a/d converter chips

[pucho812]

I got asked this the other day and since I am mainly an analog guy, I had no real idea or where to go looking for this information.

How do converter chips work? More over can anyone direct me to some good reading on how a/d chips work and so forth.

I got asked a question the other day that I couldn't answer and I forgot what the exact question was. But it had to do  with accuracy of signal vs bit depth in a 24 bit system and something along the lines of " I heard that some types of 24 bit converters amplify the signal in a sort of detection circuit so that low level analog signals get similar treatment of full bit resolution. More over if thats true how does it work?"

As far as converters go, I know them from a user standpoint and from a working tech standpoint i.e. calibrate and if it works or If it doesn't work, fix it."

 

[volker]

You could start here, that is a pretty nice series of articles. Just plug in higher numbers to the address, read at least #1-6. After that it gets a bit more in depth, then DACs are discussed. If you want to read about specific ADC architectures, that starts from #20. Audio relevant would be the sigma-delta-converter, #22/23.

 

[abbey road d enfer]

How do converter chips work? More over can anyone direct me to some good reading on how a/d chips work and so forth.

Start with that:
http://en.wikipedia.org/wiki/Analog-to-digital_converter
http://www.hardwaresecrets.com/article/317
http://www.beis.de/Elektronik/DeltaSigma/DeltaSigma.html
and then go to Lavry's website

I got asked a question the other day that I couldn't answer and I forgot what the exact question was. But it had to do  with accuracy of signal vs bit depth in a 24 bit system and something along the lines of " I heard that some types of 24 bit converters amplify the signal in a sort of detection circuit so that low level analog signals get similar treatment of full bit resolution. More over if thats true how does it work?"

That is some kind of digital companding; it exists since the '70's; was used to make 8-bit digital telephone codec's. I don't believe anyone uses them today. The last resurgences of such technology were HDCD (20bit rez from a 16bit CD) and dbx's type IV (27bit rez from 24-bit converters).

 

[JohnRoberts]

In yesterday's news was an announcement that RIM(?) or one of those mobile platform companies just settled a patent dispute with Dolby over licensing some audio compression algorithms, FWIW.

RE: A/Ds and codecs, it's magic (modern a/d/c don't even remotely resemble their older low bit ancestors technology).

I'm an old analog dog, but still an engineer with one eye on the bottom line, and it is hard to ignore 24bit stereo codecs for <$2... While probably not the best codec ever made, it makes doing stuff in the analog domain seem pricey by comparison...

JR

 

[Rochey]

Some ADC's still integrate a PGA (programmable Gain Amplifier) to amplify low level incoming signals, however, they are usually controlled from the systems main controller. These are used extensively in cellphone applications, where the amount of gain can be modified on the mic input, depnding on how loudly the user is speaking.

Other than that, most ADC's on the market oversample the incoming analog at a high rate (in the MHz range) with a low bit accuracy, then, using a digital decimation filter, will down convert the high rate to 44.1kHz/16bit or 24bit, or whatever is needed (24bit 96kHz etc)

Key thing to remember - just because a converter is specc'd at 24bits, doesn't mean that each of those bits is true. For example, the best ADC that TI makes, the PCM4222, is rated at 124dB dynamic range. 24bit audio is actually capable of 144dB dynamic range. If we approximate each bit to be 6dB of dynamic range, it means that the PCM4222's 24bit output is really only 21bits or so.

It gets worse for <$2 24bit converters!




// Interestingly, as a side note, I've seen a big difference in the sound between 24bit and 32bit DAC's... technically, on paper, I shouldn't (as human hearing isn't capable of hearing that low a noise floor)... but there is a lot of discussion about the digital filter topology used, and the data bit-width used.

 

[JohnRoberts]

Some ADC's still integrate a PGA (programmable Gain Amplifier) to amplify low level incoming signals, however, they are usually controlled from the systems main controller. These are used extensively in cellphone applications, where the amount of gain can be modified on the mic input, depnding on how loudly the user is speaking.

Yup, but these interface with common cell phone mics, so not typically capable of interfacing with professional recording mics. Also many cellphone codecs have relaxed performance on the A/D side with higher performance of the DAC side for playback of prerecorded music.

Other than that, most ADC's on the market oversample the incoming analog at a high rate (in the MHz range) with a low bit accuracy, then, using a digital decimation filter, will down convert the high rate to 44.1kHz/16bit or 24bit, or whatever is needed (24bit 96kHz etc)

Key thing to remember - just because a converter is specc'd at 24bits, doesn't mean that each of those bits is true. For example, the best ADC that TI makes, the PCM4222, is rated at 124dB dynamic range. 24bit audio is actually capable of 144dB dynamic range. If we approximate each bit to be 6dB of dynamic range, it means that the PCM4222's 24bit output is really only 21bits or so.

No kidding... but when well done the noise floor of the front end dithers the lower bits for an analog sounding noise floor, unlike the horrible noise floor of early A/D/A, while even that is a simplification.

It gets worse for <$2 24bit converters!

ya think?  .. However interesting when I can buy a nominally 24b codec cheaper than rolling my own even much lower word length conversions.

// Interestingly, as a side note, I've seen a big difference in the sound between 24bit and 32bit DAC's... technically, on paper, I shouldn't (as human hearing isn't capable of hearing that low a noise floor)... but there is a lot of discussion about the digital filter topology used, and the data bit-width used.

Again if you can reliably hear it you should be able to measure it.... Your mission is to find out and tell us what you are hearing.

JR

 

[ricardo]

For example, the best ADC that TI makes, the PCM4222, is rated at 124dB dynamic range. 24bit audio is actually capable of 144dB dynamic range. If we approximate each bit to be 6dB of dynamic range, it means that the PCM4222's 24bit output is really only 21bits or so.

The best A/Ds ju..ust manage 20 bits or maybe a little over.  See Bob Stuart's (Meridien) discussion of this.

Scott Wurcer (AD) says the best A/D he's laid his hands on was an experimental Panasonic 24b which was declared unmanufacturable.  It was true 24b but only for short periods.

There are many "24b" recording devices on the market which have the S/N performance of a cassette recorder.

A big problem is Sample Rate Conversion which a the core technology behind oversampling converters.  Many Sample Rate converters, especially those on the expensive DAWs are terrible; often with laughable anti-aliasing filters.  Aliasing IS audible and sounds like HF intermodulation.

That's probably the reason why zillion MHz bit rates sound "better" (different cos distorted) from simple 16b 44.1kHz A/D.  Especially when the end result is a CD.

There are several good comparative reviews of SRC and also A/Ds & recorders.

 

[JohnRoberts]

For example, the best ADC that TI makes, the PCM4222, is rated at 124dB dynamic range. 24bit audio is actually capable of 144dB dynamic range. If we approximate each bit to be 6dB of dynamic range, it means that the PCM4222's 24bit output is really only 21bits or so.

The best A/Ds ju..ust manage 20 bits or maybe a little over.  See Bob Stuart's (Meridien) discussion of this.

Thats why we read the data sheets.  Either 120 dB or 126dB are both respectable.

Scott Wurcer (AD) says the best A/D he's laid his hands on was an experimental Panasonic 24b which was declared unmanufacturable.  It was true 24b but only for short periods.

bummer... 24bit linearity? 24 bit Noise floor?

There are many "24b" recording devices on the market which have the S/N performance of a cassette recorder.

bummer,, a digital cassette or old phillips style?

A big problem is Sample Rate Conversion which a the core technology behind oversampling converters. 

I thought the down conversion from low bit high rate down to longer word slower sample rate was called decimation, and a different animal that sample rate conversion.

Many Sample Rate converters, especially those on the expensive DAWs are terrible; often with laughable anti-aliasing filters.  Aliasing IS audible and sounds like HF intermodulation.

Perhaps but again SRC is a different process than decimation, often compounded with dissimilar sample rates. 

That's probably the reason why zillion MHz bit rates sound "better" (different cos distorted) from simple 16b 44.1kHz A/D.  Especially when the end result is a CD.

There are several good comparative reviews of SRC and also A/Ds & recorders.

I'm just an old analog guy but I suspect there are good and bad examples of any technology. I am apprehensive about making broad sweeping generalities.

JR

 

 

[Rochey]

Meh... I'll stick to my day job defining audio converters for the home and pro audio market then.

John- your bang on the money...

 

[ricardo]

> Scott Wurcer (AD) says the best A/D he's laid his hands on was an experimental Panasonic 24b which was declared unmanufacturable.  It was true 24b but only for short periods.

bummer... 24bit linearity? 24 bit Noise floor?

Both.  For short periods.

> A big problem is Sample Rate Conversion which a the core technology behind oversampling converters. 

I thought the down conversion from low bit high rate down to longer word slower sample rate was called decimation, and a different animal that sample rate conversion.

Decimation is a form of Sample Rate Conversion and is an important stage in all SRC.  What you do is upsample to the Lowest Common Multiple fs, apply your Anti-alias filter and then decimate to the final fs.  Filtering increases the bit depth for free and some types of IIR are self dithering too.  This is sometimes called noise-shaping.  Excuse my unconventional use of certain terms but I can assure you that they are entirely accurate.

The A/D chip manufacturers like TI, AD, Philips, Crystal etc have this down pat.  The technology is now well understood  ..  except by the software people who write SRCs for the major DAWs.

The critical stage is the Anti-aliasing (the Digital Filter topology etc  as Rochey says).  There's no reason why a $2 24b converter shouldn't be very good with some attention to the analog side.  The Digital Filter side is theoretically complex but once sorted out, shouldn't cost more or give any problems.  But its two decades since I did this really dirty stuff seriously.

There's an excellent paper by my old friends (name dropping here  ;D) Michael Gerzon, Peter Craven & Peter Fryer (Bowers & Wilkins) which explains the decimation/SRC stuff.  It's actually about Speaker & Room EQ in Jurassic Days when a 200 pt FIR was huge so you had to use several bands to EQ 20 - 20k Hz.  Each band was SRC (or decimated) from the next higher one.

Michael & Peter C later did MLP for Meridien and DVD-A.

> That's probably the reason why zillion MHz bit rates sound "better" (different cos distorted) from simple 16b 44.1kHz A/D.  Especially when the end result is a CD.

I'm just an old analog guy but I suspect there are good and bad examples of any technology. I am apprehensive about making broad sweeping generalities.

Well the review of SRCs show many bad examples and only 1 good one among the expensive DAWs.  The public domain ones do better.

I should stress its SRCs from the major DAWs which are crap.  If you record at 192kHz and use one of these to make a 44.1kHz CD, you might have nasty audible intermodulation especially with cymbals.  If you recorded at 44.1kHz, the SRC (from the original oversampled signal) would be done for you by TI etc and they would get it right.

Rochey, would you care to give details of the 24b / 32b converters you heard differences with?

 

[Samuel Groner]

There's no reason why a $2 24b converter shouldn't be very good with some attention to the analog side.

Current converters are limited by the analogue part, as far as I know mostly from the sample & hold stages.

Samuel

 

[abbey road d enfer]

In particular, 144dB DR infers less than -137dBu EIN, which in turn means the source impedance must be < 35 ohms.

 

[JohnRoberts]

I should stress its SRCs from the major DAWs which are crap.  If you record at 192kHz and use one of these to make a 44.1kHz CD, you might have nasty audible intermodulation especially with cymbals.  If you recorded at 44.1kHz, the SRC (from the original oversampled signal) would be done for you by TI etc and they would get it right.

Rochey, would you care to give details of the 24b / 32b converters you heard differences with?

Thanks for the clarification... I get that they both do a down conversion, but I ASSumed oversampling convertors use more convenient simple ratios, while i guess we could encounter similar issues with A/Ds that support multiple non-integer output sample rates if they don't change the oversampling rate too. There are several cheap codecs that seem to just scale everything down from a HF clock.

So in short you are saying the chip makers (mostly) do it right, the DAW software guys mostly don't... I'm shocked :

If there have been serious investigations of real world SRC performance in DAWs, have they looked at the digital summing buses too? I have been perplexed for years by the whole "outside the box" summing approach, supposedly to deliver better performance. I spent decades intimately involved with understanding the limits of analog summing. Digital from my perspective promises improvements over this SOTA not signal degradation. Perhaps this is another dropped ball by the DAW coders (my suspicion all along). Did they happen to look at this performance aspect too?

JR

PS: I suspect the 24b linearity would be more likely to change over time/temperature than noise floor, but both are remarkable and (IMO) better than we need in practice for audio. I wonder how you would measure that? Not very easily I suspect.

 

[Rochey]

Typically converts do an integer division from the oversampling rate down to the desired sample rate. Typically, the master clock is different for multiples of 44.1kHz and 48kHz.

The main issue with the digital processing sections ( decimation for adc's and interpolation for dacs) is that traditionally, the digital density is relatively poor in high performance processes. To get the best performance (perceived sound quality), the larger thedigital circuit required.

I'm on my cell at the moment, but idid see someone ask about the difference between the 24b And 32b dacs we have (pcm1796 and pcm1795). The key things we saw was that the digital filter limited the performance of the analog circuit.
An 8x averaged fft of the output of the 24b dac showed the Noise floor reaching as low as -144dB in places. The same Fft on the 32b part showed content at -152dB. Ahhhh the fabled 25thbit!

Bear in mind also that the data path in the digital interpolation circuit allows for less truncation etc too. TruncAtion in the middle of the filter has a knockon effect .

Our Japanese Goode eared customers were able to discern between bothdevices. I don't have the same $50k monitoring setup at home to discern the difference.

Interestingly, our latest DAC (pcm5102) has a higher noise floor than our flagship, pcm1792. (112dB vs a potential 132dB) However, the digital interpolation filter and current segment (the real da bit) are similar. Our golden ear customers tell us that even though the noisefloor is higher, the tonal sound quality is very very similar. We suspect this to be down to the digital circuitry in the DAC, not necessarily the analog section.

Conclusion: there's a lot of black magic voodoo in converter design. I don't pretend or fool myself to understand much below the architecture block diagram level. What differentiates one vendor from another isnt always the silicon process, but the experience and knowledge of the definition and design teams makingthe silicon.

 

[abbey road d enfer]

So in short you are saying the chip makers (mostly) do it right, the DAW software guys mostly don't... 

That would be an extremely rude judgement...
Decimation in a converter chip (not an SRC chip) is done by taking one sample every 64, 128, 256,...clock ticks (depending on the masterclock and desired output rate). It's easy to do, and the filters are fixed so they can be fully optimized.
In a DAW, that's a different story; the basic downsampling is already done in hardware. SRC in a DAW is converting from one base SR to another, very often with a ratio that is not a power of two.
Well, in fact, two stories:
One for file processing, which is a non-real time process, and provides SRC between mathematically fixed SR's. The algorithms used there are based on upsampling to the Least Common multiple and then downsampling to the target SR. [aside] (That's similar to pre-scaling, which involves shifting bits to the right, which is much easier than applying attenuation via the ALU). [/aside]  Obviously the filters can be optimized, even more than in a chip, because there are no real-time constraints and the processing power is much larger than whatever's in the chip (although, due the the dedicated architecture, Mips are well optimized).
The second story is teal-time SRC, which involves two more or less fluctuating clocks and a permanent calculation of the real instant ratio, which is a very complex operation. Basically there are two solutions.
One is the polynomial SRC, which can be assimilated to a frequency detector delivering the instant value of the ratio, and the filters may or may not follow; most often they don't, they are fixed and the designer has to give them some slack.
The other is the system Lavry uses; I'm not sure he invented it. Basically, the converter assumes that both frequencies don't drift and operates the conversion. Obviously, after some time, one of the buffers (the input buffer that receives the incoming data and an output buffer that feeds the output when the WC tells it to) is bound to overflow, so the system will change the ratio in order to compensate.
Basically, if you consider the clock variations as two more or less parallel curved lines, the first method continuously adjusts the ratio to follow the median line, the second method does it with a succession of straight lines.
The claimed advantage of the second method is that it produces extremely low levels of jitter, because the frequency is fixed for a period of time, but from time to time, there is a large discontinuity, when the system has to change gears.
Now, in addition to that, there is a compromise, because the anti-alias filter cannot be perfect. The number of taps cannot be infinite, because of physical limitations, but also because the more taps, the longer the latency (256 taps at 44.1k is about 6ms). So the designers make their choices. Some privilege transient response at the expense of out-of-band attenuation or HF extension, some privilege roll-off at the risk of being criticized for the harshness of cymbals, some offer the choice of several optimizations (Reaper, Izotope, Samplitude,...).
Check
http://src.infinitewave.ca/

 

[Samuel Groner]

In particular, 144 dB DR infers less than -137dBu EIN, which in turn means the source impedance must be < 35 ohms.

Dynamic range is maximum level minus noise floor; thus you cannot conclude to noise floor without stating a maximum level. But you are right, designing an analogue AD driver (or DA IV converter/filter) is challenging even at 120 dB dynamic range. My best DA converter so far gets to 130 dB dynamic range in a 22 kHz measurement bandwidth.

Available recording spaces hardly support 16 bit, so fortunately this discussion is rather moot from a practical point of view...

Samuel

 

[abbey road d enfer]

That was based on a 5Vpp input (AKM style).

 

[0dbfs]

How about the Neumann digital mic's?
http://www.neumann.com/img/Linkgraphics/Solution-D_E.pdf

Any ideas how they get a reported 140dB dynamic range?

Cheers,
jb

 

[JohnRoberts]

One nuance or subtlety surrounding dynamic range discussions for A/D/A is the classic analog definition (peak signal to noise floor) ignores what is going on below this nominal noise floor. We can hear signals that exist below this nominal noise floor so there is another signal floor down at the quantization limit.  How far this quantization occurs below the noise is probably audible if close.

One thing this discussion should be making apparent is not to make conclusions about performance based simply on digital word length.

At the end of the day the old school analog metrics matter, but there are other metrics to understand if designing at the bleeding edge of technology.

JR

 

[pucho812]

guys thanks for all the comments and info. Lots to digest but all worth it. I'll point my co-worker hear to read as well.