these chips select their frequency cutoff and rolloff, but simply observing their behavior, I would say most of them are doing something similar. (The Signalyst, as best I can observe, which will be covered in greater depth in its own future review, uses only equally weighted switches and frequency cutoff can only be changed by a doubling/halving of the bitstream speed.)
Regardless of the finest details, all moving average filters are a form of FIR (Finite Impulse Response) filter, and have a delay line, taps, and accumulator. The 1-bit stream is converted into parallel streams, offset by one sample in time. The taps are the output bit-switches, regardless of whether or not they are resistor or capacitor based. This produces an analog current output that when summed together (and converted from current to voltage) is the filtered analog audio signal. Further filtering can (and likely should) be done by a following analog RC filter to remove more of the ultrasonic noise without having to resort to a steeper roll-off by the initial conversion FIR filter. This allows for higher quality analog output stage performance. This can be even more significant with higher DSD rates, which will have by nature a higher filter cutoff and can possibly allow, perhaps counterintuitively, more ultrasonic contamination in-band. With a double filter arrangement, ultrasonic noise is dealt with in a effective and clever way; the secondary analog RC filter will not add any ringing and allows the primary filter to have gentle roll-off characteristics.
I have made a project of measuring every 'native' DSD DAC I have in my current inventory at Euphonic Review. The results are quite revelatory, and actually in some cases reveal the disconnect that can exist in what we hear and how a device measures. Even the worst measuring DSD signals still sounded very, very good. Would it surprise you that the worst measuring signals often are at DSD512 and higher? That the 'sweet spot' by my best deductions poring over the data MAY be at DSD256? At least in my small sampling of DACs. However small the sample, they do represent a wide range of implementations and hardware. For example, we have the new ROHM chipset, which I don't think S.M.S.L did for its DSD performance the most favors in the output stage. Nevertheless, it still sounds very very good, but certainly could be made to sound even better. Also represented is the DSD1793 which carries the same tech pioneered by Burr-Brown in the mid 1990's for the then upcoming SACD format. AKM is here in the form of the AK4493 DAC chip, with a truly excellent analog implementation by RME. Very, very impressive. And finally, one of the most active companies in the DSD software sphere, Signalyst, is represented in a build of its open source fully discrete moving average filter DAC. (Signalyst results to come at a later date.)
Lets begin with the S.M.S.L DAC. It is the most 'tunable' of the DACs here, with 3 filter choices. However across all possible DSD rates, those 3 filters offer 3 different cutoffs per rate, for a total of 6 possible frequency cutoffs. (Remember, they overlap, adding only 1 new, higher cutoff per rate.) This comes in handy, because the greatest DSD performance swings come with the S.M.S.L DAC.
The following charts are sorted by ACTUAL FIR filter cutoff rate. NOT by the cutoff as labeled on the DAC. For instance, if the Data Rate is DSD128, the filter MARKED at 13khz is actually running at 26khz. If the Data Rate is DSD256, the filter MARKED at 52khz is actually running at 208khz. If the Data Rate is DSD512, the filter MARKED at 13khz is actually running at 104khz. Got it? It may be confusing at first, but once you grab hold of the concept, it's no big deal. DSD 64 cutoffs equal 13, 26 and 52 khz. DSD 128 cutoffs equal 26, 52 and 104 khz. DSD256 cutoffs equal 52, 104 and 208 khz. DSD512 cutoffs equal 104, 208 and 416 khz cutoffs.
So in theory, DSD128 played back with Filter 1 (13khz x 2= 26khz) will have the same filter profile as DSD64 played back with Filter 2 (26khz as marked). Indeed, they do have the same filter profile on the spectrum analyzer, but DSD128 has an advantage. The rise in ultrasonic noise starts later, therefore more of the noise is eliminated by the filter, meaning less distortion 'in-band'.
SINAD is the measurement I have chosen to show these changes in fidelity, and indeed doubling the DSD rate while keeping the filter characteristics the same causes a notable increase in SINAD performance, from 96db at DSD64 (26khz filter) to 99db at DSD128 (26khz filter). As far as SINAD in general is concerned, once the 100db SINAD mark is reached, this is an excellent result. As a matter of fact, the DSD1793 used in iFi products is rated to max out right at 100db SINAD. It is true that the latest, most state of the art chips from ESS, AKM, ROHM, etc can well exceed 100db SINAD. However once you get past the 90db SINAD mark and approach the 100db SINAD mark, you are in true high fidelity land. Beyond that we are in 'space cadet' category where performance metrics start becoming more academic and less practical, especially as we are starting to see SINAD breach the 120db line and steadily marching to the 130db benchmark.
Thankfully the S.MS.L D300 offers many choices of filters, especially if you rely on computer software oversampling to maximize your performance. No, this method of software oversampling is NOT native DSD per se, but it IS the way to get the best performance out of this DAC. I do not hold the ROHM chipset at fault here; I believe the fault is with the S.M.S.L analog implementation. Every other DAC tested here has DSD and PCM performance that match very closely. The S.M.S.L D300 goes down the path of truly exceptional PCM performance, combined with 'merely' decent to very good DSD performance, depending on the oversampling rate and cutoff frequency. I believe the observed phenomenon here is due to the lack of a well implemented secondary RC analog filter, or perhaps none at all. I could be incorrect, and am open to correction, however, something in the analog stage is preventing DSD from reaching its fullest playback potential.
However as was mentioned earlier, there is a notable drop in performance in DSD performance as compared to PCM with the S.M.S.L DAC. SINAD performance with PCM files approaches and exceeds 114db. The best DSD performance, 104db SINAD, is at DSD64 with the 13khz filter, which is quite too low a cutoff for high-fidelity audio. Indeed, the filter is of the very gentle moving average variety; nevertheless by 20khz the rolloff has exceeded 8 decibels! This will cause a very audible treble rolloff. For DSD64 files, surely the 26khz filter and its 96db SINAD is the minimum filter of choice. A far cry from the excellent 114db PCM SINAD, but thankfully it is still good enough that any loss of enjoyment is likely all in your head, because I told you so! If you had never read these measurements, I would be quick to bet if you listened to the D300 you would have heard nothing negative at all. Ah, the power of simple suggestion....
And as we stick to the same 26khz filter with 128x DSD rate (by selecting 13khz DSD filter from the menu), we see that there is a noteworthy performance increase up to 99db SINAD with DSD128 due to the ultrasonic noise rise starting much farther from the in-band frequencies (20hz to 20khz), pushing the noise to a higher level where the filter can more effectively eliminate it. Although DSD64 had the very best SINAD measurement of any at 104db, again the filter rolloff is too low to be considered as high resolution audio. S.M.S.L D300 DSD playback is MUCH improved with the DSD128 and DSD256 formats. I see no need to oversample to the DSD512 rate. The most obvious reason being the poor performance at the two highest filter settings. The next reason? There is no SINAD improvement at all over the DSD256 at 104khz filter (26khz labeled in GUI). It is simply a waste of resources to oversample to DSD512 in this case, since you will not see any performance gains. The filter profiles are the same, meaning you will likely not see any gains at all in the time domain, and the noise/distortion performance in the audible band is identical in both. Here we may be running into a very common issue in DACs and 'ultra high speed' DSD.. the ability of the logic itself to keep up. Faster is NOT always better, and it looks to me that the sweet spot for the S.M.S.L DAC DSD converter is DSD256 at either 52khz or 104khz filter. I personally would go with the 104khz filter, to take advantage of the temporal resolution DSD offers while still having good enough SINAD performance. I can totally understand the thinking that would lead one to choose the 52khz option and get that SINAD up close to 100db. Either way, the DSD sweet spot for the S.M.S.L D300 is DSD256, and choose either filter 1 or filter 2 on the GUI (13khz, 26khz).
As I mentioned before, I prefer 'native' DSD at all sample rates, what one might call 'bitperfect'. However, the S.M.S.L D300 is all over the place DSD performance-wise. I will reiterate what I believe is the source of this 'issue', and that is a lack of a secondary analog RC filter following the DSD FIR conversion filter. Yes, one can achieve 'good enough' performance without a secondary filter, and one can 'tune' in the best performance with multiple DSD filter options available on the D300. However, I think the added expense of a secondary filter (or a better performing one) would have been worth it. Bad call here by the engineers. I am on record calling this little DAC a 'giant killer', and this is still especially true for PCM file playback. Its native DSD playback, in spite of any compromises is still quite good, but as you see the performance of the next couple DACs it will become obvious exactly how good the D300 COULD have been.
Now we move onto iFi, where thankfully we can get the temporal resolution promised by the DSD format.
Unlike the 'dac du jour', Thorsten Loesch and iFi didn't just stick with whatever is currently 'popular', such as the latest ESS or AKM chipset. There is more to overall performance/sound than sub-arctic THD and/or stratospheric SINAD. There is still something to be said about overall architecture, in this particular case the Segment DAC architecture of the Burr-Brown DSD1793 chipset. The DSD1793 uses a 64+ level thermometer/unary code DAC, which allows for the top 6 bits of the 24 bit two's complement PCM to be converted directly via bit-perfect PCM (64 levels), with the remaining 18 bits are converted via a 1 bit Delta Sigma converter. This is a particularly ingenious means of conversion, as it avoids the so called PCM zero-crossing error by keeping the 'lower' 18 bits always at "full volume". Obviously it isn't at full amplitude in the time domain, however in the frequency domain zooming down to the sample level, yes, the 1-bit signal is either full on or full off. The chip will use oversampling/time-splicing for the lower 18 bits of resolution. As has been mentioned, according to Burr-Brown, SINAD is expected to be right at 100db.
Another benefit of this type of conversion is it is ready made for 'native' DSD conversion. The 64 level thermometer/ladder bit switches are all unary coded/equally weighted. The delay line for the DSD conversion is only 8 bits long, however. This allows several different combinations of the 64 switches to create 8 groups of switches that function as the 8 unequally weighted taps. These different combinations can mix and match to change the cutoff frequency and the order (steepness) of the frequency rolloff. All the while the 64 individual switches are equally weighted and allow for dynamic element matching and exceptional linearity.
Note this excellent SINAD consistency continues within the iDSD PRO. At all bitrates and their accompanying cutoff frequency, SINAD measures right at 99db.
In addition to the DSD64, DSD128, and DSD256 rates shown above,
at DSD512 the cutoff frequency is 616khz, and the SINAD stays a consistent 99db.
At DSD1024 the cutoff frequency is an astonishing 1,232khz (1.232 megahertz), and still maintains a SINAD of 99db!!!!
As you can see, the iFi products above have output stages that offer equivalent maximum performance via DSD and PCM. (The DSD1793 chipset, repeated here like a broken record, is limited to around 100db SINAD.) The iFi ZEN and iFi iDSD PRO are not configured to have switchable DSD FIR filters, however, their filter cutoff doubles with each oversampling. For example, DSD64 has a -3db cutoff at 77khz. Anything at DSD64 will always have a cutoff at -77khz. Double the speed to DSD128 however, the cutoff will double to 154khz. And this pattern continues right up to DSD1024 on the iDSD PRO. (Note the DSD1793 does indeed offer switchable cutoffs for each DSD speed rate. In recent years, however, iFi has settled on locking into what appears to be the 77khz FIR filter onboard the 1793 chipset, along with an 80khz Analog RC filter to follow.)
And finally in this particular blog entry we will take a look at the very, very impressive RME ADI-2 PRO FS R BLACK EDITION with the AK4493 chipset (pre-factory fire switched capacitor version.) Just as the iFi with its DSD1793 matches PCM and DSD performance, likewise this much more modern AKM chipset matches its PCM and DSD performance. This MAY be the highest performing true native DSD chipset on the market that is NOT made from expensive discrete parts ala the Signalyst Converter.
Note that the following AKM 4493 based DAC boasts an outstanding 114db measured SINAD with PCM material. The 'native' DSD material comes very close to this level of performance. One of the first things to note, however, is the DSD filter numbers in the GUI are incorrect. They are not 50khz/150khz. This was correct for the previous chipset used by this AD/DA, the AKM4490. The AKM 4493 used in this particular upgraded AD/DA uses DSD64 filters at 39khz/76khz.
What more can be said? Used in DSD Direct Mode (Native DSD), DSD256 using the 'high' filter, in this case at 304khz, while maintaining a 111db SINAD, is truly remarkable performance. This is an astonishingly good choice of DAC for those who use software oversampling. Oversampling in Roon or HQPlayer to DSD256 with the higher of the two filter settings (304khz) with a SINAD over 110db should offer state of the art performance, and then some.
Also, if you are more of a purist, as I am, and wish to have as little DSP touching the DSD stream as possible, The RME ADI-2 PRO FS R BLACK EDITION is also an excellent choice. You can have native rate conversion with every filter save one between 110db and 112db SINAD. The only exception, is the high 76khz filter on DSD64 material. This is not unexpected, as there is a significant amount of ultrasonic noise that escapes the lowpass filter, yet the actual performance penalty is very, very small, as the SINAD is still a high 107db!!
The RME team have built an exceptional native DSD DAC. On an aside and maybe for a future entry, I will publish the RME results with DSD in Volume Control Mode and remodulation to AKM's multi-bit Delta Sigma format. The performance here, even with DSD, is also quite outstanding. The RME continues to impress. Full review coming sometime this year.
MORE TO COME SOON, INCLUDING THE FULL REVIEW OF THE iFi IDSD PRO (part two of the review; part one reviewing the ifi iDSD NEO is currently posted. )
Also a full review of the Signalyst PURE DSD DAC is coming soon! After much listening and mixing/matching amps and headphones, the Signalyst Converter may indeed produce the very best sound I have ever heard.
The S.M.S.L D300 I reviewed was loaded with firmware 1.0. For a time, a short time, a firmware update to 1.1 was uploaded that enabled DoP (for DSD playback for those who do not have ASIO, such as on Macintosh.)
After a late night discussion with a fellow enthusiast, and many measurements, followed by more research, it looks like the S.M.S.L firmware update has been a real nightmare for some people, so bad I believe it has degraded the performance of this DAC overall. Some have even 'bricked' their D300.
Version 1.0? Yes, a GIANT KILLER DAC and a real gift to those who love native DSD conversion via analog filter. I stand by my review and findings. However, it sounds like newer firmware created a nightmare. That firmware was periodically available from the S.M.S.L website. Yes, it came and went, came back and left again. As of now, the firmware update is NOT available on the company website. It CAN be found if you look to the forums, but my advice? Stay away! If you have a version 1.0? Keep it and treasure it.
If you want to buy one? Make sure it comes with firmware 1.0! New ones did indeed ship with the updated firmware that, again, is no longer even available from the company. Bad, bad situation, and I am watching closely to see if customer service and engineering can work this out.
linked below are the user experiences. This is linked to the final page of thread, but if you want a fuller picture, go back about 10 pages and start there.
In my previous post, linked here, I noted that I was using a generic, very cheap variable Switch Mode Power Supply to power the DSC2 Converter temporarily. The supply is meant to help evaluate headphone amplifiers and their ability to power low impedance headphones, as it has up to a 5amp output at the lowest voltages. It is NOT by any means made for critical listening, and the XPower proves that. The changes in noise floor and digital 'hash' are clear when the XPower takes over duties from the generic switch mode power supply.
Now, the exact benchmark being sought at the time (jitter, IMD, SINAD, etc.) doesn't necessarily change, although the measured results CLEARLY show improvement when using the XPower. I have seen similar results from other clean power devices, especially PSU's and USB galvanic isolators. However....
Many an absolute objectivist, especially the most vocal, will point and say 'haha!', or 'gotcha!' because the jitter/IMD/SINAD etc. distortion measured result is the same. I find this to be extremely hypocritical and only self-serving. When faced with real evidence of a difference in the measurements, it is ignored and re-directed into excuses such as, 'you would never hear the difference anyway', or as mentioned before, 'the measurement you were seeking didn't change.' Ironically, they then sound exactly like the 'opposite' camp, who will point out that past a certain point, things like higher SINAD, or lower THD are meaningless because the differences are inaudible. These particular absolute objectivists seem to want their cake and eat it too, but they just can't have it both ways. One cannot laud "product Y" for its 125db SINAD, proclaiming it king of the hill, while at the same time say the obvious changes visible from "product X" are meaningless because they are 'obviously' inaudible.
The fact remains that something major DID change in the measurements, even if it is not the metric or metrics for which one searched. Perhaps what we see IS audible or contributes in some way that when combined with other metrics is audible. For all we think we know about audio reproduction, there are any number of things we don't fully know or understand. Why? Because ultimately what comes out of any audio product is judged inside an extremely complex neural network and bio-sensory system that we are only beginning to understand. If all we do is measure all day and don't listen, thinking our measurements tell us everything we need to know, this hobby is rubbish. Just as much rubbish as a half-million dollar RCA interconnect. (Yes, I find the truth to be likely found somewhere more in the middle. The fundamentalist absolutist objectivist 'know it all', and the 'used car salesman' in the store who calls amps 'class A' because he reads the Stereophile scriptures, not knowing that Class A actually IS a type of amp topology, are equally off-putting.)
Unfortunately the scientific method has been misused time and time again by closed minds to justify their 'position', or their 'ideology'. It is too bad there are not more researchers just interested in the truth, who will follow the data wherever it may lead, and entertain hypotheses in humility rather than dismiss them in arrogance, perhaps holding back and delaying progress.
Now getting off my soapbox, back to the changes the iFi xPower made in the DSC2 measurements. I will choose to focus on two in this blog entry. Intermodulation Distortion and Jitter.
With regards to the specific measurements being sought, the specific data didn't change. But many other things did. The noise floor, once full of spurious tones and quite frankly, a bit of a mess, cleaned up dramatically. And yes, listening tests performed BEFORE the measurements had already revealed a much more natural sound, with less sibilance and harshness than before.
We will start with IMD, 19khz+20khz. The first measurement is with the Generic PSU, the second is with the iFi XPower.