What would be the lowest noise connection from a balanced output dac into the QA403?
If I take both signal leads into a single input say L+ from one channel of a Topping D10 Balanced I’m getting thd of -129db and thd+n of -112db.
When I connect this up to an E1DA adc using the same adc chip as the QA403, I’m getting thd of -126db and thd+n of -118.5db. The connection in this case is xlr balanced out from the dac into xlr balanced input on the E1DA.
Is there a way I can get close to these numbers on the QA403?
The review on AudioScience forum is showing thd+n of -118.5db for this dac.
Hi @Moto, I didn’t know Topping had a balanced DAC. I just ordered to try.
To pencil this out a bit, Amir’s measurement on THDN was achieved at 4.2Vrms, which is 12.5 dBV. That means each leg is driven at 2.1Vrms which is 6.5 dBV on each. The QA403 has inputs of +6 dBV or +12 dBV.
So, I’d first go in single ended at maybe 1 dBV below 6 dBV on the +6 dBV input range and see how that looks (which means D10B out+ into QA403 in+, with in- shorted). Take a look at N-D and also 2H and 3H, and adjust up or down a bit to improve the number. Once you’ve found the optimum, connect the out- to the in-, and you signal should increase 6 dBV but your noise will only increase by 3 dBV. For that reason, click to the +12 dBV range when in balanced mode. Also, when you go balanced, the 2H should disappear.
Some things to be aware of:
Short the inputs on the QA403 and measure the noise on the +6 dBV input range. That noise will probably be around -113 dBV or so (rect, 128K fft size). That is means the noise is about -119 dB below the full scale. That is the best your THD can be considering the noise. In practice, it will degrade as the ADC noise rises as input rises.
Next, using the QA403 generator, start with the generator at 5 dBV, and the input range at +6 dBV. And then slow sneak up (using alt+click on the AMP1 button) until you hear the overload relay start clicking once per second. And then back off 0.1 dB. On my unit, note that I’m in the +6 dBV range, and yet the overload point is 7.3 dBV. The relay will click when overload is within 0.5 dB of dBFS. Now, with this overload point established, set that 0 dBR. And then do an RMS dBR.
On my unit, that gives -120.15 dBr. The point here is that even though the spec max input of the QA403 is 6 dBV, the real max input for that level is about 7.3 dBV. So,if you are trying to squeeze all you can squeeze, then in some cases your input signal could be +6.1 dBV on the +6 dBV range.
I will check out your suggestions. Thx.
Just to clarify, when you say, “ That is the best your THD can be considering the noise.”, you mean that the best thd+n can be is -119db, correct?
Yes, that’s correct. But that is the theoretical best. In other words, if your noise is -100 dBFS with no signal present, then the best the THDN can ever be is -100 dB. But in reality it will be worse because the ADC noise floor will rise when signal is applied. Between -infinity at -5 dBFS, the noise rise will be nearly zero. But as the input signal rises from -5 dBFS to 0 dBFS, the noise floor rises.
You can best see this in a typical converter plot as shown below. Usually, we look at THDN as a ratio and the result shown relative to the input signal level (eg “noise and distortion are 100 dB below the carrier”. But the ADC and DAC folks usually show the noise and distortion as a level and you get a plot such as below (taken from the AK reference design spec).
When you see a plot such as above, it should be clear what is happening: You start with just the noise, and the smallest of all signals present. And the noise stays constant as the signal level increases UNTIL you get to -10 dBFS or so. And at that point, the noise and distortion start to rise.
And so, from the AK5397 plot, we see the noise+distortion is -120 dBFS when you have a -20 dB dBFS input signal. So the THD+N will be -100 dB. And the noise is still -120 dBFS when you have a -10 dBFS, so the THD+N will be -110. But from there, it degrades. At -5 dBFS, the N+D is -110, and the signal is -5, so that gives you THDN of -105. And at -1 dBFS input, the N+D is about -100 dBFS, and the input is -1, so that is -99 dB THDN.
Another good rule of thumb is the level of the harmonics. So, if you the RMS noise you measure is -110, then your harmonics need to be about 6 dB below the RMS noise otherwise the harmonics will degrade the THDN. So, if you measure an ADC noise floor with inputs shorted, and it returns -120 dBFS RMS noise, then that tells you the harmonics need to be around -126 dBFS when you are measuring near full scale otherwise the harmonics are degrading the THDN.
So, it should be pretty easy to look at a baseline noise plot and determine what the best-case THDN could be. And when you look at a THDN plot, it should be easily to quickly tell if it is limited by harmonics or noise.
Hi @Moto, I spent some time with the D10B, and here’s a quick summary:
The D10B is set to -1 dBFS, 1 kHz, and these are balanced measurements.
First, take a look at 24 dBV full scale input. The 2H and 3H are -129 and -135 dBc, but mostly in the noise:
Now switch to 18 dBV full scale input. Here we can see the harmonics clearly (due to reduced noise floor), and the levels are at -135 and -142 dbC. This is probably a clearer read than the 24 dBV full scale input, because we can clearly see them and they aren’t competing with the noise floor to be measured. The THD is shown as -130, which looks legit because that’d roughly be the sum of -135 and -142 plus some of the other stuff we see at higher frequencies.
In the plot above, note the N-D is -99.59. That means the noise (20 to 20 kHz) by itself is -99.59 dBV, which means the best case THDN can be 11.59+99.59 = -111.18. We’re about there at -110.44. So, at this input level (+18 dBV) the noise floor is limiting the THDN.
Now switch to +12 dBV full scale input. Note the noise floor has dropped to -102.29 dBV (2.7 dB better than +18 full scale input). But the 3H has come way up as the input stage starts to overload.
I think there could be some wins here from enabling the harmonic cancellation on the ES9822 ADC, and possibly switching to mono mode. I’ll try some experiments and see.
Thx much Matt! So the firmware is such that you can set that?
By the way what set up did you use for balanced out into the qa403?
@matt I duplicated your results.
Then I left the signal generator in REW exacty the same and input it into a cosmos e1da that has an xlr input using the 4.5v setting. These use the same adc chip but have no input buffer but have had harmonic correction done at the factory.
Here are those results showing some better thd and reduced noise possibly due to the connection and cabling differences.
@matt I worked on connections and at your peak dbv got this. Why is my thd so much better than yours?
Increasing the voltage to 12.34 peak dbv at the 12dbv input level seems like it should have worsened thd but it didn’t. Why? Just better results.
Hi @Moto, it’s hard to know why your THD is better. Likely, there is an inherent THD associated with each ESS part, with some being better and some being worse than others. And then there’s a phase that is another random variable. That is, if the DAC has a 2H at 0 degrees phase, and the ADC has a 2H at 180 degrees and the levels are comparable, then they will cancel. Or if both have the 2H at 0 degrees, then they could add.
In short, the old adage that your tester needs to be 10 dB better than your DUT is very true. If they are comparable, then you will have a lot of variation. I’m not sure a $30K tester can test this DAC with enough margin to know for certain where the DAC ends and the tester begins. You’d need a tester with -130 dB THDN, and those don’t exist at any price. Or, a very, carefully designed notch and likely a JFET preamp.
But congrats on your measurement! How sensitive is that to things like position and cabling?
@matt The worst I have been able to make it measure is to put the dac on top of the qa403 then put them both right next to my laptop which is plugged in to the charging brick. That is thd-132 and thd+n -114.9. That is at 12.34 peak dbv
At 11.5dbv in that same position it measures thd -131 thd+n -114.7.
So you think that the harmonic difference then would have to be my qa403 having phase difference in the 2nd and 3rd harmonics that are cancelling those of the dac?
Also seems weird that performance is better at peak dbv 12.34 into the 12db attenuator on the qa403 unless it exacerbates the cancelling harmonics in the qa403 even more.
My connection is one channel out from the d10b split into 2 bnc connectors into the + and - L channel of the qa403. Is that the proper approach?
Hi @Moto, yes, I think your approach is correct. I’m not sure on the 2H and 3H. But it seems your measurement is pretty robust eg is not showing a lot of sensitivity. You could swap L+ and L- connections. In that case, that would upset whatever “lucky” cancellations you were seeing. And if the results change appreciably, then it was due to luck. If the results don’t change, then it’s pretty solid.
You were right. It was just luck. Thd was down to 126 from -132 worst case before( more randomness in the environment) and then I swapped L+ and L- and it went to -121. N-D stayed highly constant at around -104. Given the amount of randomness at these harmonic levels, how do we ever really trust anything?
Hi @Moto, that is why your measuring equipment needs to be 10 dB better than your DUT. When that is true, flipping the inputs won’t make a difference. But if your DUT and Analyzer are close in performance, then sometimes flipping the inputs will make things better and sometimes it will make things worse (harmonics will change, but noise won’t)
In the plot below, I ran the D10B into a balanced to single ended converter based on OPA1612 buffers, and diff amp using ORNTA1001AT1 thin film array (0.01% matching). That then goes to a passive single-ended notch made with 732 ohm R and 220nF NP0 caps, and then to a OPA1612 notch buffer with 18 dB of gain.
With 18 dBV full scale input and the inputs to the above circuit shorted, this gives:
Note that the RMS (20 to 20 kHz) is -115 dBV, and with 18 dBV full scale input, this suggests a dynamic range of 133 dB and THDN of -133 could be measured. In practice, however, the upper end won’t be useful because there will be clipping. However, the notch ensures we won’t clip.
With the D10B connected, we get the following:
Note that the noise floor has risen about 6 dB (meaning lots of margin to accurately measure the D10B noise floor). The gain of this can be a bit tricky to sort out. But with the D10B set to -50 dBFS (and keeping the 18 dBV input gain specified) and the notch bypassed, the measured 1 kHz tone is -38.94 dBV. And so, with the output set to -1 dBFS (49 dB higher), we’d expect the output to be -38.94 + 49 = 10.06 dBV. Restoring the notch gave us the -43.67 dBV peak, which is 10.06 + 43.67 = 53.73 dB below the expected value.
And remember, a TwinT will have 9 dB of suppression at 2H, and 3 dB at 3H
So, what do we know:
1H = 10.06 dBV
2H = -131.83 + 9 = 122.83 dBV = -132.89 dBc
3H = -122.88 + 3 = -119.88 dBV = -129.94 dBc
N-D = -107 dBV = -117.06 dBc
And so, I think the THDN of the D10B is dominated by noise, and is -117.06 dB, with 2H and and 3H both >10 dB away from that figure. This measurement is insensitive to swapping inputs. Averaging and a 512K FFT would buy you another dB on the noise.
And so, a credible case can be made that the THDN of the D10B is -118 dB, and that we had about 6 dB of margin when measuring that.
The circuit I described above also has an 18 dBV attenuator at the front end. When that is engaged, that gives 18 dB of atten followed by a diff to single conversion, followed by notch, followed by 18 dB of gain (net gain of 0). A plot of that is below:
Note that the noise is destroyed (due to the atten). But the harmonics improve a bit. This suggests some of the harmonic growth measured in the previous plot (not the one above) is coming from the added circuit. But this should make us even more confident that the THDN is dominated by noise.
Finally, knowing that this is dominated by noise there are a few more things to check: First, larger FFT, use 0 dBV full scale, and average 10 times:
Second, we can use cross-correlation to double check the noise isn’t related to self-noise of the QA403. Cross correlation requires that you run the DUT input to both the left and right channel. Then, from the RUN context menu, see the cross-correlation tab. What this does is average away differences that are applied to the L+R channels (due to unique channel noise), but noise and signals that are common (from the DUT) will be made stronger.
After a bit of running (183 acquisitions) we get the following:
Note above the N-D is about the same as above. And this should be expected, because our measurement noise floor is about 7 dB below the DUT noise floor. The cross-correlation confirms that.
In any case, I think a notch is important for measuring and being certain you have the margin.
@matt I can’t thank you enough for your patience, expertise and clear articulation of these issues. I will attempt to see how close I can come to duplicating your results with the twin-t notch and opa1612 amp I built earlier( need to finish balanced to se converter).
I think I speak for many of us when I say we look forward to the day that we can purchase more of your products!
Hi @Matt. I have a curiosity about this screen shot. It seems to me that no signal is applied to the input, so what you see is all noise.How come the measurement RMS and N-D are different? Should not they coincide in these conditions (numerically I mean)? Where am I going wrong? Thanks for the clarification
Hi @Claudio, the difference is due to the fact that the RMS undergoes an energy correction (to account for windowing) while the N-D does not. If you switch to RECT window, you’ll see the N-D drop below the RMS slightly, which is the expected behavior (since the N-D has parts of the spectrum notched out while RMS does not).
I opened an issue at the link below for study. thanks very much for reporting
It’s a very interesting topic–how to measure state of the art equipment with lots of margin and so I am happy to help! The process is a bit manual right now, but I think once the right combination of external circuits (eg notches+gain) and sig proc (use of lots of relay ranges to deep scrubbing to determine true harmonics) is found, it can be turned into a push-button measurement.