If you know where to get a PMD-200 HDCD chip, I would like to know. I would design it into the Sabre board that I am working on.
I have a PMD-100 chip in a DAC that I am not using. I considered making a board for it, but I estimate it requires a 128-macrocell Altera that would need to be developed.
Since I have the output of the Oppo patched to send the full 20-bit HDCD PCM out, I don't really need it. But it would be nice to be able to decode HDCD that comes from a PC media server.
I have a PMD-100 chip in a DAC that I am not using. I considered making a board for it, but I estimate it requires a 128-macrocell Altera that would need to be developed.
Since I have the output of the Oppo patched to send the full 20-bit HDCD PCM out, I don't really need it. But it would be nice to be able to decode HDCD that comes from a PC media server.
I am aware of that software decoder. I tested it several months ago.
If I were to have several thousand FLAC files on a media server, I would have to uncompress each one, test it for HDCD, and then use the decoder.
Even with the Foobar scripts mentioned in the link, I have never bothered to do all of that.
The decoded HDCD wav files result to be 3 to 6dB quiter than the rest of the wav files because of the decoded headroom.
Much easier to use a hardware decoder. It automatically adjusts the volume to match HDCD and non-HDCD.
The software decoder doesn't upsample to 88.2 either. The author said he would work on it, but seems to have given up and not worked on it since September.
sorry to stray so far off-topic.
I still want some PMD-200 chips if anyone has any.
If I were to have several thousand FLAC files on a media server, I would have to uncompress each one, test it for HDCD, and then use the decoder.
Even with the Foobar scripts mentioned in the link, I have never bothered to do all of that.
The decoded HDCD wav files result to be 3 to 6dB quiter than the rest of the wav files because of the decoded headroom.
Much easier to use a hardware decoder. It automatically adjusts the volume to match HDCD and non-HDCD.
The software decoder doesn't upsample to 88.2 either. The author said he would work on it, but seems to have given up and not worked on it since September.
sorry to stray so far off-topic.
I still want some PMD-200 chips if anyone has any.
rossl said:
Is there any advantage of using PMD200? AFAICS, it has worse
passband ripple than say a DF1706, and is still a half band filter.
I have looked at most commonly available DF's and even the
highly regarded Anagram Q5 still is half band and has 0.004
passband ripple.
Can any one shed some light on why any other std filter should be
superior to the Sabre's existing one?
Dustin, did ESS do any listening tests when implementing Sabre's
digital filter?
Thanks,
Terry
PMD-200 with Sabre
This is one of the good things about many eyes and ears involved...any permutation that can be imagined will be tried, and results shared. Hail Darwin!
I wouldn't count on MS releasing the primo implementation to the public, but it sure is worth a listen.
The 100 has a reputation as being one of the best sounding DFs, even when playing non-HDCD data, and the 200, though made largely from unobtanium, refined that sound a bit, when implemented carefully. The 100 reclocks/registers its output data for minimal jitter without external reclocking, whereas the 200 unfortunately does not. This is where the Sabre's jitter insensitivity is going to be quite welcome.
Now, whom is sitting on a secret stash of PMD-200s? Maybe somebody on this list has contact(s) within a manufacturer that has not yet cleaned off their shelves from a few years back?
Cheers,
WMS
This is one of the good things about many eyes and ears involved...any permutation that can be imagined will be tried, and results shared. Hail Darwin!
I wouldn't count on MS releasing the primo implementation to the public, but it sure is worth a listen.
The 100 has a reputation as being one of the best sounding DFs, even when playing non-HDCD data, and the 200, though made largely from unobtanium, refined that sound a bit, when implemented carefully. The 100 reclocks/registers its output data for minimal jitter without external reclocking, whereas the 200 unfortunately does not. This is where the Sabre's jitter insensitivity is going to be quite welcome.
Now, whom is sitting on a secret stash of PMD-200s? Maybe somebody on this list has contact(s) within a manufacturer that has not yet cleaned off their shelves from a few years back?
Cheers,
WMS
It is very much the question if such a digital filter should be used at all, or at all times. Since the ASRC in de Sabre is probably superior to the one on other upsamplers the addition of an extra digital filer might not be very advantageous, specially for non HDCD sources ( for most people that would probably mean 99% of their music collection I guess).
For HDCD content, the software decoder doing the 16 -> 24 bit conversion might very well be the best choice. You don't need the upsampling since the Sabre does it better anyway
For HDCD content, the software decoder doing the 16 -> 24 bit conversion might very well be the best choice. You don't need the upsampling since the Sabre does it better anyway
The only way to know that for sure is to compare :
a. 44.1k -> Digital Filter or DAC's embedded filter -> DAC
b. 44.1k -> 4x Oversampling in the PC -> DAC
c. 44.1k -> ESS Sabre
d. 44.1k -> 4x Oversampling in the PC -> ESS sabre
People seem to report option b being massively better than option a (I believe this). People also report PMD100 to be the best in the role of Digital Filter. As for option c and d, this is the unknown, and therefore needs to be tested. I will, as soon as the hardware is ready (FPGA PCBs ordered yesterday from protoexpress) and I can lay my groping fingers on a Sabre DAC.
a. 44.1k -> Digital Filter or DAC's embedded filter -> DAC
b. 44.1k -> 4x Oversampling in the PC -> DAC
c. 44.1k -> ESS Sabre
d. 44.1k -> 4x Oversampling in the PC -> ESS sabre
People seem to report option b being massively better than option a (I believe this). People also report PMD100 to be the best in the role of Digital Filter. As for option c and d, this is the unknown, and therefore needs to be tested. I will, as soon as the hardware is ready (FPGA PCBs ordered yesterday from protoexpress) and I can lay my groping fingers on a Sabre DAC.
Previously to using the Eval board, I was using a WM8741 based DAC. The Wolfson part has several (3 in hardware mode) filters with different ripple and roll-off characteristics. I found it hard to pinpoint any significant differences between them and wouldn't fancy my chances of saying which was which in a blind test. Kind of like swapping cables.
The Wolfson is a fine part, but changing to the Sabre made an instant improvement - knackered old cliche, but I really could hear information I'd never heard before, and it was quite obvious I was listening to a higher resolution DAC.
The Wolfson is a fine part, but changing to the Sabre made an instant improvement - knackered old cliche, but I really could hear information I'd never heard before, and it was quite obvious I was listening to a higher resolution DAC.
Spartacus said:
Ah! The answer at a question I asked some time ago, before being absolutely convinced that, if people were gettint that mad about this Sabre, that because it should really overperform everything existing.
Hum, just to know: does any of you have listened to it in 8 ch mode? How does it performs then?
Sabre and HDCD
For at least one the posters here, moi, the HDCD ratio frequently listened to is probably closer to 5 or 10%...and you know about the rewards of that last 10% of the performance curve. When it's good, it's really good. One way up that curve is "and" rather than "or"; I plan on both straight Sabre and HDCD fed Sabre.
Funny, the inventors of HDCD had expected it to be an interim format in the transition to a high resolution media. Thanks in part to a good old fashioned format war, it is still here a decade beyond. The PM model 2 is still regarded as one of the best, if not the best studio A/Ds for mastering; so HDCD CDs are still being released.
Cheers,
WMS
For at least one the posters here, moi, the HDCD ratio frequently listened to is probably closer to 5 or 10%...and you know about the rewards of that last 10% of the performance curve. When it's good, it's really good. One way up that curve is "and" rather than "or"; I plan on both straight Sabre and HDCD fed Sabre.
Funny, the inventors of HDCD had expected it to be an interim format in the transition to a high resolution media. Thanks in part to a good old fashioned format war, it is still here a decade beyond. The PM model 2 is still regarded as one of the best, if not the best studio A/Ds for mastering; so HDCD CDs are still being released.
Cheers,
WMS
Uhm, and also the Sabre is still a half band filter. For zero alias this is the only one that I know of at the moment.
http://audiopraise.com/vanity/overview.php
Ciao, George
Peufeu,
A halfband filter does not fulfill the Nyquist criteria: it does not eliminate all components below Fs/2, but it is instead ~symmetrical around Fs/2. All this for eliminating half of the coefficients in the FIR filter.
Why halfband is a problem, according to Bruno Putzeys:
A filter with very low inband ripple is the npc SM5842 /47.
This is the thread:
http://recforums.prosoundweb.com/index.php/m/0/14586/16/0/#msg_14586
The Sabre is half band, according to the data sheet: it starts to filter at 0.454 Fs and reaches full attenuation at 0.546 Fs.
Ciao, George
A halfband filter does not fulfill the Nyquist criteria: it does not eliminate all components below Fs/2, but it is instead ~symmetrical around Fs/2. All this for eliminating half of the coefficients in the FIR filter.
Why halfband is a problem, according to Bruno Putzeys:
A filter with very low inband ripple is the npc SM5842 /47.
This is the thread:
http://recforums.prosoundweb.com/index.php/m/0/14586/16/0/#msg_14586
The Sabre is half band, according to the data sheet: it starts to filter at 0.454 Fs and reaches full attenuation at 0.546 Fs.
Ciao, George
Zero alias filter is an expression coined by forum member Glassman..
, and I suppose that wants to describe a full Nyquist -compatible digital filter, with a lot of cells "wasted" for precision. If I'm right, it means that in it's datasheet somewhere should appear:
Passband: ~ 0.4 Fs
Stopband: ~ 0.5 Fs
Ciao, George
, and I suppose that wants to describe a full Nyquist -compatible digital filter, with a lot of cells "wasted" for precision. If I'm right, it means that in it's datasheet somewhere should appear: Passband: ~ 0.4 Fs
Stopband: ~ 0.5 Fs
Ciao, George
Yeah, I know what a halfband filter is. I was doubting because there's a typo in the datasheet with "dB" instead of "fs", page 16 of the datasheet, and also because I was wondering if the "slow roll off" was halfband or not (seems likely), and in this case would the "sharp roll off" be halfband too, etc.
Anyway I googled "zero alias filter" and found no reply except just one for-pay IEEE article so I wondered if this was some super secret technology or a name someone picked up for "a filter that is optimized to alias as little as possible", from what you say it is the second option. It sounds a bit marketspeak though.
I wonder what filter lengths this uses.
The XC3S200 has 12 18-bit multipliers, running at less than 100 MHz, let's guess 88.2M. That's anout 1 GMAC/s which is really good for a 10$ chip. Multiply both price and performance by between 100 and 500 for a V5SX, lol.
The specs of the Vanity mention 33-37 bit arithmetic and 47-67 bit accumulator. I wonder what that means wrt the number of bits used to encode the coefficients. If he went simple, he would have used 2 18bx18b mults to make a 18bx36b mult. Or 4 to make a 36bx36b mult. But he could also have extended a mult with gates to make a 24bx18b mults (using less mults). I wonder where the 33 bit arithmetic comes from.
For 24 bit data and 18 bit coeffs it would give about 6000 MACs per sample per channel. You could do some filtering with that.
Oversampling 44.1 at 4 times...there is 216 kbits of BRAM which must store the sample fifo (48 bits per sample, 2x24) and the coeffs (18 bits, 4 coeffs per sample, but the coeffs are symmetrical around 0 and identical left and right, so 18 bits per sample)...
Allowing extra circuitry for encoding SPDIF (if he used the SPDIF core from opencores, this one uses lots of resources) I would guess the FIR would apply to about 2.5-3K samples of 44.1 audio, max (which means either 3K taps or 12K taps depending on how you count). Which is a lot more than your usual hard filter chip. If he had used a XC3S100 he would have had about 1/4 of this, but he forked the extra $5 to go for the 200, so it is probably more taps than the 10 could handle. So, it gives an estimate.
I wonder why he didn't use a XC3S250E which has the same number of mults but costs less and can configure itself from a SPI flash. I would guess the uc is there to download several configurations from the SPI flash depending on the user settings, perhaps even depending if DSD or PCM is used.
I can't resist the call of reverse-engineering, lol.
Anyway I googled "zero alias filter" and found no reply except just one for-pay IEEE article so I wondered if this was some super secret technology or a name someone picked up for "a filter that is optimized to alias as little as possible", from what you say it is the second option. It sounds a bit marketspeak though.
I wonder what filter lengths this uses.
The XC3S200 has 12 18-bit multipliers, running at less than 100 MHz, let's guess 88.2M. That's anout 1 GMAC/s which is really good for a 10$ chip. Multiply both price and performance by between 100 and 500 for a V5SX, lol.
The specs of the Vanity mention 33-37 bit arithmetic and 47-67 bit accumulator. I wonder what that means wrt the number of bits used to encode the coefficients. If he went simple, he would have used 2 18bx18b mults to make a 18bx36b mult. Or 4 to make a 36bx36b mult. But he could also have extended a mult with gates to make a 24bx18b mults (using less mults). I wonder where the 33 bit arithmetic comes from.
For 24 bit data and 18 bit coeffs it would give about 6000 MACs per sample per channel. You could do some filtering with that.
Oversampling 44.1 at 4 times...there is 216 kbits of BRAM which must store the sample fifo (48 bits per sample, 2x24) and the coeffs (18 bits, 4 coeffs per sample, but the coeffs are symmetrical around 0 and identical left and right, so 18 bits per sample)...
Allowing extra circuitry for encoding SPDIF (if he used the SPDIF core from opencores, this one uses lots of resources) I would guess the FIR would apply to about 2.5-3K samples of 44.1 audio, max (which means either 3K taps or 12K taps depending on how you count). Which is a lot more than your usual hard filter chip. If he had used a XC3S100 he would have had about 1/4 of this, but he forked the extra $5 to go for the 200, so it is probably more taps than the 10 could handle. So, it gives an estimate.
I wonder why he didn't use a XC3S250E which has the same number of mults but costs less and can configure itself from a SPI flash. I would guess the uc is there to download several configurations from the SPI flash depending on the user settings, perhaps even depending if DSD or PCM is used.
I can't resist the call of reverse-engineering, lol.