One fine day I received the following e-mail:

From: lgruber@bih.harvard.edu Wed Mar 27 18:56:06 1996
X-Sender: lgruber@mercury
To: ntracy@trc.amoco.com
Subject: XDAC purchase and AD1890
Organization: Beth Israel Hospital Boston MA
Mime-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Content-Length: 4556
X-Lines: 73

With a sizable tax return, I am looking forward to purchasing a new CD transport and digital processor system soon. I am strongly considering the Rotel RDD-980 for the transport. But I am presently undecided about whether I should build the XDAC 3.0 (I'm not at all afraid of building it) or just purchase the companion Rotel RDP-980 as the processor. The reason for this is because of some comments recently in Stereophile. As a result, I am sending a letter to the editor's of The Audio Critic detailing my indecision. I value you opinion and so am including the text of this letter to them below:

After reading Robert Adams' enlightening article on the use of the Analog Devices asynchronous sample-rate converter (the AD1890) for reducing jitter in CD players (The Audio Critic, Issue 21), I thought that all was settled in my mind, since reduced jitter seems to be a desirable effect, and this chip seems to do the trick very effectively, the use of this device would be a strong consideration in my purchase of a new CD playback system. However, after reading Robert Harley's remarks in Stereophile, I am a bit confused. From his attendance in a panel discussion at the Audio Engineering Society convention last October, he says that Bob Adams "suggested that the AD1890 shouldn't be used for jitter attenuation if the sampling rate doesn't need conversion." Harley's opinion is that the AD1890 "effectively suppresses jitter but changes the value of every audio sample slightly--an undesirable side effect." He then implies that the panelists at this jitter workshop seem to hold the opinion that the dual phase-locked-loop approach was considered the best method. (Stereophile, January, 1996, p. 43). Then, in his review of the Z-Systems Z-Link jitter attenuator (March, 1996, p. 180ff), he amplifies on his opinion by saying: "the AD1890's oversampling digital filter creates entirely new samples based on the input data, some of which are chosen to appear at the output. Consequently, the input data and the output data are not identical--not a good thing. The sample values are changed slightly from what they were, which can only be regarded as a form of distortion, however low that amplitude error may be. If a jittered signal produces the right amplitudes at the wrong times, an AD1890-processed signal produces slightly wrong amplitudes at the right times."

Harley's comments about the AD1890's incompatibility with the HDCD format (it "won't pass the control code...) is less important to me. But then he goes on to say that, at the convention, "Robert Adams, publicly stated that....the chip could primarily be used when sample-rate conversion is needed, not strictly as a jitter attenuator", and says that Adams recommends the dual-PLL approach to jitter reduction. Finally, in this review article, in listening tests, Harley reports that, with the Z-Link (using the AD1890) "The bass was undeniably improved, as was soundstage focus and clarity. In terms of low-level resolution, timbral liquidity, and bloom around instrumental outlines, I preferred the sound of all the processors and transports under audition, without the Z-Link."

Larry,

lgruber@bih.harvard.edu

My response to Larry became something of an essay which I hope you will also find enlightening. NT

The AD1890 and Robert Harley's continuing comments in Stereophile

I really enjoy reading Robert Harley's reviews in Stereophile. He always describes the internal workings and chip sets of the device under review which as a card carrying gear head is what I really want to know. When it comes to the subjective part of his work, I take that with the same grain of salt pinched on any subjective review--be it literature, music, food, movies, or audio. I've been disappointed and left puzzled too often in the past to believe my personal experience will map one for one onto Robert Harley's or any other critic. Like any journalist, after we have read RH for awhile, the personal bias, preferences and causes surface. In Harley's case, its apparent the fellow really likes DACs using the UltraAnalog parts. We also know he has been the most vocal proponent of HDCD at Stereophile, which itself is the industry's strongest friend of Pacific Microsonics. That equals PacMicro's PDM1000 8x oversampling HDCD filter chip (still the one and only way to build a HDCD capable DAC or CD player). So Bob's perfect DAC becomes an 8x oversampled unit with the PDM1000 feeding data to some UltraAnalog DACs. And, sure enough, we see this architecture getting consistently high, even rave reviews from Bob H. And, I know from personal experience these do sound quite good, especially fed from $8,000 transports and $1,500 jitter filter/DSP boxes. Now along comes AD1890, a square peg when Bob has nothing but round holes. We can not put the PDM1000 upstream of the AD1890 because it only outputs 8x oversampled data. Nor can we put the AD1890 upstream of the PDM1000 because the super secret HDCD mystery bits will be lost. And none of the big guns in the High-End community use it so it is an easy target for pot shots.

One dilemma Bob Harley faces with his public questioning of AD1890 is its inclusion in a long standing Class A Stereophile Recommended Component. Referring to page 141 of the April 1996 issue, the fourth product under 'Digital Processors Class A' is the Digital Domain VSP Model S sample rate converter and jitter filter which uses the AD1890. The Z-Systems Z-Link made it to Class C. Given the VSP Model S's 3x higher price, I suspect they have the parts budget to get the critical master clock and power supplies just that much better. My take is that once a technology has shown its merit in an acknowledged state-of-the-art design, it has 'arrived'. An example of this is the integrated op-amp. For years these were shunned (and rightly so in the bad old days of slow slew limited designs) as inferior. Then modern high speed parts came along and, thanks to designers like Walt Jung and Ben Duncan, we understood how to apply them and achieve stunning results with the little 8-legged bugs. Yet you can still read in print references to discreet designs being inherently superior. All the while finding Class A anointed components like the MLAS No. 38S sitting pretty at the top of the heap using (very well selected and cared for) op-amps. The points being that audio lore dies hard--in hi-fi quality is in the execution, and minute analysis of hi-fi magazines yields inconsistencies (and the rapid onset of boredom).

A second problem Bob Harley faces with his public questioning of AD1890 is his insistence that "the input data and output data are not identical--not a good thing." This type of comment of course has at its root the strongly held high-end ideals of minimal signal processing and circuit simplicity. These ideals have their place when tempered with a dose of reality. Yet like any dogma taken to extreme they enslave and stifle. I believe digital signal processing (DSP) has enabled high-end audio DACs as we now know them and is their future. To dogmatically insist that samples in must equal samples out ignores the present state of the art and slams the door on the future. My background is in the seismic instrument side of the oil industry. Seismic recorders function to store the analog output of geophones. The seismic industry went digital in the 1960s. Present seismic recorders have A/D converters with better than 130 dB dynamic range (it helps when the bandwidth of interest is only a few hundred Hz wide). No one has looked at an unprocessed seismic map to decide where to drill an oil well in 30 years. Supercomputers often process seismic data because the oil industry has spent millions developing and proving the theory of the geosciences and knows the processed data better represents the true map of the underground world. Having witnessed DSP convert the seemingly random squiggles of a raw seismic record into a map of the earth's interior, I suppose I am less put off by DSP than some. As always in life, differing experiences yield differing points of view and belief systems.

If the AD1890 sins because its "input data and output data are not identical" then I have to insist that each and every high-end audio DAC on the market today shares the exact same sin. A provocative, and I hope thought provoking statement, which deserves expansion. In the beginning (of CD) we had two basic approaches to digital audio for the home. The Sony machines used 16 bit DACs to convert the exact 16 bit samples read off the disk at the 44.1k Hz sample rate. The Philips machines used 14 bit DACs (and later 16 bit DACs) to convert not the samples read from the disk but rather a data stream which had been resampled up to a 176k Hz sample rate using a DSP intentionally designed to interpolate the extra samples needed to get from the 44.1k to 176k Hz sample rate. Because the Philips scheme allowed the reconstruction filters to be simpler and further away from the audio band it does less damage to the music. The listeners heard this and the industry moved en mass to oversampling. If there exists on the market today a high-end CD player or DAC which converts the exact 16 bit 44.1k Hz samples from the disk I am unaware of it. Rather we all use oversampling DSPs, with most R2R types settled on 8x, and the pulse width modulation types at 128x, 256x, or more. The purist clinging to his dogma might point out "at least one out of eight samples came from the disk". He can say it but he would be wrong. For along the way we also got better DACs. From 16 to 18 and now 20 bits. In order to have data to feed these puppies the oversampling DSPs are designed to have adjustable output word lengths. We can select an output of 16 or 18 or 20 bit samples and this is set to match the DAC used. A 20 bit output sample is in no way 'identical' to a 16 bit input sample. So "input data and output data are not identical". Not in a $15,950 Levinson, not in a $90 Sony Diskman--nor should it be, for the point is not to deliver identical samples but to deliver the drama and emotion of music. And all of our collective experience gained in dragging the ear piercing sound of those early '80s players into the present state-of-the-art has shown we can do things to the data while in the digital domain which result in truer reproduction of our favorite artists for the delight of our ears.

So Mr. Harley ignores the fact that all of his preferred and recommended DACs make no effort to output samples identical to their input. Least of all his favored son HDCD with its buried data tagging along on the data samples and top secret processing. What of his oft repeated (by him and others) criticism of the AD1890:

"If a jittered signal produces the right amplitudes at the wrong times, an AD1890-processed signal produces slightly wrong amplitudes at the right times." This statement shows a lack of complete understanding of how a AD1890 works and is used in a DAC system. The AD1890 allows us to feed data into it using one clock and read data out using a second ASYNCHRONOUS clock. How this all works will be explained later, but for now I will simply assert: "If a jittered signal produces the right amplitudes at the wrong times, an AD1890-processed signal produces lower over all system distortion by delivering the right amplitudes at the (asynchronous) right time obediently upon demand from a spectrally pure local master clock."

AD1890 and AD.

Before expounding on how the AD1890 can lower distortion, what about Robert Adams' comments? After all he was head of the AD1890 design team at Analog Devices and is a Fellow of the AES. Given that we have only a second hand report of his comments, perhaps taken out of context, I prefer not to speculate on what he said or why he said it. I will point out that Adams' published article in 'The Audio Critic' (Issue 21, "Clock Jitter, D/A Converters, and Sample-Rate Conversion") was strongly in favor of the use of sample rate converters in this role. Analog Devices data sheets and white papers on AD1890 all advocate its use as jitter filter. Finally the latest audio DAC from Analog Devices (AD1859) on which Dr. Adams was again design chief integrates many of AD1890's technologies right into the DAC chip itself.

I can not close out the discussion of others comments on AD1890 and go onto discuss the details of my own product, the X-DAC, without pointing out a bit of irony. I initially became interested in using the AD1890 after reading about its announcement in Stereophile's Industry Update section. Following up on this lead, I tried the AD1890 and liked it so much I designed it into my new companies' premier product, the X-DAC. Only now to find AD1890's most influential critic a Stereophile writer. Fetch my rusty old armor, I have a windmill to charge!

AD1890 from pipe dream to real world systems.

In May of 1993 when I first obtained beta samples of AD1890 chips I shared the concern that the DSP functions which allow it to filter jitter and sample-rate-convert would introduce audible side effects. Thus the first order of business was to insert it in a known DAC and listen. I had been refining a design I dubbed the X-DAC based on the Crystal Semiconductor CS8412 and CS4328 chip set and selected it as the test mule. For the initial trials I wanted to hear just the effects of passing data through AD1890. To separate that from its jitter reduction capability I wired the test adapter to retain a synchronous master clock. To be specific, the CS8412's PLL derived master clock output was used throughout the test DAC to clock data both into AND out of AD1890 as well as the all important master clock for CS4328 DAC chip. Please recall that in PWM DACs like the CS4328, the most important input signal as far as jitter is concerned is the DAC's master clock (in ladder DACs the slower Latch Enable clock is the more critical). Thus with its jitter reduction trump card removed the best I hoped for AD1890 in that first test was that it be inaudible. I though it quite likely it would have some negative side effects, after all the output samples were different from the input! It so happened that the test circuit was ready while my listening room was dismantled for painting. To get by I had setup my alternate speakers, Stax ELS-F81 (first series) in the bedroom driven by a pair of mono connected McIntosh 225 tube amps. Listening in the near field it was like sitting in the worlds biggest set of headphones! Detail & Resolution was the name of that game. Imagine my surprise as it became clear that contrary to my expectations, the AD1890, even in synchronous mode, improved the sound of the X-DAC! I build my development DACs with lots of sockets and adapters so I could switch back and forth and listen with and without AD1890 inline. My intellectual understanding of the time told me the best a synchronous connected AD1890 should be able to do was a draw, yet my ears begged to differ. I was very intrigued to hear the improvements as they were in the exact areas which, up to then, had held the CS4328 behind the best results I had obtained using ladder DACs fed into tubed analog stages. To be specific, the CS4328's subtle lack of midrange transparency and 'fat' bass were greatly improved upon. I sent an excited e-mail to Robert Adams detailing these results which as I recall he politely dismissed as yet more audio maniac rants. Nevertheless with the zeal of a new convert I developed the CS8412-AD1890-CS4328 circuit. The addition of a low jitter local master clock to feed the CS4328 and clock data out of AD1890 asynchronously yielded a leap in performance beyond any of my ladder type DACs. With further refinements this became the X-DAC 3.0.

After three years experience with the AD1890 and X-DAC 3.0 along with much discussion on 'the net' and elsewhere I believe I know why AD1890 shines so brightly as a system with the CS8412 and CS4328. What one hears is the effects of both jitter reduction AND going from 16 to 18 bit input for the CS4328 yielding an overall system resolution enhancement. This is due to two facts. First, all samples produced by the AD1890 are in 24 bit format and second, its allowing of the use of an asynchronous spectrally pure local clock.

Highly integrated DAC chips like the CS4328, which combine their oversampling filters with the DAC in a single chip, have many advantages and, until now, one disadvantage. Advantages include the ability to design exactly complimentary filters and the greatly reduced RFI effects inside and outside the system. The disadvantage they bear in comparison to separate filter and DAC solutions has been the lack of bits beyond the 16th at the DAC's inputs. Any modern 8x oversampling filter chip likes NPC's, PM's, or Sony's can be configured to produce samples with a 16, 18, 20, or greater bit resolution to match the following DAC. Likewise, solutions using general purpose DSP chips are easy to program for greater than 16 bit samples at the output. Prior to the AD1890 the only way to exercise the CS4328's full 18 bit resolution has been to feed it 18 to 20 bit master tapes! Now with the AD1890's 24 bit samples resulting from computations at 27 bit accuracy we can hear the results in its full 18 bit glory. I believe this is the improvement I heard in my bedroom back in 1993. Skeptics and purists will complain that truncating the 24 bit samples from the AD1890 to 18 bits at the CS4328's input (it actually just loads the top 18 MSBs and ignores bits 19 to 24) is a known 'no no' in digital hi-fi. This bothered me also, yet I continued based on the evidence of my ears. Frankly it was not until the X-DAC 3.0 was completed and in production that it occurred to me why the AD1890 and CS4328 are such a synergistic pair. The CS4328 is a pulse width modulation (PWM) DAC. The PWM DAC has only been made applicable to hi-fi applications due to another DSP trick known as the noise shaper. In PWM DACs the multi-bit input sample is greatly oversampled while being decimated down to a single bit, or Pulse, the Width of which is modulated to recreate the desired analog signal. If done simply, this would result in a huge amount of audible noise. However, the redeeming technology for hi-fi PWM DACs is noise shaping, so called because within the DSP the noise is shifted to the ultrasonic range beyond the audible spectrum where it is easy to filter out. It is my theory that this noise shaping within the CS4328's onboard digital filter effectively rounds off and dithers the truncated signal it reads from the AD1890. Thus we are treated to our 18 bit cake and get to eat it to!

This brings us to the second big advantage bestowed by AD1890--true asynchronous operation for outboard audio DACs. By now the negative effects of jitter on digital hi-fi music reproduction are well known (see references 1,2,4-8). Likewise the compromises inherent in the S/PDIF AES/EBU single wire serial interface have been analyzed and thoroughly documented (reference #3). And Robert Harley along with other reviewers continue to document just how monumental a technical challenge it is to extract a spectrally pure clock signal from our standard S/PDIF signal to the desired pS resolutions within synchronous DACs. All this attention to jitter is justified because available audio frequency DACs have improved to the point that the jitter on their clock inputs can very well become a dominate factor in the remaining distortion. The lesson that has been learned is that it is not enough to supply a state-of-the-art DAC with a high quality data stream, clean power, and inspired analog sections if the clocks are contaminated with jitter. Because the AD1890 allows its output samples to be clocked out asynchronously the clock driving its output can be derived from a fixed frequency crystal clock circuit rather than a voltage controlled oscillator (VCOX) as part of a phased locked loop. The advantage here is two fold. First oscillators built with quartz crystals can be designed to have lower inherent jitter. After all, the very nature of a VCOX is to have a changeable output in response to its control voltage. Second, and I believe this is more audibly significant, what jitter remains in a fixed frequency crystal clock circuit will NOT be correlated with the musical signal as it is in typical synchronous DACs. Technically this results in a lowering of intermodulation distortion. To our ears the result is a dramatically clearer, cleaner, more transparent and precise reproduction.

References on Digital Audio & Jitter

1. A Digital Discourse, Dr. Malcolm Hawksford; HiFi News & Record Review Feb, April, June, Aug, 1990
2. The Effects of Sampling Clock Jitter on Nyquist Sampling ADCs, and on Oversampling Delta Sigma ADCs, Steven Harris; Journal Audio Engineering Society, July 1990
3. Is the AES/EBU-S/PDIF Digital Audio Interface Flawed?, Chris Dunn & Dr. Malcolm Hawksford; Audio Engineering Society preprint
4. Jitter: Specification and Assessment in Digital Audio Equipment, Julian Dunn; Audio Engineering Society preprint
5. The Jitter Game, Robert Harley; Stereophile January 1993
6. Jitter & The Digital Interface, Remy Fourre (Ultra Analog); Stereophile October 1993
7. A Transport of Delight, Robert Harley; Stereophile Nov. 1993
8. Clock Jitter, D/A Converters, and Sample-Rate Conversion, Robert Adams; The Audio Critic Issue 21

Preprints are available for $5 ea. from:
Audio Engineering Society
60 East 42nd Street
NY NY 10165 USA.