Voltage reference in a TM4313 GPSDO. Why?

[tverbeure]

After checking out countless hobby GPSDO projects online and planning to build one, I gave in an bought a $95 TM4313 on eBay. It's working well enough for my needs (because I basically don't have anything that needs to be measured with high accuracy.)

I did a teardown of the thing (blog post is here: https://tomverbeure.github.io/2023/07/09/TM4313-GPSDO-Teardown.html), and one thing that struck me is the fact that they're using a MAX6192 voltage reference to create a stable 5V power rail for the OCXO and the opamp that is part of the PWM-based DAC.



I don't think I've seen that in any of the popular hobby GPSDO designs, such as the Lars design and the budget GPSDO, or really anything else.

It's weird that one of the cheapest commercial GPSDOs available uses a pretty expensive component ($1.73 on LCSC at 1000+ volume) for no good reason.

IMO it's definitely not needed for the OCXO power supply itself (they usually have a very low temperature coefficient), but it's also used for an opamp that buffers the result of the PWM R/C filter. So maybe there?

Still if others can get away without it and get results that are probably better than the TM4313...

Thoughts?

[MasterT]

Likely reason is to have low noise power rail, to null out phase jitter of  the main oscillator. Looking into  DS of max619x I see miserable 25 uVp-p & 65 uVrms , what make me think that:
- design was altered later on with wrong part number;
- or marking is a fake.
 Testing noise performance of the ADC, I discovered that the best price/ noise spec. ratio has TL431B from OnSemi, parameters close to LTC6655 but cost 1/50.

[PCB.Wiz]

Likely reason is to have low noise power rail, to null out phase jitter of  the main oscillator. Looking into  DS of max619x I see miserable 25 uVp-p & 65 uVrms , what make me think that:
- design was altered later on with wrong part number;
- or marking is a fake.
 Testing noise performance of the ADC, I discovered that the best price/ noise spec. ratio has TL431B from OnSemi, parameters close to LTC6655 but cost 1/50.

Interesting.  I would also expect an OCXO to have a oscillator regulator inside the package ?

I did see a claim about regulator noise effects on oscillator here (likely not for a crystal VCXO)  :
https://www.nisshinbo-microdevices.co.jp/en/about/info/20230404.html
They compare  'others' 135uV regulators with theirs 8uV

[srb1954]

+1 for the MAX6192 supplying a "low noise" reference voltage for the OCXO control loop.

On the face of it the MAX6192 would appear to be a poor choice for such an application. They have installed a precision, low drift reference but neither of these factors is relevant to the circuit operation as the GPSDO feedback loop will take out any errors related to the initial reference voltage tolerance and any slow thermal drift. So the money spent on this relatively expensive reference chip is largely wasted.

What the control loop can't do is compensate for any low frequency noise on the reference voltage, particularly in the flicker noise region under 10Hz, as the GPSDO control algorithm is likely to be too slow to respond and cancel out such noise. As MasterT points out the MAX6192 has a less than stellar noise performance in the critical low frequency range <10Hz. It is even worse above 10Hz but the PWM filter will take out most of the noise in that region.

It might also pay to take a closer look at the PWM filter. If they have used 5 equal RC stages this will give the worst performance for a filter in terms of settling time and ripple rejection. This due to interactions between stages with subsequent stages loading down the previous stage. The preferred option is to isolate each stage with op amps or, as a minimum, step up the impedance level at each subsequent stage so as to minimise the loading on prior stages. It would be instructive to do some SPICE simulations on the performance of this part of the feedback loop.

One final comment on the PWM filter is the use of ceramic caps for this critical part of the circuit. With the values and sizes of these caps they are likely to be X7R dielectric or similar. Such dielectric materials are subject to microphonics and will generate spurious voltages in response to vibrations. These spurious voltages will generate sidebands on the output of your OCXO which can't be removed by the GPSDO control loop. If you are using this GPSDO in any application where phase noise is critical don't place the unit on top of any equipment with a fan in it or with a buzzing power transformer.

[tverbeure]

On the face of it the MAX6192 would appear to be a poor choice for such an application. They have installed a precision, low drift reference but neither of these factors is relevant to the circuit operation as the GPSDO feedback loop will take out any errors related to the initial reference voltage tolerance and any slow thermal drift. So the money spent on this relatively expensive reference chip is largely wasted.

Ok, so at least I'm not alone in questioning this choice.

It might also pay to take a closer look at the PWM filter. If they have used 5 equal RC stages this will give the worst performance for a filter in terms of settling time and ripple rejection. This due to interactions between stages with subsequent stages loading down the previous stage. The preferred option is to isolate each stage with op amps or, as a minimum, step up the impedance level at each subsequent stage so as to minimise the loading on prior stages. It would be instructive to do some SPICE simulations on the performance of this part of the feedback loop.

The 4 R value are definitely the same: 33k. I measured at least 2 caps with a value of 2.5uF, while the other was 3.5uF or something of that order. I measured this in-circuit with a regular multi-meter, so who knows how accurate those C measurements really are.

The hobby GPSDOs that I've looked at, like the Budget GPSDO (https://www.eevblog.com/forum/projects/budget-gpsdo-a-work-in-progress/msg3685390/#msg3685390) or the Lars GPSDO (https://www.eevblog.com/forum/projects/lars-diy-gpsdo-with-arduino-and-1ns-resolution-tic/msg1278672/#msg1278672) also use R/C filters, without even an opamp.

Wouldn't it be true that the settling time and ripple rejection are not a big deal when dealing with huge time constants and update rates that are measured in seconds or even minutes?

Such dielectric materials are subject to microphonics and will generate spurious voltages in response to vibrations. These spurious voltages will generate sidebands on the output of your OCXO which can't be removed by the GPSDO control loop. If you are using this GPSDO in any application where phase noise is critical don't place the unit on top of any equipment with a fan in it or with a buzzing power transformer.

It's hard to see in the picture, but during normal operations, the GPSDO is permanently tucked away in the drawer of my desk, to reduce temperature variations. My experience with these kind of small OCXOs is that they are quite sensitive to changes in ambient temperature when compared to, say, an oscilloquartz 8663. Of course, it's already in an enclosure so that may be less of a concern, but the drawer is a good way to keep my desk less messy and it won't hurt.

Of course, I'm not using it at all for anything critical. I'm just playing with test equipment and learning.

[srb1954]

It might also pay to take a closer look at the PWM filter. If they have used 5 equal RC stages this will give the worst performance for a filter in terms of settling time and ripple rejection. This due to interactions between stages with subsequent stages loading down the previous stage. The preferred option is to isolate each stage with op amps or, as a minimum, step up the impedance level at each subsequent stage so as to minimise the loading on prior stages. It would be instructive to do some SPICE simulations on the performance of this part of the feedback loop.

The 4 R value are definitely the same: 33k. I measured at least 2 caps with a value of 2.5uF, while the other was 3.5uF or something of that order. I measured this in-circuit with a regular multi-meter, so who knows how accurate those C measurements really are.

The hobby GPSDOs that I've looked at, like the Budget GPSDO (https://www.eevblog.com/forum/projects/budget-gpsdo-a-work-in-progress/msg3685390/#msg3685390) or the Lars GPSDO (https://www.eevblog.com/forum/projects/lars-diy-gpsdo-with-arduino-and-1ns-resolution-tic/msg1278672/#msg1278672) also use R/C filters, without even an opamp.

Wouldn't it be true that the settling time and ripple rejection are not a big deal when dealing with huge time constants and update rates that are measured in seconds or even minutes?

The settling time may not be an performance issue when you have a slow control loop but it may still need to be taken into account when considering the stability of that negative feedback loop. The time delay through the filter adds extra phase shift to the feedback path and potentially could degrade the phase margin of the feedback loop. This may or may not be a problem depending on the implementation of the control algorithm in the processor.

Any extra ripple on the O/P of the filter is more problematic as this will modulate the OCXO and produce extra phase noise on your 10MHz O/P; the control loop is too slow to do anything to reduce this extra phase noise. If you can filter the PWM signal and attenuate the control voltage ripple down to a point where it has negligible effect, compared to the inherent internal phase noise of the OCXO, then the overall GPSDO performance will be better for phase noise critical applications.   

I am not saying that you can't use cascaded RC filters in such an application but that if you do you such a configuration there are some steps you can take to get better performance. People tend to throw together N cascaded RC stages and think that they will get N times the performance of a single RC stage. This is not the case due to interactions between stages and you get a very soft initial roll-off in the frequency response and increased settling time. You can get a little better performance by scaling the impedance in later stages so that there is less interaction between stages. For example the 5 identical stages of 33k resistor and 2u5 capacitor could be better implemented with 5 different stages of (33k, 2u5), (120k, 680n), (560k, 150n), (2M4, 33n) and (10M, 8n2). An additional advantage of having smaller capacitors in the later stages is that NP0 capacitors can be used, which are less susceptible to microphonic problems.

My comments about the performance of cascaded RC filters are more of a general nature and not necessarily just specific to GPSDO designs.  Again, it is worthwhile playing around with SPICE to compare the effect of different filter configurations.

[tverbeure]

I am not saying that you can't use cascaded RC filters in such an application but that if you do you such a configuration there are some steps you can take to get better performance. People tend to throw together N cascaded RC stages and think that they will get N times the performance of a single RC stage. This is not the case due to interactions between stages and you get a very soft initial roll-off in the frequency response and increased settling time. You can get a little better performance by scaling the impedance in later stages so that there is less interaction between stages. For example the 5 identical stages of 33k resistor and 2u5 capacitor could be better implemented with 5 different stages of (33k, 2u5), (120k, 680n), (560k, 150n), (2M4, 33n) and (10M, 8n2). An additional advantage of having smaller capacitors in the later stages is that NP0 capacitors can be used, which are less susceptible to microphonic problems.

I'll keep that in mind when (if?) I design my own GPSDO. It's one of those things for which it's quick and easy to throw a Spice simulator together and observe the behavior when you make changes.

[MasterT]

Likely reason is to have low noise power rail, to null out phase jitter of  the main oscillator. Looking into  DS of max619x I see miserable 25 uVp-p & 65 uVrms , what make me think that:
- design was altered later on with wrong part number;
- or marking is a fake.
 Testing noise performance of the ADC, I discovered that the best price/ noise spec. ratio has TL431B from OnSemi, parameters close to LTC6655 but cost 1/50.

Interesting.  I would also expect an OCXO to have a oscillator regulator inside the package ?

I did see a claim about regulator noise effects on oscillator here (likely not for a crystal VCXO)  :
https://www.nisshinbo-microdevices.co.jp/en/about/info/20230404.html
They compare  'others' 135uV regulators with theirs 8uV

Don't know about regulator inside the package. I was referring to this diagram linked in the OP:
https://tomverbeure.github.io/assets/tm4313/TM4313_schematic.svg
Any noise from power rail easily gets into Tune pin via PWM-inverter or even OPA PSRR weakness.

[MIS42N]

Just came across this post. The hardware in this design bears a lot of resemblance to my own design https://www.eevblog.com/forum/projects/budget-gpsdo-a-work-in-progress/. I also have provision for a MAX voltage source to help stabilise the tuning voltage. I use the OSC5A2B02 which I think is similar to OC5SC25. The information in the teardown is wrong, the OSC5A2B02 is a 5V HCMOS square wave output.

The OSC5A2B02 has very sensitive V/Hz tuning, 1 part in 10^-6/V. Higher spec OCXO are around 1 part in 10^-7/V. So the OSC5A2B02 needs a stable tuning voltage. I tried 2 solutions, one was to split the power supply so the OCXO and tuning voltage come from one supply (on the basis that the OCXO is pretty constant current draw so the supply voltage would not fluctuate short term), the other was to supply the tuning voltage from its own stabilised supply. Some sort of isolation is required because there is enough fluctuation in a common supply rail to upset the OCXO. Such as attaching a lead, plugging in a USB etc.

It turns out that the quality of GPS signal is dominant. Either supply option will hold better than 1 part in 10^-9 accuracy (i.e. 10MHz±0.01Hz) with poor GPS signals, improving to better than 1 part in 10^-10 with good quality signals. The choice of a tuning voltage supply does seem to make a difference with excellent GPS (e.g. from a well positioned U-blox lea-M8T) when it is in the 2 to 4 parts in in 10^-11 range. I haven't tested extensively as my original design parameter was 'cheap' and an M8T second hand is the same cost as the whole of the GPSDO excluding the MAX and including a neo-7m. 1 part in 10^-10 is better than most people need. I did some tests with an M8T board but I haven't modified it yet to interface properly.

[GigaJoe]

my guess .....    using such Vref,  it allows to operate without antenna with reasonable accuracy. when it transferred to the place without gps signal, after it takes some time to adjust itself.

location of Vref under OCXO can ,  influence by temperature, so it indirectly warm up, decreasing affect of environment temp. fluctuation ... ;  I would say it a good design.

if possible i would suggest a little mod of shielding ocxo by additional thermal isolation, foam , cellulose, anything to minimize  heat dissipation to environment,  it will minimize jitter and save some energy,  increase quality of adjustment.

when turn on  - standalone oscillators usually stop drifting about 2 weeks ;  after power interruption , a few hours.  for Vref , a bit worse - initial drift a few hundred hours, and can be jump in Vref after power interruption.

tl431 extremely unstable comparing to MAX , and it nothing to do with output noise, stability of  Vcorr  has to be fraction of millivolt

[MIS42N]

my guess .....    using such Vref,  it allows to operate without antenna with reasonable accuracy. when it transferred to the place without gps signal, after it takes some time to adjust itself.

location of Vref under OCXO can ,  influence by temperature, so it indirectly warm up, decreasing affect of environment temp. fluctuation ... ;  I would say it a good design.

if possible i would suggest a little mod of shielding ocxo by additional thermal isolation, foam , cellulose, anything to minimize  heat dissipation to environment,  it will minimize jitter and save some energy,  increase quality of adjustment.

when turn on  - standalone oscillators usually stop drifting about 2 weeks ;  after power interruption , a few hours.  for Vref , a bit worse - initial drift a few hundred hours, and can be jump in Vref after power interruption.

tl431 extremely unstable comparing to MAX , and it nothing to do with output noise, stability of  Vcorr  has to be fraction of millivolt

These are good thoughts.

I find the OSC5A2B02 is not much affected by temperature and it is not necessary to add thermal isolation except to reduce the current draw.

The OSC5A2B02 does not have good retrace (i.e. if you turn it off and on with the same V_corr it doesn't always return to the same frequency). It will be close - usually within 0.1Hz. Moral, don't turn it off.
It also doesn't like to be moved. Turn it upside down the frequency changes. Moral, don't move it.

"stability of  Vcorr  has to be fraction of millivolt" YES - a change of 1mV changes the frequency by 1 part in 10^-9.

It is unrealistic to expect a $5 OCXO to perform like a $100 OCXO. They are not the best, but amazing performance for the price.

[CBrandin]

In case anybody is interested there is a way to modify the TM4313 to replace the 1PPS output with a 10MHZ TTL output. The mod is easy and, as you can see, a fairly decient TTL signal is produced. Note, however, that you should probably not use the sine and TTL outputs at the same time. I tested with my PSG2400A signal generator's external clock input (which is TTL) and it worked fine.

[andre_p]

Hey guys,

reading the last post I was wondering if there is a way to keep the 1PPS output but instead of having a 5V level get a 3.3V reference?

Thanks
Andre

[kray]

Some more info about TM4313 GPSDO https://rkraevskiy.blogspot.com/2024/01/tm4313-gpsdo.html