Hi! Everybody knows Cirrus Logic DACs CS43131 and CS43198. Once I noticed a discrepancy between the sound of my CS43131 and its measurements. After some research, I came to the conclusion that its brilliant results are fake. I have not any source with CS43198, so it would be interesting if someone checked my findings with it.
Fake 1. 352kHz and 384kHz support
Both chips accept such streams, but few people noticed the fixed setting of digital filters for these modes: 0.263Fs. In other words, frequencies above 92.8/101kHz will not be played at all.
Fake 2. Noise I (THD+N)
The noise floor in wideband measurements, which is also rarely used, has a steep rise above 40 kHz.
On the FFT, the signal harmonic bases are more wide compared to, for example, es9038q2m.
I'm not an expert, just an advanced user. If this is not a consequence of active noise shaping, then at least it looks that way. That is, the real THD+N should be noticeably worse.
Fake 3. Noise II (DR)
According to the standard test, CS43131 shows a DR of at least 130dB(A), similar to good professional chips. It was unclear why I couldn't hear it. The following test was carried out. A 30kHz@0dB signal was mixed into the standard 1kHz@-60dB. 30kHz is far enough from the [20Hz-20000kHz] range so as not to interfere with the measurement too much.
Result DR=129.8dB(A)
The noise floor rose by almost 20dB!
Result DR=110.3dB(A)
This confirms the rumors about variable gain, I suppose.
I very briefly repeated this test on the ES9038Q2M to avoid an error.
Result DR=122.8dB(A)
Result DR=120.9dB(A)
In general, there is no difference of 20dB here.
Logic of Currus
Of course, I can't know it for sure. I believe the dynamic gain is needed for outstanding DR. The noise shaping helps to improve THD+N measurement. And the huge noise hump is reduced by a non-switchable filter. I don't want to blame the CS developers for anything. I just don't think the CS43131 is a high quality HiFi chip.
PS Many thanks to IVX for exellent equipment!
Fake 1. 352kHz and 384kHz support
Both chips accept such streams, but few people noticed the fixed setting of digital filters for these modes: 0.263Fs. In other words, frequencies above 92.8/101kHz will not be played at all.
Fake 2. Noise I (THD+N)
The noise floor in wideband measurements, which is also rarely used, has a steep rise above 40 kHz.
Fake 3. Noise II (DR)
According to the standard test, CS43131 shows a DR of at least 130dB(A), similar to good professional chips. It was unclear why I couldn't hear it. The following test was carried out. A 30kHz@0dB signal was mixed into the standard 1kHz@-60dB. 30kHz is far enough from the [20Hz-20000kHz] range so as not to interfere with the measurement too much.
Result DR=110.3dB(A)
This confirms the rumors about variable gain, I suppose.
I very briefly repeated this test on the ES9038Q2M to avoid an error.
In general, there is no difference of 20dB here.
Logic of Currus
Of course, I can't know it for sure. I believe the dynamic gain is needed for outstanding DR. The noise shaping helps to improve THD+N measurement. And the huge noise hump is reduced by a non-switchable filter. I don't want to blame the CS developers for anything. I just don't think the CS43131 is a high quality HiFi chip.
PS Many thanks to IVX for exellent equipment!
I suspect the "wider harmonic bases" is simply the phase-noise skirt, i.e. an indicator of more jitter in the sampling time. Note that some fraction of this skirt will be the Blackmann Harris window profile - for measurements of such high dynamic range maybe you should be using a window function with even higher dynamic range, HFT144D or HFT169D perhaps (which are flat top windows so will get the peaks accurate in a spectrum, see here: https://holometer.fnal.gov/GH_FFT.pdf).
The rise in noise above 40kHz will indeed be noise shaping, but you don't care about this as its ultrasonic - it can be filtered out digitally whenever you want. The benefit of noise shaping is the vastly lower noise within the audio band it can give. Similarly I don't the issue with the 384kSPS mode filtering out ultrasonics. There are many many ways to structure a decimation filter chain in an ADC, so long as the audio spectrum is represented cleanly without aliasing onto it, that's all you need. I'd worry more about spurs in the audio band under full signal input, ie SFDR upto 20kHz.
The noise floor with the 30kHz tone - do you get the same response if the 0dB tone is 15kHz, not 30kHz?
The rise in noise above 40kHz will indeed be noise shaping, but you don't care about this as its ultrasonic - it can be filtered out digitally whenever you want. The benefit of noise shaping is the vastly lower noise within the audio band it can give. Similarly I don't the issue with the 384kSPS mode filtering out ultrasonics. There are many many ways to structure a decimation filter chain in an ADC, so long as the audio spectrum is represented cleanly without aliasing onto it, that's all you need. I'd worry more about spurs in the audio band under full signal input, ie SFDR upto 20kHz.
The noise floor with the 30kHz tone - do you get the same response if the 0dB tone is 15kHz, not 30kHz?
I mentioned about these skirts as a trails of noise shaping caused by finite precision of processor. But it's only my guess. Thanks for comment.
Of course this HF noise hump is not a big problem. Though I'd prefer the unshapen variant. The noise floor is about -110dB and fully inaudible. So losing some HF signal under hump is meaningless. It also leads to unability of comparing such artifical results to other DACS. That's a main disadvantage. You are right about SFDR but I don't sure that harmonics will be still reduced by noise shaping in the real signal.
Yes. The noise floor will rise gradually with any frequency from -30dB.
Of course this HF noise hump is not a big problem. Though I'd prefer the unshapen variant. The noise floor is about -110dB and fully inaudible. So losing some HF signal under hump is meaningless. It also leads to unability of comparing such artifical results to other DACS. That's a main disadvantage. You are right about SFDR but I don't sure that harmonics will be still reduced by noise shaping in the real signal.
Yes. The noise floor will rise gradually with any frequency from -30dB.
This chip doesn’t truly support 384 kHz based on the filter and analog performance. However,
I don’t think it’s a real issue.
I don’t think it’s a real issue.
Indeed, the most of hi-res records fits in 100kHz band and the value of higher frequencies is not so clear. But the low actual DR is the real, clearly hearable issue.
No. Look at the measurements of Sonata BHD Pro in ASR.
If you attenuate the signal by 41dB the SNR will not degrade:
In the multitone measurement of TANCHJIM SPACE you can't see any spurious components:
Such measurements are made at a level of -60dB not -30 or higher. I showed my similar measurement with the result of 129.8 dB(A).
If this is done, the real DR will degrade by the same 41dB. And yes, DR will be about 130(noise)-41(peak)=89дБ. Actually my result 110.3dB(A) is much better. Still, this is not the declared 130dB.
It's looking good. I believe that multitones are not very precise measurements. Due to intermodulation, the multitone level is well below 0dB, no matter what you see in this picture. Because of this, the harmonics are also low. In addition, they are suppressed by CS43131 noise shaping. In a real signal, the spectrum is much denser and the illusion of a large range will disappear.
You played a full scale signal at 30 khz with the nos filter. No wonder the result was bad. And the frequency you used for your measurement has nothing to do with music reproduction.
Do you know the difference between DR and SNR? With every digital attenuation the signal decreases but the noise level will be the same. This behavior is expected. Compare the chart on ASR:
The signals are all 0dbFS which beside you everyone else can see.
The noise floor shape in your DR measurements looks strange. It should have similar shape as at 0dBFS. What is the measurement setup?
No. It's a premixed file. THD+N(1kHz)=-85.1dB[20..20000Hz]
That's what I'm talking about. This statement is not true for a CS43131.
For example, you may write 0dB 1kHz file for THD measurement, but not for SMPTE. You may check this.
Perhaps the reason is in the displayed FFT ranges. They're differ between CS and ES. All floors was flat (to 40kHz for CS43131) before AW calibration.
Setup was E1DA 9038D6K - E1DA Cosmos ADC@mono. My ADC is too noisy for precise DR measure, but 30kHz@0dB made impossible to use either a preamp or a notch.
Noise floor shape should be the same regardless of FFT range or input level. What do you mean by AW calibration? The noise floor shape suggests that you are using some calibration file in REW that is totally incorrect.
I mean A-weight calibration in REW. What's wrong with it?
6.3 85.4 0
8 77.6 0
10 70.4 0
12.5 63.6 0
16 56.4 0
20 50.4 0
25 44.8 0
31.5 39.5 0
40 34.5 0
50 30.3 0
63 26.2 0
80 22.4 0
100 19.1 0
125 16.2 0
160 13.2 0
200 10.8 0
250 8.7 0
315 6.6 0
400 4.8 0
500 3.2 0
630 1.9 0
800 0.8 0
1000 0.0 0
1250 -0.6 0
1600 -1.0 0
2000 -1.2 0
2500 -1.3 0
3150 -1.2 0
4000 -1.0 0
5000 -0.6 0
6300 0.1 0
8000 1.1 0
10000 2.5 0
12500 4.3 0
16000 6.7 0
20000 9.3 0
40000 17 0
8 77.6 0
10 70.4 0
12.5 63.6 0
16 56.4 0
20 50.4 0
25 44.8 0
31.5 39.5 0
40 34.5 0
50 30.3 0
63 26.2 0
80 22.4 0
100 19.1 0
125 16.2 0
160 13.2 0
200 10.8 0
250 8.7 0
315 6.6 0
400 4.8 0
500 3.2 0
630 1.9 0
800 0.8 0
1000 0.0 0
1250 -0.6 0
1600 -1.0 0
2000 -1.2 0
2500 -1.3 0
3150 -1.2 0
4000 -1.0 0
5000 -0.6 0
6300 0.1 0
8000 1.1 0
10000 2.5 0
12500 4.3 0
16000 6.7 0
20000 9.3 0
40000 17 0
Ok. That type of calibration file is not normally used as the resulting curve is not comparable to datasheets (or measurements made by other members). REW gives you absolute numbers. A-weighted DR/SNR is approximately 2-3dB higher.
I really can't get where exactly you see the problem. As I said, it was an approximate measurement, a little spoiled by ADC noise (-126.3dB(A) in mono mode). Actual measured DR of my DAC is 125.8dB(A), this unit was calibrated by IVX as 126dB(A). So with nearly the same ADC noise the DR result should be 3dB lower, exactly around 123dB(A).
As I said that type of calibration file is not normally used so the noise floor looks odd. Nothing else. Can you repeat the measurements without the calibration file?
LOL, nick_I measured the Meizu Hifi Pro Dac with DSD128 and the E1DA with PCM:
It's not worth to discuss this.
It's not worth to discuss this.
I mean A-weight calibration in REW. What's wrong with it?
6.3 85.4 0
8 77.6 0
10 70.4 0
12.5 63.6 0
16 56.4 0
20 50.4 0
25 44.8 0
31.5 39.5 0
40 34.5 0
50 30.3 0
63 26.2 0
80 22.4 0
100 19.1 0
125 16.2 0
160 13.2 0
200 10.8 0
250 8.7 0
315 6.6 0
400 4.8 0
500 3.2 0
630 1.9 0
800 0.8 0
1000 0.0 0
1250 -0.6 0
1600 -1.0 0
2000 -1.2 0
2500 -1.3 0
3150 -1.2 0
4000 -1.0 0
5000 -0.6 0
6300 0.1 0
8000 1.1 0
10000 2.5 0
12500 4.3 0
16000 6.7 0
20000 9.3 0
40000 17 0
8 77.6 0
10 70.4 0
12.5 63.6 0
16 56.4 0
20 50.4 0
25 44.8 0
31.5 39.5 0
40 34.5 0
50 30.3 0
63 26.2 0
80 22.4 0
100 19.1 0
125 16.2 0
160 13.2 0
200 10.8 0
250 8.7 0
315 6.6 0
400 4.8 0
500 3.2 0
630 1.9 0
800 0.8 0
1000 0.0 0
1250 -0.6 0
1600 -1.0 0
2000 -1.2 0
2500 -1.3 0
3150 -1.2 0
4000 -1.0 0
5000 -0.6 0
6300 0.1 0
8000 1.1 0
10000 2.5 0
12500 4.3 0
16000 6.7 0
20000 9.3 0
40000 17 0
Did you see the real levels of these Archimago multitones: -12.98dBFS, -12.99dBFS, -11.97dBFS? You just prove my words.
I didn't use DSD anywhere in these measurements. Please go to ASR and laugh at Amir too.
I did not realize you were using A-weight calibration in REW as you did not mention it in your OP. Anyhow I would not use it when sharing measurements as practically nobody does.
I don't see your point about noise shaping as every DS DAC/ADC uses it.
Here is ES9038Q2M in SYNC mode at 128fs. Noise shaping in ES9038Q2M becomes apparent with lower MCLK.
