[ Home | Staff & Contacts | DIY & Tweaks | Listening tests | HiFi Playground | Music & Books ]

DIGITabilis: crash course on digital audio interfaces

Part 1.5 - Synchronize your... clocks

[Italian version]

Clock jitter measures

For what regards instead clock measurements, the frequencies we are talking about are in the range of 10-40MHz, so a 100MSa/s spectrum analyzer can supply a reasonable spectrum; the issue here is to achieve a reasonable precision on the frequency scale, which means to have at least a 1M (better 10M) points FFT. Unfortunately it is difficult to find low cost (and even reasonable cost) units with such a large sampling memory.

There are however at least two workarounds.

First of all, in case a digital spectrum analyzer with a wide enough bandwidth at any sampling rate and no antialias filter is used, then it is possible to take advantage of the aliasing effect: in practice undersampling the clock, that is using a sampling rate lower than the double of the clock frequency, in the analyzer span (0-Fs/2) will appear a downconverted alias of the clock spectrum.

[jitter]

These aliases can be easily analyzed in detail, given the reduced sampling frequency. This makes a direct clock jitter analysis available even with low cost analyzers.

[jitter]

Unfortunately the dynamic range of low cost spectrum analyzers is very limited, while jitter clock components are expectedly at very low levels. Moreover, some kind of averaging is required to reduce random noise. And in any case, it is hardly enough, especially in case of analysis of high quality clocks.

[jitter]

In these cases, it is probably better to force the analyzer a little into saturation, disregarding the higher harmonics so generated, which are in any case normally present due to the typically squared waveform of clocks. As a matter of fact, a few tests of mine seem to show that the side bands relative level does not change even with hard saturation, possibly due to the fact that jitter affects mainly the timing of the rising and falling edges of signal.

To end with, note that a sound card is not an option, here: any decent sound card has an antialiasing input filter, and the clock frequency is therefore filtered away before reaching the convertion circuits.
As an alternate solution, it is possible to use an RF mixer to combine the clock under exam and a second reference clock. A mixer is a device whose output is the product of the two inputs. If we assume that the two signals are sinusoids with frequencies a and b, for sure you do remember that

cos(a*w*t) * cos (b*w*t) = 1/2 * [cos([a+b]*w*t) - cos([a-b]*w*t)] (Werner)

that is the resulting mixed signal contains two sinusoids, whose frequencies are the sum and the difference of the original signal and reference frequency. By the way...ehm... what are you saying? You did not remember that? very strange indeed... how could you do without it?

Choosing adequately the reference clock it is possible to get any desired center frequency for the difference, in practice translating the original signal without any compression on the frequency axis. The resulting signal can be easily analyzed in detail without the need for a very large FFT.

Obviously the second clock should be perfectly stable: if not, its jitter will add up with a quadratic low to the jitter of the clock under test. The mixer in principle could also be as simple as a single XOR port with an adequately stabilized power supply: such a mixer implemented with an integrated port works perfectly, but its sensitivity/precision is not sufficient for our needs.

Note that in practice the signals are not sinusoids, but square waves, and a little more care is required to be sure to avoid taking harmonics for jitter components...

However, the most important application of this principle consists in making both oscillators have the same frequency.
At this point, the difference frequency goes to zero, and what remains around zero is just oscillators phase noise.

[jitter]

I have used this technique to assess a few audio superclocks, and the results are at least interesting.

This is the measure setup schematic. The mixer must be an RF mixer, but must sport a frequency response extending down to very low frequencies (fraction of herts, ideally DC). All circuits must be powered with ultra regulated and low noise supplies. Special care must be taken to avoid ground loops.

[jitter]

Given the high stability these clocks should have (and normally in facts have), for a visual qualitative assessment of these clocks there is no real need for a PLL. A very fine tunable oscillatore, and a lot of patience, is however required.

You need to wait for both the clocks to reach electrical and thermal stability (which can take as little as few minutes or as much as hours...) and then tune the two clocks in order to see the clock difference signal disappear into zero. The fact that we are mixing square waves, however, helps a lot in achieveing a precise tuning, because all the harmonics interact just like the fundamental, but having a multiple frequency their difference frequency is multiple of the same order too of the central frequencies difference: having often 30-50 harmonics available, you have a natural error amplifier available.

The analysis of the resulting mixed signal can be done with a sound card. However, even a 24 bit one is quite at the limit.

A possible variation on this of this principle makes use of a PLL oscillator, with a huge filter constant in order to make it almost completely stable. If you lock it to the clock under test, the difference frequency will be zero, and you can directy read the jitter spectrum with a simple FFT (including, however, both clocks jitter components, as above).

[jitter]

At this point you can analyze the phase error voltage for example with a sound card or a spectrum analyzer: this as a matter of fact represents the PLL output phase noise, and not the input clock one, but with an adequate design of the circuit and in particular of the low pass and PLL filters frequency response it is normallypossible to extend the analysis to the full range of interest.

Clock Jitter: how much is it important?

One of the most "ancient" and diffused tweakings for CD player consists in the installation of an high quality clock. But how much important is to have a better clock?

This is no easy question, and I am not sure we will arrive at any point in the discussion. Testing is complicated and time consuming, and would require the availability of several units to mod, which is not so easy to have.

According to my experience, using an high quality clock can produce very significant improvements in the sound of a low cost player. Also from the sound point of view, the changes are exactly the ones you would expect in case of a drastic jitter reduction, as described above: essentially a far better pace and sens of rythm, which increase in a very significant way listening pleasure.

I had recently the fortune to be present at a public demonstration of the LClock XO II, one of the many superclocks available (a previous version was also reviewed years ago here on TNT) at Milano HiEnd 2004. Fabio Camorani, LC products distributor in Italy, had set up a system with two externally identical Pioneer players, one in the original version and one with the clock installed. The public had to recognize which was the tweaked one. Even though I must admit that it was not a real blind test, and that the system was not "scientifically" verified, I must say that the differences were absolutely in line with the ones I remembered from the review. In any case, after a couple of source commutations the interested listeners had correctly identified the tweaked unit. This is, in standard exibition condition, a synonym for from worst case, as you know...

This should eliminate any doubt on the validity of this kind of tweaking. But...

Recently, I spent a couple of days in trying to understand how much the setup of a commercial "superclock" could affect jitter at audio level. The clock was NET Audio RockClock (next on these pages), kindly made available by David Pritchard. The RockClock circuit is enclosed in a cast(!) metal box connected to ground, which should make the clock generation circuit widely insensible to the specific positioning. It was mounted in the usual Pioneer PD-S 505, and connected to the motherboard with a 8cm long section of UTP CAT5 internal twisted couple. The clock had its own separate PSU, the NET Audio CD Clock Power Supply.

A note on the measured values. Measuring jitter means to evaluate the level of signals in the range -105 / -130dBFS (0.7-11 microvolt): the numbers you get are not perfectly stable. Each test has been performed and the result recorded from 5 to 8 times. The values presented hereafter are the average value, followed by the minimum and maximum values enclosed in parenthesys, just to give an idea of the spreading of the results.

Initially I started with the units positioned in the easiest way inside the box. Jitter measured at this point was 358ps (340-370ps). However, the clock case was near the power control chips of the transport, which appeared less than optimal.

So I raised the clock just a few centimeters (5-6cm), leaving all the rest as above. Jitter dropped to 334ps (329-341ps).

Moving the clock away from the board as much as possible given the short connection made the jitter utterly go down to 328ps (322-332ps). Having the clock output twisted pair go along the transport power chips had scarcely any effect.

Then I moved the PSU back towards the main power supply output stages and cabling, but there was an immediate increase of jitter up to 346ps (338-351ps), so I moved the PSU into another position nearby: this quieter position was just above the main PSU rectifying diodes.... and don't ask me why this appeared a quieter position! Only supposition is that the power supply rails are subject to very fast current pulses caused both by the digital components and the transport engines; these current pulses are sourced from PSU filter capacitors, which in practice prevent the pulses to reach the rectifiers. Ok, ok, science fiction requires less phantasy, I know...

Now I started trying to optimize the clock output connection with the motherboard. First, I added a 43ohm resistor in series to the hot clock line (as suggested by other clock designers) which should provide some kind of matching of the output impedence to the twisted couple impedence (approx. 100ohm). The results where definitely interesting, with jitter going down to 311ps (302-318ps).

At this point, it was clear that there was also some kind of impedence matching problem. I eliminated the 43ohm resistor, and added a 100ohm termination resistor at the motherboard input pins (the motherboard had been obviously modified by eliminating the crystal and connected capacitors, and inserting in their place two stout pins: a very good idea, given the extra work they have been carrying out these days!)

The result was beyond expectation: the measured jitter went down to 288ps (278-302ps).

Adding again a 43ohm resistor at the input did not change the values significantly.

So, just with apparently very minor changes, it is possible to reduce an already not so high jitter by 20%. Taking into account how much jitter is probably added between the clock and the conversion circuit because of interference and power supply problems, this amount is far more than I could expect.

An even more extreme experience. I was testing my TNT 1541 DAC (temporarily modified to bypass decimation circuit), but I was continually getting quite different jitter figures in what in my opinion was exactly the same configuration. This was strange: actually, even though the figures are not completely stable, they have normally a limited range of variability.

At the end I was able to track the source of the problem: the jitter measurs changed heavily depending on the presence or position of the cover of the cabinet, that in the meanwhile was been moved on and off to find out the reason of the differences. The test has been repeated several times, and the results proved definitely consistent.

[jitter]
Effects of cover of TNT1541 prototype on jitter
Green = iron cover on Cyan = aluminum panel on
Yellow = iron cover, insulated, on Orange = no cover

What does all this mean? That conversion jitter is a wild beast: it is the effect of so many different components that it results extremely sensitive to a huge number of different electrical, environmental and mechanical factors.

So, for a CD player to have a very good pace, a good clock is mandatory, but far from enough: all the treatment of the clock signal throughout the system must be specifically designed to make available the clock where needed in the best conditions, and all power supply lines must be designed to avoid disturbs generated especially by the digital circuits to spread trough them.

These days there are clock designers boasting 2-3ps jitter for their products. From the tests it seems clear that such a clock can give a significant benefit only in the best units, designed specifically to achieve the best performances under this aspect.

In the average or low cost ones, these clocks oustanding pace would well be distorted and polluted by the not so cured clock path, by not correct matching, by the not perfect ground loops, the dirty power supplies, the low cost op-amps and so on.

In the most recent systems, in particular low cost DVD players, clock quality seems to be better than in CD players of just a few years ago. However, the sampling jitter remains rather high (500-600ps), and is essentially made of noise caused by digital or video circuits.

Note also that most of the issues listed above are structural, depending on initial design tradeoffs and motherboard layout. There is very little you can do to solve this kind of problems (or, at least, not that much). In these cases the choice of a very high quality clock (few ps jitter) either at design time or later as a tweak seems just a waste of money.

On the other end of the scale, it is expected that high end designers already include very low jitter clocks in their design. In any case, I do not think that so many people would like to have their own high end player tweaked, unless the tweak is sponsored by the manufacturer. So, in my view, here too there is a question about utility of real "super" clocks.

So, high quality clocks are no use at all? Nothing's more wrong.

There is in facts a huge number of medium quality player, and all low quality players, that can get a really huge benefit from a new clock. The effect is evident, clear, there is no discussion about this. So the utility of these clocks is definitely out of discussion.

A real issue, however, is the general price tag of these units, compared with the cost of the players. In facts, the price of many clocks is higher then the price of an entry level player; what's worse, with DVD players, a complete tweaking would require 2 clock units. On the other side it is out of discussion that the effect of a good clock on an entry level player is absolutely notable, evident; a low cost player with tweaked clock normally sports a far more natural and rythmic sound than an untweaked player of double cost.

So the balance is really hard to draw, essentially depending on the original character of the player and personal tastes of the user.

What is clear, at least IMHO, however, is that, with most medium and average level players, a tweaker using an ultra high quality clock with jitter in the range of few picoseconds without specific instrumentation to check for installation optimization risks to achieve the same results as with any clock with a reasonably low jitter (less than 50ps). But it's also true that with a better clock there are chances to achieve a better result.

Bibliography

Rewind to: Part 1.1 | Part 1.2 | [Part 1.3] | [Part 1.4]

© Copyright 2005 Giorgio Pozzoli - www.tnt-audio.com

[ Home | Staff & Contacts | DIY & Tweaks | Listening tests | HiFi Playground | Music & Books ]