Mike, Mike Monett wrote: > The Allen deviation is used to describe the performance of a stable > clock. Measuring the performance of a good clock requires a counter > with resolution down to picosecond levels. As Dr Griffith points > out, some modern counters may have internal signal processing that > makes them unsuitable for this task.
Hold on! What modern counters do is to use various means to improve on frequency and period measures. This makes the frequency and period measures unsuitable for futher processing as well as evaluation of expected performance when doing measures for Allan Deviation. However, what we do is usually not that measure, we to Time Interval measures for individual trigger points. When doing those measures these smoothing methods cannot be utilized. Then you are running on the bare-bone hardware performance with only the normal (traditional) translation skews. I specifically cautioned Ulrich from making Allan Deviation performance estimate from the frequency performence for this reason. The smoothing will make such rule-of-thumb comparisions much harder. > Another thread discussed using a mixer to generate the difference > frequency between two oscillators, then measuring the stability of > the resulting beat note: > > http://www.febo.com/pipermail/time-nuts/2005-July/019006.html > > The basic principle is sound. If the oscillators were running at > 10MHz, and the frequency difference was 1 Hz, then the beat note > would be 1 Hz. > > This represents one part in 10 million, or 1e-7 of the original > frequency. If the beat note is measured with 1 microsecond > resolution, the overall resolution is 1e-7 / 1e-6 = 1e-13. This is > beyond the capability of most commercial counters. > > The difficulty with this approach is the output of a mixer is at a > fairly low level, perhaps 50 millivolts or so. The frequency would > also be very low, perhaps 1 Hz. This means the counter would have to > trigger accurately on a very slow-rising, low amplitude signal. This is not the actual problem. The actual problem is the slew rate of the signal. Even if the amplitude was several volts peak-to-peak the slew rate of the beat note is the main problem as the wideband noise of at the output added with the wideband noise of the counter input causes a random additive voltage modulation which can pre/post trigger around the ideal position with a RMS value of t_jitter = N_total / SR (this is a traditional trigger jitter formula). The gain stages / slew rate amplifiers that Bruce and I have discussed contributes a significant gain which significantly goes beyond what a can come out of a mixer. Signal is clipped and filtered in order to improve signal to noise properties such that a minimal of noise is amplified while the slew rate is raised significantly. > This is a very difficult measurement problem. The accuracy will be > degraded by noise, such as the 60Hz AC line frequency and its > harmonics, switching noise from the pc power supplies and monitors, > radiation from nearby fluorescent lighting, plus thermal noise from > the mixer and input stage of the amplifiers. Not too hard really. The thing which makes it complex is that good signal to noise is needed both at the carrier frequency and beat frequency. Some knowledge of suitable measures should give adequate measures. > This low-level noise is very difficult to eliminate, especially when > coax cables are needed to transfer the desired signal from one place > to another. The result is the measurement system is not as good as > it could be. Is it? Fighting ground loops to handle H fields is no big magic. Using mixers which ports is galvanically isolated helps. E fields is easier to handle at lower frequencies. For the output port, the difference frequency needs the signal to noise properties. Traditional diffrential signal handling deals with both E and H field issues to such a level that other sources will dominate. It should also be pointed out that carefull adjustment of both input port levels and the loading on the output port will have impact on performance as recorded in literature. > There is a solution to this problem. Another kind of mixer called a > "digital mixer" is ideally suited for this application. It uses a > d-flop, with one signal going to the clock pin, and one going to the > "D" input. The resulting signal on the "Q' output is the frequency > difference between the two signals. > > The output signal is a full logic level swing, perhaps 5 Volts, with > a risetime of a couple of nanoseconds. This is an ideal signal to > pass on a terminated coax cable to the counter. The schematic and > waveforms are shown in the attached GIF. You will not solve the requirements for good dynamics. The digital input is highly non-linear and thus behaves like a mixer so due care is still needed, both at the beat frequency and carrier frequency. The benefit is the high slew rate. It behaves like a sample and hold system, but with the quantization occuring before the sample action rather than after, which would could debated which is best, but the sampling action is certainly the mixing action causing problems, regardless of methods it is realized through. Regardless of method, to achieve performance, you would need to maintain a certain degree of signal hygene to achieve the inteded or possible limit of the system. > The output of the first d-flop is passed to a second d-flop to > eliminate glitches due to metastability in the first stage. This can > occur when the signal on the "D" input is exactly on the switching > threshold when the clock transition occurs. The resulting glitch can > severely disrupt the following logic stages. > > In practice, it might be difficult to offset two stable oscillators > by 1 Hz. In this case, the frequencies can be multiplied to some > higher value. For example, the frequencies could be multiplied by a > factor of 10 to 100MHz, and offset by 1 Hz. > > There may be some jitter in the leading edge of the beat note since > the d-flop may or may not catch the transition as it crosses the > threshold on the "D" input. Instead of the standard +/- 1 clock > ambiguity in digital circuits, the output could be several clocks > late. However, if the counter had a resolution of 100 nanoseconds > (10MHz clock), the extra delay is much less than the counter > resolution and should have no effect. No, they add up unless you use the CLK signal for couting the beat frequency, in which case this is really the input trigger of a non-interpolating counter. If you use different clocks the beating pattern between them will introduce the additional noise signal. > The overall resolution in this example would be 1e-8 / 1e-7 = 1e-15. > > This is achieved in one second, which is an impossible task for a > counter. This means the Allen deviation can be measured much faster > than before, and with much higher accuracy. > > A simple LTspice analysis is included in the attached ZIP. Where is the added noise? What D flip flops do you use? What is the ref inverter and what properties does it have? It's an analogue world after all, so we need to evaluate it as an analogue system. Cheers, Magnus _______________________________________________ time-nuts mailing list -- time-nuts@febo.com To unsubscribe, go to https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts and follow the instructions there.