[time-nuts] Re: Build a 3 hat timestanp counter (hans-ge...@lehnard.de)
Hi, It looks as if you have a higher noise floor with the TLV3501. I see two effects, both higher slope (usually but not always due to gaussian noise) and then also a higher systematic noise. The later could be from power-suppy for instance, but any form of RF and LF frequency pickup. The actual noise of the TLV3501 I just fail to spot in the datasheet on a quick look, but also the bandwidth. I assume that the trigger point is good, that is that you trigger on the highest slew-rate point of the curve. For a sine that would be at the through-zero point, but for square-wave it is actually closer to the previous level for actual signals. A DC blocker and keeping the trigger close to zero usually suffice. Then just not loosing amplitude going in, as amplitude convert to slew-rate. You might benefit of doing spectrum analysis on the data to locate RF frequencies and track them down in the analog domain. Consider using amplification stages to increase slew-rate before hitting an input. I remember once a design where the hardware guys had an ECL "comparator" setup so it in one state gave a solid signal but the other acted as a linear amplifier of all the noise on the board. While it may seem like adding hysteresis would cure it, it will only cure it for the non-timing parts (which will be the amplitude part) of the signal where as the timing part would still be affected. Also, hysteresis shifts the trigger point to one which has somewhat less ideal slew-rate for timing purposes. As always the timing/phase and amplitude parts of the signal is on orthogonal parts. Cheers, Magnus On 2022-06-06 15:19, Hans-Georg Lehnard via time-nuts wrote: Hi, I tested the TLV3501 with the HP E1740A TIA and there is a visible difference. First test an OCXO on reference and directly on the input. Second test OCXO via the TLV32501 on the input. ___ time-nuts mailing list -- time-nuts@lists.febo.com To unsubscribe send an email to time-nuts-le...@lists.febo.com ___ time-nuts mailing list -- time-nuts@lists.febo.com To unsubscribe send an email to time-nuts-le...@lists.febo.com
[time-nuts] Re: Build a 3 hat timestanp counter (hans-ge...@lehnard.de)
Hi, I know Michael and his work, we have private email contact and I also have a meter with TDC7200 from him. I would like to solve 2 problems. 1.) I have 4 working Temex MCFRS-01 Rb oscillators (stability up to 4*10E-11 per month) and no better reference against I can measure in the range of seconds. Therefore the idea with the 3-hat principle. 2) I have 2 HP 10811 OCXO and want to measure the short time stability (< 1s) in the range of 10E-12. With the BG7TBL FA2 counter I can make long time measurements over hours against my Samsung UCCM GPS and with the HPE1740A short time measurements up to 1.6s with 50ps resolution. Both problems can not be solved with the FA2 and also not with the Mino counters. I will now try to optimize only one channel as recommended by Magnus and bring the raw timestamps to the Pc and evaluate them there. Then we can see if the faster sampling and averaging can solve my problems. Hans-Georg Am 2022-05-31 22:53, schrieb zfe via time-nuts: > Acam, now ScioSense, has also the 4 channel GPX2 that achieves up to > 35Msps over LVDS. It has the same time resolution as the AS6501. > > Michael Nowak (http://www.mino-elektronik.de/) build nice simple little > counters with the TDC7200 and the AS5601 (the later is not on his > website, I think) that do gapless measurements and have the option to > use internally linear regression (the AS6501 version with more than > 100.000 sps). He kindly provided me with several revisions of prototypes > of both counters for testing. > > In loopback measurements the improvements of linear regression are as > expected. The problem is that the performance of linear regression can > break down when DUT and reference have the same (stable) frequency with > a small unfavorable offset. Unfortunately this is easily the standard > situation if you test two 10MHz oscillators against each other. Things > get worse the shorter the measurement time is. I think the three > cornered hat approach will not necessarily eliminate that problem, it is > more for overcomming the limits of the reference oscillator. > > Achieving a good part of the higher performance is possible if you > adjust the frequencies of DUT and reference carefully and if they are as > stable as supposed. But that is a messy process where I am always in > doubt if the displayed result is the performance of the DUT/reference or > a problem with the counter, unless I do several cross checks. > Best would be the possibility to use reference with a frequency skew to > the DUT. > > I attached a quick example with Michael Nowaks AS6501 counter: > > There are three loopbacks one without linear regression, and two with > linear regression (with 1s and 10s measurement time). > > Then I compare a 10MHz OXCO, with good short time stability, to a well > performing FE-5680 > Without linear regression the Allan deviation below 100s is due to the > AS6501 resolution and not due to the oscillators. > Next measurements with linear regression. > If I match the frequencies to about 1mHz the limits come mostely from > the FE-5680. > If I adjust to 0.5Hz offset the linear regression breaks down to the raw > AS6501 performance and worse. > > To demonstrate that that OCXO performs well with 0.5Hz offset, I use it > as reference to measure a fine 16.384MHz OCXO. What you see in the plot > is, more or less, the performance of the 10MHz-OCXO. > There is also a plot of the 16.384MHz OCXO against the FE-5680, with and > without linear regression. > Combining both you see that linear regression gives improvements - you > get at least about 10^-12 with 1s measurement time. I have no better > 10MHz oscillator to explore the limits with 1s measurement time. > But if you decrease the measurement time you fast get in messy > measurements that give you headache with attributing the culprit. > ___ > time-nuts mailing list -- time-nuts@lists.febo.com > To unsubscribe send an email to time-nuts-le...@lists.febo.com ___ time-nuts mailing list -- time-nuts@lists.febo.com To unsubscribe send an email to time-nuts-le...@lists.febo.com
[time-nuts] Re: Build a 3 hat timestanp counter (hans-ge...@lehnard.de)
Acam, now ScioSense, has also the 4 channel GPX2 that achieves up to 35Msps over LVDS. It has the same time resolution as the AS6501. Michael Nowak (http://www.mino-elektronik.de/) build nice simple little counters with the TDC7200 and the AS5601 (the later is not on his website, I think) that do gapless measurements and have the option to use internally linear regression (the AS6501 version with more than 100.000 sps). He kindly provided me with several revisions of prototypes of both counters for testing. In loopback measurements the improvements of linear regression are as expected. The problem is that the performance of linear regression can break down when DUT and reference have the same (stable) frequency with a small unfavorable offset. Unfortunately this is easily the standard situation if you test two 10MHz oscillators against each other. Things get worse the shorter the measurement time is. I think the three cornered hat approach will not necessarily eliminate that problem, it is more for overcomming the limits of the reference oscillator. Achieving a good part of the higher performance is possible if you adjust the frequencies of DUT and reference carefully and if they are as stable as supposed. But that is a messy process where I am always in doubt if the displayed result is the performance of the DUT/reference or a problem with the counter, unless I do several cross checks. Best would be the possibility to use reference with a frequency skew to the DUT. I attached a quick example with Michael Nowaks AS6501 counter: There are three loopbacks one without linear regression, and two with linear regression (with 1s and 10s measurement time). Then I compare a 10MHz OXCO, with good short time stability, to a well performing FE-5680 Without linear regression the Allan deviation below 100s is due to the AS6501 resolution and not due to the oscillators. Next measurements with linear regression. If I match the frequencies to about 1mHz the limits come mostely from the FE-5680. If I adjust to 0.5Hz offset the linear regression breaks down to the raw AS6501 performance and worse. To demonstrate that that OCXO performs well with 0.5Hz offset, I use it as reference to measure a fine 16.384MHz OCXO. What you see in the plot is, more or less, the performance of the 10MHz-OCXO. There is also a plot of the 16.384MHz OCXO against the FE-5680, with and without linear regression. Combining both you see that linear regression gives improvements - you get at least about 10^-12 with 1s measurement time. I have no better 10MHz oscillator to explore the limits with 1s measurement time. But if you decrease the measurement time you fast get in messy measurements that give you headache with attributing the culprit. ___ time-nuts mailing list -- time-nuts@lists.febo.com To unsubscribe send an email to time-nuts-le...@lists.febo.com
[time-nuts] Re: Build a 3 hat timestanp counter
Hi Magnus, I got the E1740 last year on ebay for 80EUR and wrote a quick to dirty software for it. This is my thread about it: https://www.eevblog.com/forum/testgear/hp-e1740a-time-interval-analyzer-292384/msg3718660/#msg3718660 It has 2 channels and can timestamp any edge up to 80 Mhz. It is possible to measure with one channel, with 2 channels and the difference between them. I hope to see a difference between the direct input and the TLV3501. I also have 2 Acam 2 channel TDC AS6501 with DDR-LVDS outputs and 10ps resolution. With this it should be possible to sample up to 12,5 MHz with a FPGA. This will be my next project. There is also the AS6500 which has 4 channels, but can only be read out via SPI. For the LTC6957 I will use the circuit from here: https://ohwr.org/project/wr-low-jitter/uploads/74305092be60598e9bffd39d59489594/Daughterboard_V2.pdf Page 4. Hans-Georg Am 2022-05-29 18:35, schrieb Magnus Danielson via time-nuts: > Hi Hans-Georg, > > On 2022-05-28 17:04, Hans-Georg Lehnard via time-nuts wrote: > >> Hello Magnus, >> >> I understood that simply sampling 3 channels fast and averaging does not >> solve all the problems ;-). > Sure, I just want to illustrate how various approaches could allow you to get > the most out of the hardware you have. > >> I have a HP E740A time interval analyzer that I might use for my >> oscillators. >> The HP-TIA has 50ps resolution, can sample 10 Mhz directly but has only >> 512K sample memory. For longer recordings I can use the histogram >> function. I still need to do some work on my software for that, but that >> might be the fastest way to get results. > Does the E1740A have a high speed time-stamping port? The 5371A and 5372A > have that as an option, so if one could have some suitable hardware process > on those time-stamps, it would be something. > >> Attached is a picture of my HP10811 and another one that someone made >> for me as a comparison to other TIAs. >> >> The third pic shows another meter i have with TLV3501 and TDC7200 >> without averaging compared to an FA2. >> >> I will keep trying with the TDC7200 and maybe better with the LTC6957 >> and only one channel. > > Yes, to try different approaches. The LTC6957 is a very cool little chip. > > Cheers, > Magnus > ___ > time-nuts mailing list -- time-nuts@lists.febo.com > To unsubscribe send an email to time-nuts-le...@lists.febo.com ___ time-nuts mailing list -- time-nuts@lists.febo.com To unsubscribe send an email to time-nuts-le...@lists.febo.com
[time-nuts] Re: Build a 3 hat timestanp counter
Hi Hans-Georg, On 2022-05-28 17:04, Hans-Georg Lehnard via time-nuts wrote: Hello Magnus, I understood that simply sampling 3 channels fast and averaging does not solve all the problems ;-). Sure, I just want to illustrate how various approaches could allow you to get the most out of the hardware you have. I have a HP E740A time interval analyzer that I might use for my oscillators. The HP-TIA has 50ps resolution, can sample 10 Mhz directly but has only 512K sample memory. For longer recordings I can use the histogram function. I still need to do some work on my software for that, but that might be the fastest way to get results. Does the E1740A have a high speed time-stamping port? The 5371A and 5372A have that as an option, so if one could have some suitable hardware process on those time-stamps, it would be something. Attached is a picture of my HP10811 and another one that someone made for me as a comparison to other TIAs. The third pic shows another meter i have with TLV3501 and TDC7200 without averaging compared to an FA2. I will keep trying with the TDC7200 and maybe better with the LTC6957 and only one channel. Yes, to try different approaches. The LTC6957 is a very cool little chip. Cheers, Magnus ___ time-nuts mailing list -- time-nuts@lists.febo.com To unsubscribe send an email to time-nuts-le...@lists.febo.com
[time-nuts] Re: Build a 3 hat timestanp counter
Dear Hans-Georg, The sooner you can supply ADEV and MDEV phase-plots from your device, we can provide more detailed recommendations and point you to specifics. I try to raise your awareness in dry-simulation before any measurement is available. Please be aware that comparators is inherently slew-rate limited, such that the slew-rate convert the amplitude noise to time-noise, which is later time-tagged by the time-tagger circuits. This follows the classic trigger noise formula: t_n = e_n / SR where t_n is the time-noise, e_n is the voltage noise and SR is the slew-rate. You can stress-test this using two methods: 1) change trigger voltage to alter voltage point and thus slew-rate on signal shape. In practice this is mostly useful as you operate a scope, but very illustrative as you see the fuzzyness increase and decrease due to the trigger noise. 2) alter amplitude of signal, such as inserting a 6 dB damper, which will double the slew-rate limited noise. To complicate matters, there is inherent noise of any channel, so that comes on top of the slew-rate limited noise. Alternate slew-rate to separate the effects is straight-forward. So, with that basic on trigger jitter, I aim to complicate matters for you. You can make the trigger process better than comparators. There is a paper by Collins that covers it, but it has been further investigated. See the contributions to the field available online by Bruce Griffiths, a fellow time-nuts in New Zeeland. The basic reasoning builds on four observations: First, the noise of a circuit depends on the bandwidth of the input. Quite simply, if we have a 300 K noise-source, how much voltage we get depends also on the bandwidth. Secondly, the slew-rate we have is limited by the bandwidth, since the reciprocal of bandwidth provides access to the rise-time and slew-rate is rise-time limited. Third, we can increase slew-rate by using gain. Fourth, each gain-stage will add noise. In a reasoning similar to, but not quite matching up to the Friis formula for noise figure is relevant here. So, rather than going straight into a comparator, which is a high-bandwidth input with high gain, you can have multiple stages of amplification and successively increasing bandwidth to support the slew-rate. Eventually you have a slew-rate so high, that a straight comparator or even digital input will not care. Another complicating factor is that the quantization noise and random noise actually interact. I did analysis on that and presented at a conference. The paper for it is just unreadable and below my standards, but the learnings is important and I will have to revisit the topic. It can actually be good to have more noise than the quantization step has, if you do averaging. Actually, the HP5328A with Option 040, 041 or 042 or a HP5328B will intentionally add noise to the signal to improve the precision as you do averaging. It claims 10 ps resolution achievable if you look in the catalog. Turns out it is better than that, because it assumed a sqrt(N) benefit. Now, observe how I just contradicted myself in methods. This is, in Bob Camps lovely terminology, a "it depends" issue. Which way to go depends so much on where your limits are and how you choose to approach it. You mention that it may be good to separate things. I can state clearly that it does. As these signals go through the same chip, ground-bounce issues cause cross-talk. This cross-talk looks similar to a capacitive coupling between the signals, at the point of most slew-rate, the cross-talk is worst. This cross-talk causes any time-wise nearby signal to shift it's apparent time, causing a non-linear shift of time-difference between the signals. This will cause the RMS error to increase as measured over all phase-relationships, and it looks like apparent lock-up of the oscillators, when it is in fact the measurement device which causes the imperfection. One trick to be used is to use a short bit of coax and simply time-stamp a delayed version independently. This way you can estimate the effect and it's impact on your measurement. This usually leads to work on the counter to reduce coupling. Notice that this non-linearity increases with increased slew-rate. Care in how traces and it's ground-imaging variant is traced, de coupled etc. can significantly help to reduce it. Also, decreasing slew-rate where not needed helps. Again, contradictory to what one would think of. The magical trick of mixing with a signal is that you subtract frequency but maintain phase, which causes the time-difference amplified. This is the core of the DMTD measurement. Now, what bites you is the slew-rate is reduced by the same factor. But sure, to some degree with the other tricks you can gain something. Can be worth testing. Now, with the high-speed ADCs you can approach this differently, convert the IQ samples through arctan, and then decimate the data to arb
[time-nuts] Re: Build a 3 hat timestanp counter
Thanks for your answer and the many suggestions what can be improved. The first picture shows my concept for a prototype. The input shaper consists of a 4:1 transformer with differential output and a TLV3501 comparator. The digital part with divider and start/stop logic fit for all 3 channels into a XCR3064XL CPLD. Maybe it is better to separate the channels later. The MC is a STM32H743 and runs with 450 MHz (pll), clocked from the reference frequency. TDC7200 used as TDC. The measurement with spi readout takes about 5µs, so I decided for 10µs (100 kHz) sample time. The second picture shows the measuring timing inside CPLD. The TDC7200 runs in mode 1 and supplies only the fine time. The MC runs a 10 MHz (reference) counter with 3 capture channels as coarse time. So I only have to read the fine timer and the calibration register from the TDC. The TDC cannot measure from 0, so a reference cycle is added. (t = x +100ns). For the averaging I had thought of a linear regression. Hans-Georg Am 2022-05-25 01:18, schrieb Magnus Danielson via time-nuts: > Hi, > > The first limit you run into is the 1/tau slope of the measurement setup. > This is often claimed to be white phase modulation noise, but it is also the > effect of the single-shot resolution of the counter, and the actual slope > level depends on the interaction of these two. > > So, you might want to try a simple approach first, just to get started. > Nothing wrong with that. You will end up want to get better, so I will try to > provide a few guiding comments for things to think of and improve. > > So, in general, try to use as high frequency as you can so that as you > average down, your sqrt(f/f0) gets as high as possible as the benefit will be > 1/sqrt(f/f0) where f is the oscillator frequency and f0 is the rate after > average. > > As you do ADEV, the f0 frequency will control your bandwidth. > > The filter effect of the averaging as you reduce and sub-sample will help to > some degree with anti-aliasing, but rather than doing averaging, consider > doing proper anti-aliasing filtering as the effect of aliasing into these > measures is established and improvements into the upcoming IEEE Std 1139 > reflect this. In short, aliasing folds the white noise and straight averaging > tends to be a poor suppressor of aliasing noise. > > For white phase modulation (WPM) the expected ADEV response depends linearly > with the bandwidth of the measurement filter. It's often modelled as a > brick-wall filter, which it never is. For classical counters, the input > bandwidth is high, then the sampling rate forms a Nyquist sampling frequency, > but wide band noise just aliase around that. Anti-aliasing filter helps to > reduce or even remove the effect, and then the bandwidth of the anti-aliasing > filter replace the physical channel bandwidth. If the anti-aliasing is done > digitally after the counter front-end, you already got some aliasing > wrapping, but keeping that rate as high as possible keep the number of > overlays low and then filtering-wise reduce it will get you better result. > > For aliassing effects, see Claudio Calosso of INRIM. Great guy. > > This is where the sub-sampling filter approach is nice, since a filter > followed by sub-sampling removes the need to produce all the outputs of the > original sample rate, so filter processing can operate on the sub-sampled > rate. > > As your measures goes for higher taus in ADEV, the significant amount of the > ADEV power will be well within the pass-band of the filter, so just making > sure you have a flat top avoids surprises. For shorter taus, the > anti-aliasing filter will be dominant, so assume first decade of tau to be > waste. > > I say this to guide you to get the best result with the proposed setup. > > The classical three-cornered hat calculation has a limitation in that it > becomes limited by noise and can sometimes result in non-stable results. The > Grosslambert analysis is more robust, since it is essentially the same as > doing the cross-correlation measurement. The key is that you average down > before squaring where as in the three-cornered hat to square early and is > unable to surpress noise of the other sources with as good quality. For > Grosslambert analysis, see François Vernotte series of papers and > presentation. François is another great guy. I spent some time discussing the > Grosslambert analysis with Demetrios the other week. I think I need to also > say that Demetrios is a great guy too, not to single him out, but he really > is. > > There is another trick up the sleeve thought. If you do the modified Allan > deviation (MDEV) processing, it actually integrate the sqrt() trick with > measurement, achieving a 1/tau^1.5 slope for the WPM. This will push it down > quicker if you let it use enough high rate of samples, so that you hit the > flicker phase-modulation slope (1/tau), the white frequency modulation
[time-nuts] Re: Build a 3 hat timestanp counter
Hi, The first limit you run into is the 1/tau slope of the measurement setup. This is often claimed to be white phase modulation noise, but it is also the effect of the single-shot resolution of the counter, and the actual slope level depends on the interaction of these two. So, you might want to try a simple approach first, just to get started. Nothing wrong with that. You will end up want to get better, so I will try to provide a few guiding comments for things to think of and improve. So, in general, try to use as high frequency as you can so that as you average down, your sqrt(f/f0) gets as high as possible as the benefit will be 1/sqrt(f/f0) where f is the oscillator frequency and f0 is the rate after average. As you do ADEV, the f0 frequency will control your bandwidth. The filter effect of the averaging as you reduce and sub-sample will help to some degree with anti-aliasing, but rather than doing averaging, consider doing proper anti-aliasing filtering as the effect of aliasing into these measures is established and improvements into the upcoming IEEE Std 1139 reflect this. In short, aliasing folds the white noise and straight averaging tends to be a poor suppressor of aliasing noise. For white phase modulation (WPM) the expected ADEV response depends linearly with the bandwidth of the measurement filter. It's often modelled as a brick-wall filter, which it never is. For classical counters, the input bandwidth is high, then the sampling rate forms a Nyquist sampling frequency, but wide band noise just aliase around that. Anti-aliasing filter helps to reduce or even remove the effect, and then the bandwidth of the anti-aliasing filter replace the physical channel bandwidth. If the anti-aliasing is done digitally after the counter front-end, you already got some aliasing wrapping, but keeping that rate as high as possible keep the number of overlays low and then filtering-wise reduce it will get you better result. For aliassing effects, see Claudio Calosso of INRIM. Great guy. This is where the sub-sampling filter approach is nice, since a filter followed by sub-sampling removes the need to produce all the outputs of the original sample rate, so filter processing can operate on the sub-sampled rate. As your measures goes for higher taus in ADEV, the significant amount of the ADEV power will be well within the pass-band of the filter, so just making sure you have a flat top avoids surprises. For shorter taus, the anti-aliasing filter will be dominant, so assume first decade of tau to be waste. I say this to guide you to get the best result with the proposed setup. The classical three-cornered hat calculation has a limitation in that it becomes limited by noise and can sometimes result in non-stable results. The Grosslambert analysis is more robust, since it is essentially the same as doing the cross-correlation measurement. The key is that you average down before squaring where as in the three-cornered hat to square early and is unable to surpress noise of the other sources with as good quality. For Grosslambert analysis, see François Vernotte series of papers and presentation. François is another great guy. I spent some time discussing the Grosslambert analysis with Demetrios the other week. I think I need to also say that Demetrios is a great guy too, not to single him out, but he really is. There is another trick up the sleeve thought. If you do the modified Allan deviation (MDEV) processing, it actually integrate the sqrt() trick with measurement, achieving a 1/tau^1.5 slope for the WPM. This will push it down quicker if you let it use enough high rate of samples, so that you hit the flicker phase-modulation slope (1/tau), the white frequency modulation slope (1/tau^0.5) and finally flicker frequency modulation (flat) quicker. The reference levels will be different from ADEV for the various noise-types, but that you can look up in tables and correct for. Cheers, Magnus On 2022-05-24 18:37, Hans-Georg Lehnard via time-nuts wrote: Hi, my Name is Hans-Georg Lehnard from Germany and I'm new here, worked as a developer for hardware then for software and last as a system developer. Now I'm retired and I can play with hardware again ;-). I have: 4 x 20MHz Rubium (TEMEX MCFRS-1), 2 x 10MHz HP10811-60111 1 x Samsung UCCM GPSDO 1 x FA2 counter. lots of OCXO and try to build a house standard that I can trust and qualify my oscillators. Reproducible measurements with the FA2 in 10s precision mode I trust to 10E-11. The short-term stability of the HP oscillators cannot be measured with it, or both are defective. The FA2 is not suitable for short-term measurements of 0.01 ... 1s. For measurements against a reference frequency, the stability of the reference must be 5 to 10 times better than the measured frequency, and I don't have that. Now there are 2 options DMTD mixer or 3-hat measurements. Because I'm a digital person I chose the