Hyldal said:
Better impulse response is a myth:
http://recforums.prosoundweb.com/index.php/t/14651/0/
oshifis said:
Then someone must have forgotten to tell Alex Peychev that, and yet his NWO-3.0GO is considered by many to be the best digital source ever made.
Bruno Putzeys is right on all accounts of course (I would expect no less from The Man !).
I will not repeat what he said about the impulse response, since this is basic sampling theory.
True, non-os will sound better than badly implemented oversampling, with lots of jitter, and digital filters designed to minimize silicon area and cost at the expense of precision, as is almost ALWAYS the case.
The only two digital filters I would consider for inclusion in the "properly implemented" list are the PMD100 and the filter in the Sabre DAC. All others have compromises like not enough bits of precision in the intermediate calculations, bean counting in other words !
Ironically a small FPGA from Xilinx with support logistics will cost the same than say, a DF1704 and will allow a much better filter to be implemented, and also you can tweak it !
Also even if using those chips which were designed by People Who Know What They Are Doing, or a well designed DSP or FPGA filter, note that there are approximately a zillion ways to f*uck the performance at the implementation stage, the most popular one being using a high-noise clock from a VLSI chip to feed a sensitive DAC.
As evidenced by the last product I opened, the Terratec DMX 6Fire USB, where clocks coming from low-cost oscillators fed from unfiltered and noisy power supplies, are routed through a FPGA (adding immense amounts of jitter) whose power supply is also inadequately decoupled, then through approx 8cm of trace and finally to the DAC itself. Needless to say this device sounds absolutely horrendous.
The irony of it is that this € 200 device could have been made excellent sounding with a budget of € 3 : some placing and routing changes to the board (free), one extra picogate (€0.2), a few caps (€0.1), a low noise reg from Micrel for the clock (€0.5) and high quality opamps instead of JRC bottom level ones...
I will not repeat what he said about the impulse response, since this is basic sampling theory.
True, non-os will sound better than badly implemented oversampling, with lots of jitter, and digital filters designed to minimize silicon area and cost at the expense of precision, as is almost ALWAYS the case.
The only two digital filters I would consider for inclusion in the "properly implemented" list are the PMD100 and the filter in the Sabre DAC. All others have compromises like not enough bits of precision in the intermediate calculations, bean counting in other words !
Ironically a small FPGA from Xilinx with support logistics will cost the same than say, a DF1704 and will allow a much better filter to be implemented, and also you can tweak it !
Also even if using those chips which were designed by People Who Know What They Are Doing, or a well designed DSP or FPGA filter, note that there are approximately a zillion ways to f*uck the performance at the implementation stage, the most popular one being using a high-noise clock from a VLSI chip to feed a sensitive DAC.
As evidenced by the last product I opened, the Terratec DMX 6Fire USB, where clocks coming from low-cost oscillators fed from unfiltered and noisy power supplies, are routed through a FPGA (adding immense amounts of jitter) whose power supply is also inadequately decoupled, then through approx 8cm of trace and finally to the DAC itself. Needless to say this device sounds absolutely horrendous.
The irony of it is that this € 200 device could have been made excellent sounding with a budget of € 3 : some placing and routing changes to the board (free), one extra picogate (€0.2), a few caps (€0.1), a low noise reg from Micrel for the clock (€0.5) and high quality opamps instead of JRC bottom level ones...
I would like to hear your oppinion on a set up with a interface resciver feeding an asyncronius upsamplingsconverter and finaly a 24 converter DAC, and leave out the oversampling filter.
Would that be worth to try.
The noise should move up in frequence, making it easier to filter out with analog filter after the DAC.
(Sorry about my newbe questions, but I learn from you guys every time a ask! Thank you!)
Would that be worth to try.
The noise should move up in frequence, making it easier to filter out with analog filter after the DAC.
(Sorry about my newbe questions, but I learn from you guys every time a ask! Thank you!)
"The only two digital filters I would consider for inclusion in the "properly implemented" list are the PMD100 and the filter in the Sabre DAC."
This is really interesting. Studio equipment is using up to 72 bit of precision for digital signal processing; I doubt that any chip implementations used in CD players comes even close. Peufeu, do you know what precision is commonly used?
I also think that the lenght of the FIR-filter is usually not adequate. Experiments have shown that going to 4096 or 8192 taps yields a significant improvment in sound quality. I lost the reference to that, but it may even have been somewhere here in DIYaudio. Such a large number of taps makes only sense if you can implement it with the required precision, which requires very powerful signal processing.
I don't doubt that many NOS-implementations sound very good. One reason may be that they are implemented by skilled DiYers with love and attention to detail, without penny pinching, often combined with a very good I/V-stage and output buffer. However, I think that the sound quality would be most noticable in the midrange. To keep the highs near the Nyquist frequency clean and uncontaminated by aliasing and higher order residuals is quite difficult without oversampling.
This is really interesting. Studio equipment is using up to 72 bit of precision for digital signal processing; I doubt that any chip implementations used in CD players comes even close. Peufeu, do you know what precision is commonly used?
I also think that the lenght of the FIR-filter is usually not adequate. Experiments have shown that going to 4096 or 8192 taps yields a significant improvment in sound quality. I lost the reference to that, but it may even have been somewhere here in DIYaudio. Such a large number of taps makes only sense if you can implement it with the required precision, which requires very powerful signal processing.
I don't doubt that many NOS-implementations sound very good. One reason may be that they are implemented by skilled DiYers with love and attention to detail, without penny pinching, often combined with a very good I/V-stage and output buffer. However, I think that the sound quality would be most noticable in the midrange. To keep the highs near the Nyquist frequency clean and uncontaminated by aliasing and higher order residuals is quite difficult without oversampling.
Yep.
Nope
The precision for the Sabre is mentioned in the Sabre's topic. I don't remember what it is, something like 32-36 bits.
Also most filters use the smallest number of taps they can get away with, and the quantization of the coefficients is not mentioned, along with the type of dither used. All of those are important...
No. Probably as cheap as possible though.
Well not necessarily. Suppose you use 4096 taps to oversample 44.1k to 176k. Each 176k sample only "sees" 1024 taps. So you get 176 million MAC/s per channel. With a crummy 50 MHz DSP this would be a problem of course. You mght want to spend a few bucks more. Blackfin might be borderline for stereo.
However, a low-cost FPGA contains many multipliers, each of them capable of many times that power. FPGA mults are 18x18 so you can use 2 to make 18x36 (quantizing the coeffs at 18 bits and keeping the full signal precision) or even use 4 to make 36x36 bit mults, but I doubt 36-bit coeffs would be useful. You could also use one multiplier and some slices to add the missing bits.
In order of increasing price, from low-end to pure overkill, with the GMAC/s ie billions of multiply-accumulates per second on 18x24 data :
Xilinx Spartan-3E 250 ($12) : 12x multipliers @ 100 MHz (0.6 GMAC/s)
Altera Cyclone EP3C5 ($15) : 23x mults @ 250 MHz (2.8 GMAC/s)
Xilinx Spartan-3E 500 ($20) : 20x multipliers @ 100 MHz (2 GMAC/s)
Altera Cyclone EP3C40 ($90) : 126x mults @ 250 MHz (15.7 GMAC/s)
Xilinx Spartan-3A DSP 1800 ($110) : 84x DSP units (MACs) @ 250 MHz (10 GMAC/s)
Xilinx XC5VSX95T ($2600 looool) : 640x 25-bit by 18-bit mults with 48-bit result and accumulator @ 550 MHz (352 GMAC/s this isn't the top end model mind you)
FYI "very powerful signal processing" is the last one in the list, ie. DSP on multichannel cellphone signals in baseband coz 100 gigaflops is cheaper than a bunch of analog RF filters these days. Each DSP unit draws 1.38 mW/100 MHz by the way.
So if you wanna oversample 44 to 176 in stereo with 1024 MACs for each 176k sample you need what ? 352 million MAC/s ? half that if your impulse response is symmetrical, so you need a few BRAMs to store your coeffs and samples and you need a few multipliers. Basically you need a $15 FPGA (in qty 1). Make that $20 with the logistics (power supply, 50 smd caps, etc). Since you're going to use 20 pins take a small SMD package and you can even fit it on a double sided PCB.
With the Xilinx tools you just click and instantiate a FIR from the FIR core generator which comes from the free webpack and you're all set. With Altera you have to pay or write it yourself which is prety easy.
I will (of course) put some custom FIR filters in my FPGA ethernet DAC... it's progressing nicely.
After liking it I grew bored of the NOS sound.
Re: Re: Nos Dac ves oversample DAC
There's even a more fundamental and theoretical reason why this is so.
Sorry, have no time to explain now, have to pick up the kids, but I may check in later and give it a try.
Home assignment: all please re-read the proof of the sampling theorem.
Cauhtemoc said:
There's even a more fundamental and theoretical reason why this is so.
Sorry, have no time to explain now, have to pick up the kids, but I may check in later and give it a try.
Home assignment: all please re-read the proof of the sampling theorem.
Thanks Peufeu for your valuable information; you have researched the available chip options very well. I wasn't aware that so much signal processing power is available at such comparatively modest cost. Indeed, I thought that you probably would end up with heat sinks!
So why isn't this used to the fullest advantage in the more expensive CD-players? One can only guess.
Now, that many are convinced that the CD is at the end of its lifespan, we finally begin to understand how to realize its full potential. Ironic. And only now do we have the required technology.
One other remark concerning the FIR-filter. Usually, an analog filter is used after the DAC to eliminate the higher order garbage. This is much easier to do with oversampling, but will still have some effect on impuls response and phase response in the audio band. This degradation can be eliminated by predistortion in the FIR filter, i.e. you optimize the coefficients for the joint system including the analog filter. Unfortunately, the coefficients will not be symmetric any more with such an approach. While this would be the clean way to go, I have no idea if the sonic improvment would really be noticable.
Finally, let's not forget to do the clocking right. The low-jitter master clock should clock the DAC and FIR-filter; the drive mechanism should then be slaved to that clock, and not the opposite way as done in most CD-players.
So why isn't this used to the fullest advantage in the more expensive CD-players? One can only guess.
Now, that many are convinced that the CD is at the end of its lifespan, we finally begin to understand how to realize its full potential. Ironic. And only now do we have the required technology.
One other remark concerning the FIR-filter. Usually, an analog filter is used after the DAC to eliminate the higher order garbage. This is much easier to do with oversampling, but will still have some effect on impuls response and phase response in the audio band. This degradation can be eliminated by predistortion in the FIR filter, i.e. you optimize the coefficients for the joint system including the analog filter. Unfortunately, the coefficients will not be symmetric any more with such an approach. While this would be the clean way to go, I have no idea if the sonic improvment would really be noticable.
Finally, let's not forget to do the clocking right. The low-jitter master clock should clock the DAC and FIR-filter; the drive mechanism should then be slaved to that clock, and not the opposite way as done in most CD-players.
> I wasn't aware that so much signal processing power is available at such comparatively modest cost.
Yeah, we're going to see the rise of the FPGA as a low-cost replacement to ASICs. Lots of souncards have FPGAs in them already, like RME, etc, because if you don't manufacture a million units, it's a lot cheaper than custom silicon...
> Indeed, I thought that you probably would end up with heat sinks!
I have a 75% full spartan-3E 500 with a microblaze and lots of other stuff, when running it draws about 0.5W. Far from a heat sink...
Also yes compensating the phase response of the analog filter in the digital domain is a good idea... I'll try that.
Yeah, we're going to see the rise of the FPGA as a low-cost replacement to ASICs. Lots of souncards have FPGAs in them already, like RME, etc, because if you don't manufacture a million units, it's a lot cheaper than custom silicon...
> Indeed, I thought that you probably would end up with heat sinks!
I have a 75% full spartan-3E 500 with a microblaze and lots of other stuff, when running it draws about 0.5W. Far from a heat sink...
Also yes compensating the phase response of the analog filter in the digital domain is a good idea... I'll try that.
The SAA7220 is also an example of everything that can go wrong in a digital filter :
- undithered output (quantization error)
- 16 bit accumulator (lots of rounding errors)
- 12 bit quantized coefficients (lol)
- 120 taps
In a word, completely worthless. But that's technology from 1985 !
It was designed to be manufacturable for an acceptable price at that time.
The SAA7378 in the CD723 is no better with its 16 bit arithmetic and undithered output, this kind of processing causes a great loss of detail and ambience in the recording.
Of course if you remove those and make the DAC non-os, it's going to sound a lot better, there is no question about it !
- undithered output (quantization error)
- 16 bit accumulator (lots of rounding errors)
- 12 bit quantized coefficients (lol)
- 120 taps
In a word, completely worthless. But that's technology from 1985 !
It was designed to be manufacturable for an acceptable price at that time.
The SAA7378 in the CD723 is no better with its 16 bit arithmetic and undithered output, this kind of processing causes a great loss of detail and ambience in the recording.
Of course if you remove those and make the DAC non-os, it's going to sound a lot better, there is no question about it !
Hi everyone,
I wrote an article about it. Listening digital filter
I have on my audio system many DAC. One of then is a Wadia 1000 who use reduce algorithm (4 taps). This digital solution gives very high definition sound. Conventional digital filter, with high taps, produce foggy sound.
Eric
Using high number of taps is non sense in audio application. Music is not periodical signal. This kind of distortion is very audible.Javin5 said:
I wrote an article about it. Listening digital filter
I have on my audio system many DAC. One of then is a Wadia 1000 who use reduce algorithm (4 taps). This digital solution gives very high definition sound. Conventional digital filter, with high taps, produce foggy sound.
Eric
OK, so you compared the digital filter in your player's AKM DAC with the stand-alone digital filter (is it the DF1700 ?) and it resulted the DAC's internal filter was better. That's interesting.
It would be interesting to do the same comparison with a no-compromise analog stage and low-jitter clock.
It would be interesting to do the same comparison with a no-compromise analog stage and low-jitter clock.
Hi Peufeu,
Like you can imagine, general audio result is very poor. But, difference than I can ear comes from software comparison.
Twenty years ago, we were very happy when we listen system with very depth space. In fact, we were listening default coming from digital filter!
Eric
The Philips DVD player use internal processor to perform digital filter upsampling for CD, SACD decoding and HDCD (bad) decoding. There is no DF1700, DF1704, DF1706 or other known circuit inside. Listening test performed compare software solution from single chip.peufeu said:
Like you can imagine, general audio result is very poor. But, difference than I can ear comes from software comparison.
Twenty years ago, we were very happy when we listen system with very depth space. In fact, we were listening default coming from digital filter!
Eric
Eric,
Hm, a similar experience can be had with a Northstar192 dac: you can switch upsampling on/off. When off, then you listen to the internal filter in the cs4396 dac. When on, you listen to the more sophisticated filter in the SM5849 and the crystal filter is switched to a basic level. Though both filters are higher tap number than in your case, still the definitely higher tap number NPC filter gives a much better result...
They paid attention to jitter in that dac, and it is the same for sure in both cases.
Are you sure about the quality of the included digital filter in that DSP? Not tap numbers, but precision, filter characteristics etc?
Ciao, George
Hm, a similar experience can be had with a Northstar192 dac: you can switch upsampling on/off. When off, then you listen to the internal filter in the cs4396 dac. When on, you listen to the more sophisticated filter in the SM5849 and the crystal filter is switched to a basic level. Though both filters are higher tap number than in your case, still the definitely higher tap number NPC filter gives a much better result...
They paid attention to jitter in that dac, and it is the same for sure in both cases.
Are you sure about the quality of the included digital filter in that DSP? Not tap numbers, but precision, filter characteristics etc?
Ciao, George
Hi George (Joseph K),
Later I have ability to test high-end DAC like Wadia 1000, Wadia X32, Dax...
It seems that DAC using low taps filter sounds more natural with more details. It's great with good recording, it's difficult with bad recording.
I'm designing a top-end NOS DAC and will probably confirm my impress. It seems that :
- With conventional DSP (with high tap filter) distorsion due to digital filter is inside music. (like particular colorature and lost of details).
- With NOS or low tap filtering, distorsion is outside music. Sounds seems 'piquancy'. Like if you add to music constant 1KHz signal. Human earing don't summ this two signal, but listening become very tiredness.
(I will confirm these impress in few mounth).
Eric
Philips DSP is low cost solution. SPDIF outputs constant 48KHz. DSD is converted to PCM before sending to AKM DAC. I never take time to mesure sample frequency when playing DSD, but I don't be surprise sample rate be 48KHz too.Joseph K said:
Later I have ability to test high-end DAC like Wadia 1000, Wadia X32, Dax...
It seems that DAC using low taps filter sounds more natural with more details. It's great with good recording, it's difficult with bad recording.
I'm designing a top-end NOS DAC and will probably confirm my impress. It seems that :
- With conventional DSP (with high tap filter) distorsion due to digital filter is inside music. (like particular colorature and lost of details).
- With NOS or low tap filtering, distorsion is outside music. Sounds seems 'piquancy'. Like if you add to music constant 1KHz signal. Human earing don't summ this two signal, but listening become very tiredness.
(I will confirm these impress in few mounth).
Eric
Hi Eric
Do your impress of your top-end NOS DAC confirm that low taps filter sounds more natural with more details ?
Can you give us some details about those impress ?
And what about the soundstage of your NOS DAC ?
Thanx
Paul
Hi Paul,
Each digital filter adds coloratura. An analog coloratura is easy to detect because it acts only for some frequencies. Digital coloratura introduced by digital filter acts at all frequencies. It's very difficult to detect it.
Take time to listen NOS DAC during one week or more and forget your old sound's reference. Don't search what you like in your previous DAC and appreciate qualities that you never hear before.
During comparison, you could not predict result. In some case you fell more bass, in other you fell less, in some case you feel more dynamic, in other you feel less...
I ask my friends to come with 3 CD; the first one have very good recording, the second one is preferred music and the third one is an awful recording.
When listen on NOS DAC, the third CD is not so awful
About NOS DAC, I can note constant impress:
- Each singer and musician stays at the same place. Each of one doesn't disturb the other when power is growing.
- There is a lot of information (ambiance, harmonics, depth...) but you can't speak about details. Each 'detail' is so natural, at the good position, so naturally integrated into music that you totally forget it. It's only when you lose all those details that you realize they were presents.
I hope that one day I discover a better DAC, it will be very instructive. If you have one, it will be a great pleasure...
Don't forget that a good DAC must be used with good source (CD Player). Even the best one still so poor if badly feed.
Yes. Digital filter add echoes (shown on Dirac impulse response: undulation before and after pulse can be shown as echoes). Adding echoes on sound cancel details.
Each digital filter adds coloratura. An analog coloratura is easy to detect because it acts only for some frequencies. Digital coloratura introduced by digital filter acts at all frequencies. It's very difficult to detect it.
Comparing OS DAC and NOS DAC is not easy. If you quickly compare for two or three pieces of music, you probably prefer OS DAC because it sounds more familiar. If you don't listen music from vinyl, oversampled sound is the only one you ear since 1990!
Take time to listen NOS DAC during one week or more and forget your old sound's reference. Don't search what you like in your previous DAC and appreciate qualities that you never hear before.
During comparison, you could not predict result. In some case you fell more bass, in other you fell less, in some case you feel more dynamic, in other you feel less...
I ask my friends to come with 3 CD; the first one have very good recording, the second one is preferred music and the third one is an awful recording.
When listen on NOS DAC, the third CD is not so awful
About NOS DAC, I can note constant impress:
- Each singer and musician stays at the same place. Each of one doesn't disturb the other when power is growing.
- There is a lot of information (ambiance, harmonics, depth...) but you can't speak about details. Each 'detail' is so natural, at the good position, so naturally integrated into music that you totally forget it. It's only when you lose all those details that you realize they were presents.
I take great pleasure to listen to it. At this time I never ear better (OS or NOS). But I didn't compare my realization with all manufactured DAC available on market.
I hope that one day I discover a better DAC, it will be very instructive. If you have one, it will be a great pleasure...Don't forget that a good DAC must be used with good source (CD Player). Even the best one still so poor if badly feed.
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.