Sony "staggered" twin TDA1541

Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.
A myth seems to have arisen concerning the operation of the two TDA1541 DACs in Sony CD players such as the CDP337ESD. Sony claims 8X oversampling from "staggered" operation. The myth is that they are really just connected in parallel.

In fact, data inputs and current outputs are connected in parallel. However, the digital filter is set for 8X oversampling: double the processing rate of each DAC. One DAC WS input is clocked by the WS output of the filter, and the other by the not-WS, so they do indeed alternate, each processing 4X, resulting in 8X output.

Parallel TDA1541/A has a bad reputation that has been extended, unfairly, to the Sony implementation, which if nothing else does indubitably have the benefit of the higher rate.

I've never quite grasped why monotonicity is important, but it occurs to me that alternating monotonic dacs may not combine to give a monotonic result. That is, an incremental rise in the input could lead to a fall in the output, or vice versa.

Anyway, just to put the record straight. The historic and illustrious name of Sony needs all the help it can get.
 
Just had a scan through my CDP-507ESD service manual, and you do seem to be correct about the dac arrangement. Whether it's advantageous versus straight parallel remains to be determined. I've just acquired a 507ESD, which I consider fundamentally better than the far more expensive(even now) 707ESD, and plan to see just how good I can make it sound sometime in the next few months. It's interesting that the 507ESD uses half of each TDA1541A-R1 for one channel, and the other half for the other, rather than the more common use of a whole dual dac per channel. Makes for simpler l/r clocking, I suppose, and easier layout.
 
It's interesting that the 507ESD uses half of each TDA1541A-R1 for one channel, and the other half for the other, rather than the more common use of a whole dual dac per channel.

Are you sure? My interpretation is that both DACs get all 8 oversamples for both channels, and each ignores alternating L/R pairs. That is, they both operate normally at 4X.

But there is an apparent anomoly...8X output steps each with the period of 4X takes twice as long as real time. :confused:

So the outputs overlap like bricks, and consequently the output level from the IV should be double, just as for normal parallel DACs. This, it occurs to me, is why it looked so bad when tested with RMAA...it was overloading the input to my rudimentary soundcard. Now I've made a buffer/headphone amp, so I can test it again with the necessary attenuation.

As for the sound, it's delicious. How much that comes from pride of ownership, I can never tell. I got worried about mechanical noise from the irreplaceable KSS190A and now I'm using an Arcam Alpha. When I swap back I'll be in a better position to compare. I guess the biggest difference will be between the analogue filters.

Also using, with my headphone amp, a delightfully simple little CDPM75...single TDA1541 and just one dual opamp serving both channels...sounds good too.
 
I was merely saying that each dac chip is used r channel to r chan, left chan to left chan, not commenting on the sampling in that regard. Most parallel dac implementations use a dual dac's left & right channels in parallel with each other for one channel, then same for other.
As to KSS190 acoustic noise, no cause for concern, IMO, as it is pretty darn certainly just the focus gain and/or tracking gain being turned up higher than needed. All you need to do is play the cd/cdr the player takes longest to start up(if there is a difference), then turn each gain pot down(which ever way reduces noise, usually ccw) to the point where play starts failing, then turn pot back up about 20 degrees of rotation. Then verify it has no trouble starting up any other discs, or skipping tracks, etc. Other thing to cause excess noise is if focus bias(focus offset) is off. Usually the best set point corresponds to minimum noise point, but not always, and it really should not be adjusted without an oscilloscope, using which one shoots for maximum RF pattern amplitude & clarity.
 
Ah. I didn't know what most parallel implementations do. I've seen plenty of TDA1543 stacked in parallel, where all the Rs do the R, and the Ls the L. AFAICS for the Sony arrangement that's the only way to achieve the 8X, which is why I mentioned it.

As for my KSS190A, I did all the adjustments by the book. Had 0.8V RF only. Renewed all relevant caps and readjusted, to get barely 1V, just within spec. Finally adjusted FO by ear, but there's still an episodic tickytacky noise from the mech. Not intrusive like it was, which is why I'm going to reinstall it.

Search and play for all disks has never been a problem...almost instant thanks to the linear sled motor. The eye pattern is sharp and clear, but there's spikes on the crests and, during the noisy spells, its DC offset jumps in time with the rotation-speed tick. I guess it's the focus reaching the end of its travel. No idea what the spikes are about. Philips are fuzzy, Sony are spiky, I thought.

One thing I'm about to check is whether the CD is allowed to tilt beyond allowable limits. The tilt could vary with the jogging of the sled, explaining the episodic nature of the noise.

I'm hoping someone else will rebuild one of these heads before I have to.
 
In fact, data inputs and current outputs are connected in parallel. However, the digital filter is set for 8X oversampling: double the processing rate of each DAC. One DAC WS input is clocked by the WS output of the filter, and the other by the not-WS, so they do indeed alternate, each processing 4X, resulting in 8X output.

The TDA1541A has two possible input modes, the I2S mode won't accept more than 4X OS as there's a limit on the bit clock speed (6.4MHz). The mode which works faster has separate L and R data inputs. Check to see if pin4 has any activity - if so then its using the simultaneous data mode.

If the I2S WS pin is merely inverted then there will be smearing - adding together of adjacent samples - because the update of the output is always done on the clock cycle subsequent to the falling edge of WS. To handle this correctly without distorting the output (incurring roll-off) would require a delay line.
 
The output from the filter chip is set to 8X, I2S mode. I assumed the TDA1541 are each 4X, I2S. They alternate and overlap to achieve the 8X. There's no delay. I assume, but perhaps should check, that there is none between the WS and not-WS outputs of the filter, but it doesn't seem likely.

Checking the TDA1541 datasheet, it's set for I2S it seems, with pin 4 tied to pin 2 (BCK) and pin 27 held high.

What I'm uncertain about is exactly how the DACs operate their input registers and latches. The data, and presumably the WS signal, arrive and disappear at twice the rate expected by each DAC. Unexpected signals are presumably ignored.

Smearing would be quite an oversight on Sony's part. I think I see what you mean, but don't see how it applies here. The outputs are completely overlapped. With two 4X chips sharing 8X data, there's no time for them to be anything else.
 
It certainly applies to all I2S DACs and it bit me when I tried this trick of swapping channels, using TDA1387s. You have updates going on at 8X but the DAC with the inverted WS does not just have the channels swapped over - because WS is used to time the updates and frame the data.

I2S is sent left first, then right. When you invert the WS you don't suddenly get right sent first, rather you end up treating the right channel of the previous sample as your new (current) left channel. This gets paired with the current left, re-labelled as current right. Hence the stereo pair isn't a true pair any longer.

An easy oversight to make, but it remains an oversight nevertheless.
 
There is nothing in the CDP-507ESD schematic to suggest that the WS for the second dac is merely inverted. What you're describing would not have been missed by Sony's design staff, I'm sure. And no channel swapping is going on. Each 1541's left channel is used for the left channel, and each right channel is used for the right channel. I believe the WS signal is simply "demultiplexed", one pulse fed to one dac chip, the next pulse sent to the other. Each dac thus simply handles half of the samples. What is not clear to me is if the output currents sum or not. Seems to me that the output current from the two dacs should stay the same as for one dac, since they are alternating, like two people hammering a nail, so you get twice as many hits, not twice the force. I will be able to confirm or debunk all of the above once I have time to upgrade the 507ESD.
 
Presumably each DAC's output produces a constant current until it receives its next sample. Therefore, you will have two DACs producing current per channel, which basically doubles the current compared to a single-DAC circuit.

Interleaving two four times oversampled DACs like this does give you eight times oversampling, but you get some extra low-pass filtering for free. That is, as each DAC keeps its output constant for 1/(4 fs) instead of 1/(8 fs), the zero-order hold filter transfer goes to zero at all multiples of 4 fs rather than 8 fs. This makes life easier for the analogue post filter and probably reduces sensitivity to jitter somewhat, but it needs to be compensated for in the digital oversampling filter.
 
Its that 'free low pass filtering' that gives the clue as to this not really being 8X OS. The point of 8X OS is to move the image frequencies out further from the wanted frequency band - so they would be centred around 384k, but they aren't in this case as you point out. So to me its unclear what the technical benefits might be.
 
Hi Abraxalito,

Why do you think the images are not centred around 8 fs? I would expect that the images are centred around 8 fs, but the zero order hold suppresses both 4 fs and 8 fs.

This is all assuming perfectly matched DACs. If you have mismatch between the DACs, you will get some imaging around 4 fs back, but much less than with just a single DAC running on 4 fs (depending on how well the conversion gains of the DACs match).

Regards,
Marcel
 
Hi Marcel,

It strikes me that to get frequencies centred around 8Xfs you do need to update the DACs with new data every 8X clock, not have them holding the same data for two of these clocks. Otherwise seems too much of a free lunch. But if you're right then this would allow sub-unity OS ratios, as follows...

Suppose I wanted to do a NOS DAC but instead I ran the two DACs at 22kHz, but interleaved - it would be a cool solution if it worked, but I can't see it. You reckon this would work? If so I plan to design the first sub-NOS DAC :D How far could we extend it? Run 4 interleaved DACs at 11kHz? As I said, seems too good to be true.

<edit> Here's my hand-waving argument for why it can't work. An 8X OS DAC reproduce frequencies up to 8 * 22kHz = 176kHz. To create a 176kHz sinewave at full scale it needs to swing from positive full scale to negative full scale on each cycle. This is the worst case. But if only one of two DACs (whose outputs are summed) is updating at this rate then the combined output can only swing down to 0V because the first DAC is still holding the previous positive full-scale value. The output can only swing to -ve full scale after both DACs have been updated, which of course takes 2 clocks. So no longer can we create 176kHz at full scale. Perhaps we could create 176kHz at half scale (-6dB) though if I thought carefully about it :)
 
Last edited:
Hi Abraxalito,

The problem with your sub-NOS DAC is that the first zero of the response of the zeroeth-order hold gets rather close to the band of interest, even in the band of interest when you have more than two DACs. You could solve that using return-to-zero DACs, but on top of that, the requirements on DAC matching will be impossibly tough.

My reasoning is as follows:

Start with a digital signal with sample rate fs. According to discrete-time signal processing theory, its spectrum is repetitive, having spectral copies around all multiples of fs. When you put this into a theoretical perfect impulse DAC, its output will be a series of infinitely high and infinitely narrow (Dirac) impulses having a charge or flux that corresponds to the digital number and, of course, having a sample rate of fs.

Now comes the digital interpolating low-pass filter with sample rate 8 fs. If the filter works properly, its output signal only has spectral copies around multiples of 8 fs. When you put this into a theoretical perfect impulse DAC, its output will be a series of Dirac impulses with a sample rate of 8 fs.

When you demultiplex the signal across 2 perfectly matched perfect impulse DACs running at half the clock rate, whose clocks are shifted by 1/(8 fs), and add the outputs, you get the exact same series of impulses as you would have got with the single impulse DAC. Hence, the spectrum is also the same, and spectral copies only occur around multiples of 8 fs.

For a number of practical reasons, a real-life DAC does not output infinitely high and narrow impulses. Instead, it may output a pulse with finite height and width T. Mathematically, such a signal could have been obtained by passing the impulse signal through a zeroeth-order hold filter, that is, a filter having an impulse response that is constant during a time T and then drops to zero. Being linear and time invariant, such a filter cannot create any frequencies that weren't there before, it can only filter the frequencies that are already there. In fact it acts as a low-pass having a sin(pi*f*T)/(pi*f*T) magnitude response.

Assuming perfect DAC matching, the only difference between a conventional system with a single DAC running at 8 fs and the Sony system with two interleaved DACs running at 4 fs is the different value of T. Conventionally, T=1/(8 fs), but for Sony, T=(1/4 fs). Hence, the Sony zeroeth-order hold filters a bit more.

Now with mismatch. For simplicity, start with total mismatch: one DAC works and the other doesn't. The working DAC sees an input signal with sample rate 4 fs. Hence, its input signal spectrum has spectral copies at all multiples of 4 fs. The whole system then degrades to a four-times-oversampling system.

With partial mismatch, the output signal spectrum will be something in between the cases with no and with complete mismatch. For example, suppose the first DAC has 1 % more conversion gain than intended while the second DAC has exactly the right conversion gain. You could get the same signal by taking two DACs with the intended gain, and adding one DAC running at 4 fs with 1 % of the desired DAC conversion gain. The spectral copies at odd multiples of 4 fs will then reoccur, but at only 1 % of the level you would have had in a single-DAC system. As the desired signal is twice as strong as in a single-DAC system, the ratio of desired signal to 4 fs spectral copy is still 200 times (46.02... dB) larger than with a single DAC running at 4 fs.

By the way, it is true that you can't make 176.4 kHz anymore with the Sony system. That is the consequence of the time T of the zeroeth-order hold increasing to 1/(4 fs): the first zero of the filter response will be at 176.4 kHz when fs=44.1 kHz. For CD that doesn't matter because the desired signal is below 22.05 kHz anyway.

Regards,
Marcel
 
Last edited:
It certainly applies to all I2S DACs and it bit me when I tried this trick of swapping channels, using TDA1387s. You have updates going on at 8X but the DAC with the inverted WS does not just have the channels swapped over - because WS is used to time the updates and frame the data.

I2S is sent left first, then right. When you invert the WS you don't suddenly get right sent first, rather you end up treating the right channel of the previous sample as your new (current) left channel. This gets paired with the current left, re-labelled as current right. Hence the stereo pair isn't a true pair any longer.

An easy oversight to make, but it remains an oversight nevertheless.

OK, woke up next day and you're making sense now...as does the following discussion. My original thought was that feeding an inverted WS to one chip would lead the two chips to latch at equal intervals, so everything would work out somehow. But for a double-rate WS that's not true. They should be triggered by alternate WS cycles.

Perhaps I misread the symbols...the pin I took to be inverted WS is marked WS with a bar. The filter is CXD1144. I'll look at the datasheet more closely when I've got some thinking time, and look with a 'scope.

For now, I note that the 337ESD manual illustrates the waveforms at the two WS DAC pins with a single picture.

I came to praise Sony...
 
sony's first staggering DAC circuit was employed in DAS-R1.( it was a DAC with transport CDP-R1)
According to an article of a japanese audio magazine about them,
it is suggested that the CXD1144 have poor attenuation around 4fs,
so it will be some help the attenuation of 0th order hold of each staggered 4fs DAC.

sony developed CXD1244 next year ,which have sufficient stopband attenuation.
so they do not need to employ staggering in next model, haha.
 
Last edited:
Looking more closely at the 507ESD schematic.
The WS pins are marked LE/WS and APT/WS. Not very revealing, but certainly seems to indicate that one is not the inverse of the other.
The other thing that I'd not noticed before is that, while each dac chip's AOL pins are summed to feed left out, and each chip's AOR pins are directly summed to go to right out circuit, i.e., each 1541 feeding it's r ch to r ch, left ch to left ch, as I'd said before, the CXD1144 data L is fed to only one dac, and the data R is fed only to the other. That is a brain teaser.
Also, the dac chips, which are labeled TDA1541-R1 in the physical reality, are designated TDA1541-R5 on the diagram, which suggests to me that the chips are custom matched.
Darned interesting setup, which I would not have paid attention to until it came time to mod my 507, had the OP not brought it up.
 
LE and APT are Sony, non-I2S options. The chip is set to I2S. One of your WS should be not-WS, like this
 

Attachments

  • CXD1144pinout.jpg
    CXD1144pinout.jpg
    191.2 KB · Views: 269
sony's first staggering DAC circuit was employed in DAS-R1.( it was a DAC with transport CDP-R1)
According to an article of a japanese audio magazine about them,
it is suggested that the CXD1144 have poor attenuation around 4fs,
so it will be some help the attenuation of 0th order hold of each staggered 4fs DAC.

sony developed CXD1244 next year ,which have sufficient stopband attenuation.
so they do not need to employ staggering in next model, haha.

So if Sony used stagger to mitigate the attenuation problem, that might mean that the CXD1144 has features specifically designed to support that mode of operation, including the two outputs to the DACs WS pins.

When I've finished dinner, I'll get the 'scope out. Dual trace but not dual beam. What's the best way of investigating the relationship between the two WS signals?
 
Status
This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.