My yet another finding was perhaps the strangest. In sync mode ES9038Q2M seems to have some noise floor related issues depending on MCK frequency. At 24.576MHz the noise floor is steady up to about -1.5dBFS. With higher input noise floor rises about 20dB. This seems to be repeatable as it occurred with all my 3 ES9038Q2M chips.
Here is 12kHz@48k at -1.5dBFS.
And here the same but with -1.4dBFS.
With 12.288MHz clock the noise floor shift occurred at lower input level (about -2dBFS).
What could be the reason for this?
Here is 12kHz@48k at -1.5dBFS.
And here the same but with -1.4dBFS.
With 12.288MHz clock the noise floor shift occurred at lower input level (about -2dBFS).
What could be the reason for this?
No, but the boards are naked in front of my LCD monitor so it is possible that there is some LF or RF pickup.
No.
Sync mode.
Without laying out the specifics of the listening session used on SQ assessment, it's a moot point because the results vary greatly depending on the types of listening session.
Yes, MCK needs to be 128fs for ES9038Q2M to work properly in sync mode. Thanks JensH & Markw4.
When MCK is 256fs or 512fs in sync mode ES9038Q2M seems to work properly unless the input signal is above 7kHz with level at about -2dBFS or above. So with 24.576MHz MCK clk_gear needs to be 1/4 for 48k and 1/2 for 96k. But strangely it did not work at all for 192k with clk_gear at 1 (default).
When MCK is 256fs or 512fs in sync mode ES9038Q2M seems to work properly unless the input signal is above 7kHz with level at about -2dBFS or above. So with 24.576MHz MCK clk_gear needs to be 1/4 for 48k and 1/2 for 96k. But strangely it did not work at all for 192k with clk_gear at 1 (default).
Last edited:
Yes, bits 3:2 to 00. I need to check if BCK frequency has some impact.
I've been running the ES9038Q2M at 1/2 or 1/4 MCK due to lower high order distortion and power consumption though with higher noise. For these noise measurement I set clk_gear to 1 for lower noise but did not realize it had a side effect.
I've been running the ES9038Q2M at 1/2 or 1/4 MCK due to lower high order distortion and power consumption though with higher noise. For these noise measurement I set clk_gear to 1 for lower noise but did not realize it had a side effect.
Interesting about the noise. Long ago when I tried using clk_gear to reduce ES9038Q2M MCLK frequency, I thought it sounded worse when divided. Didn't much care why, only cared about better or worse SQ but more noise might be the reason why.
Also might have been that what I was hearing with undivided MCLK was what I now call 'false clarity,' when higher order HD has the effect of cutting through otherwise muddy sound. The effect can be mistaken for true increased clarity. That was a few years back before figuring it out. My professional designer friends have since confirmed that they are aware of the effect too.
Also might have been that what I was hearing with undivided MCLK was what I now call 'false clarity,' when higher order HD has the effect of cutting through otherwise muddy sound. The effect can be mistaken for true increased clarity. That was a few years back before figuring it out. My professional designer friends have since confirmed that they are aware of the effect too.
Last edited:
If I interpret the datasheets correctly, the ES9038 typically works with a 3.3 V external reference supply and the AK4490 with 5 V. For a given reference voltage noise density and given full-scale output levels, that would give the AK4490 an advantage of 20 dB*log10(5 V/3.3 V) ~= 3.6 dB. Was the difference of that order?
This is normal. High-K ceramics (X7R etc, not C0G/NP0) are piezoelectric microphones. If you use X7R cap on a high impedance node like CSET on LT3042 then every vibration on the board will be faithfully reproduced in the output. The piezoelectric process converts stress (when the cap gets squeezed) into charge (ie, current). So it creates a voltage that is proportional to product of capacitor vibration stress, and the impedance of the node the cap sits on. On a low impedance node like power supply decoupling, the effect is not noticeable. On a high impedance node, it can be quite hilarious.
Note on the skirts: the amplitude of jitter and phase noise skirts is proportional to the derivative of the signal, therefore its frequency. Noise on VREF simply multiplies with the signal, so the amplitude of the skirts is constant no matter what the frequency is. So you can distinguish between the two.
Another simpler method is to notice the output of a DAC is proportional to VREF. The output is the product of the digital sample value and VREF. After all, in order to output analog, a DAC needs to know what a volt (or a milliamp) is, otherwise what would it do? That's what the reference is for.
Most DACs are differential. So, when outputting zero digital samples, the DAC outputs a differential signal with both positive and negative polarities being equal. Usually it's something like VREF/2 (AVCC/2 for ESS DACs, you get the idea). So whatever noise is on VREF is present on both sides of the differential output, and the next stage, which is usually a diff amp, does its job and rejects it. Therefore, signal to noise measurements done while playing zero samples are useless. They measure the CMRR of the differential IV, not the noise of the DAC. It's pretty hilarious.
If you want to know the noise on your VREF, you can make the DAC play a wav file which contains constant full scale digital samples. Something like +32767 in 16 bit. Then, the differential output is maxed out, the negative outputs 0V, and the positive outputs VREF. And the analog output of your DAC is equal to VREF, which you can then measure without even opening the case.
If it has a digital highpass filter, that won't work.
It is interesting to do this measurement while playing a test signal on the other channels. If the designer didn't do their job, for example they didn't notice ESS DACs draw constant current on VREF if the I/V is differential, but they draw signal-dependent current if the IV is not differential and leaves one of the outputs unused... then you'll see the response of the VREF regulator to this current on the channel that is set to output a constant DC value. You should see a harmonic of the test signal pop up, second or third, I don't remember. Probably third.
We were all thinking about microphony as an explanation, but are not sure if that is it. There must be random vibrations in the 0 to 10 Hz frequency range then. Shock mounting the PCB and measuring in quiet surroundings should help if it is the cause.
Good point.
Noise doesn't get masked at all during silence, so this is a very relevant measurement as far as I'm concerned - but it indeed doesn't show how good or how bad the reference is.
You also need a DC coupled DAC then. If it has an analogue high-pass filter, there will be a hole in the spectrum (bohrok2610 is interested in offset frequencies from 0 to 10 Hz).
No as my noise measurement setup (Groner LNA, ES9822PRO ADC) is probably good down to about 10Hz.
Yes, 3.3V for ES9038Q2M and 5V for AK4490. The difference was closer to 10dB.
I made some further tests. The 22uF/25V X5R I used has only about 60% of nominal capacitance at 5V bias. This seems to be common among the various 22uF/25V X5Rs I have. Then I replaced the 22uF Cset with 10uF tantalum polymer. Noise skirts were very similar with 22uF tantalum polymer so still much better than with X5R. So it seems that the bias voltage dependent capacitance of X5R is not the issue. The cause may well be piezoelectric microphony. I still find that quite odd since my workplace does not have any significant LF (or RF) noise sources. Shock mounting the DAC board seems like an overkill.
It could, with the old version. And it does improve the dynamic range. I tried it, but unfortunately the distortion at high levels was much higher, so I went back to 5V.
Don't use a 7 V supply with the AK4490REQ, which is the new version (after the fire at AKM).