DAC output filters need explanations

[mr.duck]

1) What do the "digital filters" built into the DAC chip do? If they are for filtering high frequency noise from the oversampling and D to A conversion, then how can they be digital filters and not analog filters?

2) Since there is this filter already in there, why do people put a low pass filter on the DAC output? If it is for a reconstruction filter then is it really needed? And does the DAC's built in filter do any reconstruction?

3) If the low pass filter is really needed, then why isn't a high order filter such as 8th order filter used to do a really good filtering job?

4) If a 2nd order low pass filter with buffer is recommended over a filterless design, what would the passband frequency need to be? Would 30 KHz be good or would 100 KHz be a better choice, and why?

[dusfor99]

Originally Posted by mr.duck

1) What do the "digital filters" built into the DAC chip do? If they are for filtering high frequency noise from the oversampling and D to A conversion, then how can they be digital filters and not analog filters?

2) Since there is this filter already in there, why do people put a low pass filter on the DAC output? If it is for a reconstruction filter then is it really needed? And does the DAC's built in filter do any reconstruction?

3) If the low pass filter is really needed, then why isn't a high order filter such as 8th order filter used to do a really good filtering job?

4) If a 2nd order low pass filter with buffer is recommended over a filterless design, what would the passband frequency need to be? Would 30 KHz be good or would 100 KHz be a better choice, and why?

Hi Mr.Duck,

I will try explain the best I know.

1. TH filtser is to get rid of the "images" of the singal that appear when you oversample a signal. Basically when you oversmaple a digital signal sample at a vertain rate, you will get these. For example.

It you are sampling at 44.1kHz, (CD's) and you are sampling a 20kHz tone, that tone is actaully indistiguishable from its images, one which is located at 24.1kHz (44.1kHz - 20kHz). So if you are trying to play a 20kHz through a DAC with no filtering, you will actually get tones at these image frequencies coming out of the DAC!!!. Mathematically if you have an input tone called "TONE" and a smapling rate "FS", you will have images at N*FS +/- TONE, where N is an integer. SO from this you can see if you sample a 20kHz tone at 44.1kHz, you not only just get the image at 24.1kHz, but also at 64.1kHz (44.1k + 20k) , 68.2kHz (2*44.1k - 20k), (2*44.1k + 20k) 108.2kHz...... The oversamplong filters job in life it to remove these as they were never in the origanal recording. SInce the oversmapking filter take the rate up to a much higher rate, it still too suffers from images, but now since the sample rate is so high, the first image can be over 1 MHz, easily filters by a simple R-C ananlog filter. for example, if you have an 32x oversamping filter, the 44.1kHz gets upsampled to 1.4112MHz., if the oversmapling process sucessfully got rid of all the images cause from the 44.1kHz initial sample rate, you now only have to worry about filtering freuquencies above 1.3912MHz (32*44.1kHz - 20kHz) This is were a R-C filter would doa great job.

2. As eluded to above, the R-C filter is to get rid of the images of the oversampling filter itself, but these images are at a much much higher frequency and a simple 1st order is usually sufficient. Also, the analog filter will be used to filter someong else out. The quantization noise from the noise shapeing modulator. (I assume were not talk about non-oversampled DAC here) Since the morder DAC are now all multibit noise shaped types that can have up to 135DNR, (AFAIK) you simple cannot just build a R2R ladder to get this, what you do is make a 5/6/7/8/9bit DAC that runs at a MUCH MUCH higher frequency than the sample rate (this is why you need the oversmapling filter) and re-create the signal with 6-9 bits only. THis will have quantization noise, the modulator in the DAC takes this noise and moves it into the 100kHz to 50Mhz or so frequency band, the analog filter must get rid of this before trying to drive into a pre-amp or power amp.

3. Practically speaking, if we could built a sharp 8th order analog filter, maybe oversampling wouldn't be needed, (with the exception of something else ill mention later), but how would one make that filter with the pole and zeroes right where they want them? Im sure it cane be done on a PCB, but the yield would not be that good do you variances in the R's and C's. The next reason is because if one builds a sharp 8th order analog filter successfully, the group delay will not be constant since it is an IIR type filter. Non constant group delay means that you cymbal hits (high frequencies) and bass drum kicks (low frequencies) dont come out of the filter with the same amount of delay. So if a cymbal was hit at the instant the bass drum was kicked, and you play that track back trough this wicked 8th order analog filter, they will not longer be in sync upon play back, this TOTALLY messes up the soundstage.

4. My choice would be to have the filter at 100kHz, this is to minimise the phase shift at 20kHz and also to have a "resonably" constant group delay over the 20 - 20kHz range.


Just my thoughts.

Dustin

[kevinkr]

Great explanation Dustin, but the many typos and misspellings would make it hard reading for many.

[tritosine]

http://www.lavryengineering.com/white_papers/sample.pdf

_________________________

if we start with this paper, I dont see the need for 50mhz sampling rate, nor 135DNR. These are numbers beyond belief to me. The best delta sigma ADC is like 127DNR. Assuming the best real world case, say 120, thats no problem to a single bit quantizer. Wont need 50mhz.

[dusfor99]

Originally Posted by kevinkr

Great explanation Dustin, but the many typos and misspellings would make it hard reading for many.

Hi Kevin,

Yes, I know im terrible at the spelling thing, I've been told this many times, I really should slow the typing down.

Thanks

Dustin

[dusfor99]

Originally Posted by tritosine

http://www.lavryengineering.com/white_papers/sample.pdf

_________________________

if we start with this paper, I dont see the need for 50mhz sampling rate, nor 135DNR. These are numbers beyond belief to me. The best delta sigma ADC is like 127DNR. Assuming the best real world case, say 120, thats no problem to a single bit quantizer. Wont need 50mhz.

Hi Tritosine,

I was just pointing out what the DAC can do. Maybe the ADC will catch up soon I know there is one that does do 135dB DNR , its the ES9018 (Sabre 32) I have measured this myself on the Audio Precision 2722.


Thanks

Dustin

[oshifis]

Digital filters "restore" or calculate the values of the oversampled data between the real sampled data. For example if the sampling rate is 44.1 kHz and we have 4x oversampling, we need to calculate 3 values between each two original samples. This is called interpolation, and can be linear or based on higher-order filters, which is precisely what digital filters are for.

[tritosine]

Originally Posted by oshifis

Digital filters "restore" or calculate the values of the oversampled data between the real sampled data. For example if the sampling rate is 44.1 kHz and we have 4x oversampling, we need to calculate 3 values between each two original samples. This is called interpolation, and can be linear or based on higher-order filters, which is precisely what digital filters are for.

seems to me you confuse oversampling with curve fitting. Perhaps read the Lavry paper, it also explains why non oversampling is bad for your waveform. Oh, and there is pre ringing on most CD material, embedded.