A Look at ADC Noise Limits and Distortion - Or "How much difference does the Cosmos APU make?" And a look at 2-channel Room EQ Wizard measurements!

I received a few questions as a follow-up to last week's post about the E1DA Cosmos APU basically around the question: "When do I need a pre-amp/notch filter like that if I have a high resolution ADC?"

That's a great question! Let's spend some time looking into this...

Now, the first thing to be mindful of is that our DACs and ADCs are not perfectly linear devices, nor have limitless low-noise performance of course. This is why we measure the limits of DACs to gauge the extent of "accuracy". And ADCs likewise have limits as well to their intrinsic noise level and potential for distortion. As we approach the limit of the ADC's performance, it will start introducing its own characteristics into the captured signal which could look like noise anomalies as well as harmonics - like DACs, there is an ADC THD+N.

In this post, let's run some practical "experiments" of sorts using the Cosmos ADC and RME ADI-2 Pro FS to examine these ADC limits when measuring DAC dynamic range and THD+N/SINAD with and without the aid of the Cosmos APU.

As a start, let's look for hints in the data sheets of where the limits might be.

If we look at the information for the ESS ES9822Pro ADC inside the Cosmos ADC, we see that for stereo operations it's DR 124dB DR / THD+N -117dB, and in mono mode these values improve to DR 127dB / THD+N -118dB. These are nominal chip specs and in real-life we might run into "golden" versions of the chips that perform even better; hopefully not worse. (This explains to an extent why there are "Grade A"/"Grade B" and "Grade C" versions of the Cosmos ADC.)

As for the RME ADI-2 Pro FS (including the newer RME ADI-2 Pro FS R Black Edition), the ADC is the AKM AK5574 operating in 4-to-2 channel mode with DR 124dB and this improves with 4-to-1 mode up to DR 127dB. However, the THD+N spec fluctuates around -110 to -112dB depending on sampling rate (best at 48kHz).

Already these numbers give us an idea of the limits which we should be careful of surpassing to maintain reliable measurements but only at the level of the chips themselves, not in the environment of the whole ADC box with all its support circuitry.

All things being equal, assuming the ADC is incorporated as part of a good design, the Cosmos ADC's ES9822Pro should perform a bit better with the same input level. A caveat is that we should keep in mind the Cosmos ADC's lower input impedance. Previous testing indeed does show the Cosmos ADC has lower noise and is able to record/measure the input signal with greater resolution.

Armed with this information, let's see if I can use my actual devices to demonstrate the limits.

Like last week, let's compare using the Topping D10s (unbalanced RCA) and D10 Balanced (XLR) DACs. 

We'll start with dynamic range measurements with and without the Cosmos APU's pre-amp (+60dB gain setting) using a -60dBFS 1kHz tone. For ease of comparison, let's set the 0dBFS unbalanced RCA output from the Topping D10s as around -3dBFS on the two DACs, and overlay this -60dBFS 1kHz graph in REW (applying a -60dB compensation for APU measurements):

As you can see, using the Cosmos APU helps achieve a consistent measurement result from both the Cosmos and RME ADCs. Looking at the Cosmos ADC-only and RME ADI-2-only results, we can see that the Cosmos ADC does allow for a lower noise floor but the difference is <2dB.

Already, even with a modern hi-res USB-powered single-ended DAC, for accurate dynamic range measurements, we can see the value of using the APU.

Okay then, let's make it even more challenging! What if we use the Topping D10 Balanced instead doing the same "experiment"?

This time, notice that I've got 2 other measurements captured by turning on "MONO" mode with the Cosmos ADC and "M/S Proc" with the RME ADI-2 Pro FS. By doing this, we are able to improve noise level and indeed differences are measurable. By doing this, and with the DAC's 0dBFS level around -1.25dBFS into the ADC this time, the Cosmos ADC is able to surpass a result of 120dB dynamic range and the RME gets us to 118dB. As expected, compared to the chip specs sheets, I see these as the "real world" numbers of ADC performance.

Again, with the APU, readings from both the Cosmos and RME ADCs performed equivalently well, allowing us to appreciate a dynamic range for the D10 Balanced at around 123dB.

Another little detail to notice is that the RME's AKM ADC noise floor has a tendency to creep up in the higher frequencies as a reflection of the sigma-delta noise shaping. It is a small amount and not an issue up to 20kHz in any event. By amplifying the noise with the APU pre-amp, we lift it up and above the ADC's noise modulation.

Moving along to THD+N testing, from measurements using the Audio Precision, we know that the D10s measures with a THD+N of around -112dB and the D10 Balanced around -118dB. That's a significant gap in performance between the two, reflective of about 6dB between unbalanced and balanced performance (numerically at least, even if audibly the difference might be subtle at best!).

Again, starting with the Topping D10s, let me show you the actual FFT so we can appreciate the numeric results in detail:

As you can see, the lower 2 "APU + ADC" graphs are clearly measuring the same device resulting in exactly the same THD+N of -112.5dB. Without the APU, the top 2 "ADC-only" measurements are showing harmonic variations affected by the limits of the ADCs. I've put a red asterisk to indicate obvious harmonics that appear to be "off" compared to the consensus APU measurements.

Notice that the anomalous harmonic levels are all below -130dB for both ADCs. Let's keep that number in mind. When looking at a single number, the THD+N value achieved by the "ADC-only" measurements are off by about 1-2dB with the D10s.

Now let's look at the D10 Balanced knowing it's an even higher resolution DAC. As above, let's also include the Cosmos' MONO mode and turn on the RME's "M/S-Proc" to improve noise floor:

The APU's 1kHz notch filter is a great "equalizer"! Using this, both the Cosmos and RME ADCs result in almost perfect concordance for THD+N measurements yet again. With some peeping into the pattern of harmonics, we can again appreciate that there are some low-level (below -130dB) distortions that differ when performing the test with ADC-only. These again would be examples of distortions introduced by the ADCs when fed the strong 0dBFS signal.

Compared to the lower resolution unbalanced Topping D10s, we see a wider gap with and without the APU in the case of the D10 Balanced. With the Cosmos ADC, the gap is about 2dB in stereo mode, and 1dB "MONO". The RME gap is about 5dB stereo, and 4dB "M/S Proc".

In summary then: "How much difference does the Cosmos APU make?"

As expected, the difference is a function of the ADC we're using and the resolution (low noise level and low harmonic distortion) of the DAC we're measuring. Let me offer some observations:

1. As a tool to investigate the noise floor, the Cosmos APU improved precision consistently using both ADCs and in these measurements can already demonstrate performance beyond 123dB dynamic range (in my early look, I already suggested up to 145dB even without bridging the APU). With both DR and THD+N measurements, it performed very well with the Cosmos and RME ADCs. I expect that you can use it with a wide range of decent ADCs thanks to the low-noise +34/60dB pre-amp and +20/26dB 1kHz notch. Having said this, remember that DAC measurements are not just DR and THD+N/SINAD, so you'd still need a good ADC for general measurements of characteristics like distortion across frequencies, output level linearity, frequency response, jitter, etc.

2. As expected, the RME's AK5574 ADC isn't as resolving as the Cosmos's ES9822Pro when used without the APU. My suspicion is that one would get accurate results with devices up to DR 115-118dB using the RME with the higher number using "M/S Proc" mode and balanced 4Vrms input, and DR 116-120dB with the Cosmos ADC alone.

As for the distortion test, the RME likely can accurately measure devices down to -110 to -114dB THD+N, and the Cosmos ADC around -111 to -116dB THD+N. Beyond these values, the ADCs would be introducing significantly more of their own distortions. Note that these estimates are with proper measurement technique such as making sure the 0dBFS 1kHz tone from the device being tested is at least higher than -6dBFS at the ADC's input to maximize the available dynamic range and being mindful of the Cosmos ADC's variable input impedance.

3. As for anomalies of low-level harmonics when looking at the THD+N FFTs, they tend to be around -130dB or less with both the Cosmos and RME ADCs. Harmonic levels higher up will generally be accurately captured (make sure to calibrate the Cosmos ADC as best you can BTW).

To demonstrate the limits discussed with a "proof of concept", let's use a lower resolution DAC - the Topping D10 discussed a number of years back. Let's measure this using the RME ADI-2 Pro FS ADC with and without the Cosmos APU:

Here, with a lower resolution DAC (THD+N higher than -110dB), the unaided RME ADC is able to measure the device with results <1dB difference in DR and THD+N compared to the ADC+APU combination. Also, notice that the pattern of the harmonics in the THD+N graphs are essentially the same as opposed to the higher resolution D10s and D10 Balanced since we're not straining down to the limits of the ADC.

Obviously, unless we're reproducing audio for cats, dogs, dolphin, and bats, it's important to correlate objective measured performance with human auditory/cognitive abilities. As a species, our brain resources allotted for sensory perception tend to be biased towards the visual; thankfully we're not "as blind as a bat". Even within what I believe are the accuracy limits of the RME ADI-2 Pro FS at around DR 115dB and THD+N -110dB, would I argue that a measurement device hitting such a limit would not already be "perceptibly perfect" (assuming other factors like frequency response checks out)? No, of course not! Seriously folks, big numbers are nice but their benefit is more for "bragging rights"; part of the competitive nature especially for us men!

As expressed even recently, assuming other tests like frequency response are OK, a THD+N of -100dB (SINAD 100dB, 0.001%) would already be great; I strongly doubt blind testing would reveal an issue even with an army of trained young 20-year-olds listening to actual music at normal volumes. The resolution limits of amplifiers, headphones and speakers are orders of magnitude greater than this (not to mention normal room acoustics and ambient noise level). Nonetheless, it's good to be as accurate as possible when running measurements especially these days as DACs extend to better than THD+N -120dB! Very nice that the Cosmos APU has "democratized" the ability for home audiophile hobbyists to achieve even greater measurement accuracy to keep up... As I said, it's about bragging rights now. And IMO, Ivan has a lot to be proud of at this kind of price point with his E1DA products!

One last thing about the hardware, as published previously, the Audio Precision APx555 B-Series uses dual AKM AK5394A ADCs for its detailed measurements up to 96kHz (24-bits and maximum samplerate of 192kHz). DR on that chip is spec'ed at 123dB with THD+N -110dB. Like what we're doing here pairing the APU + ADC, it's not so much the ADC chip itself that can peer into such low noise levels, but rather the techniques employed around that ADC to enhance low-noise performance.

--------------------

To end off, over the last 2 weeks I've been looking at pre-production hardware in the form of the E1DA Cosmos APU. Well, here's a quick look at currently pre-release software!

The other day, I E-mailed John Mulcahy, the author of Room EQ Wizard wondering if there would be a way to perform dual-channel measurements using his software. Lo and behold a few days later he sent me a beta version 5.20.8 to test out with dual-channel features. Note that Room EQ Wizard Pro upgrade for multichannel (US$100) is needed for the feature I'm showing here.

Turn on "Multiple inputs" in the input selector (white double arrow icon) in the RTA window; here's an example about to look at 0dBFS THD+N:

Notice I'm applying +30dB 1kHz notch compensation to the left channel which is receiving input from the APU.

And here's a look at 1kHz 0dBFS from the Topping D10 Balanced into the Cosmos ADC, left channel (blue) going through the APU 1kHz notch filter, and the right channel (red) directly into the Cosmos ADC:

Here's a close-up of some results:

Nice! Realtime monitoring of the 2 channels showing the improved noise floor thanks to the APU. This will be helpful in the days ahead to simultaneously capture stereo output. [Note that the performance from the 2 channels of DACs and ADCs are not exactly the same so I'm not expecting identical results here, just showing the kind of magnitude differences one sees using the APU and without.]

Not only can we look at 2-channel FFTs concurrently, but multichannel stepped sine measurements can be done:

In the example above, I'm running an amplitude stepped sine measurement from -140dBFS to -60dBFS, left channel going to the Cosmos APU's +60dB pre-amp, and the right channel straight to the Cosmos ADC.

With that data we can have a peek at the DAC's output level linearity all the way down from -60 to -140dBFS:

These days many DACs already can essentially achieve near-perfection down to -120dBFS which is the case with the Topping D10 Balanced. -120dBFS is usually where I see Audio Precision measurements end. We can see the difference the APU makes in helping to extend the accuracy of very low levels down to -140dBFS. I'm certainly not saying that we need to measure linearity down to the 23rd bit of course. But it's cool I think to look down that far at the vibrations of electrons! ;-)

To have REW be able to capture the data in both channels in a single pass allows the process to be done in half the time. Thanks John also for adding the linearity normalization option to the software feature set.

Hmmm... With the ability to measure both channels concurrently, I think I need to get a second Cosmos APU now! ;-)

Lovely seeing progress made to substantially empower the home objectivist interested in performing his/her own measurements, both in the hardware and software domains. Neither of which requiring any huge expense.

Alright friends, now I really really need to get back to work instead of playing with these audio toys! Hope you're enjoying the music.

Addendum - June 26, 2022:

It took me a bit of looking around and asking John M. eventually to figure out that if I need to apply calibration to Stepped Sine measurements, there's a setting in Preferences / Analysis to tick off. This is useful for example with the Cosmos APU to apply the 1kHz compensation: