High fashion crystal oscillator?

Yup. It's all about the time scale of the Allan variance. ;)

Cheers

Phil Hobbs

For some applications (software defined radio), that's a killer. One must hope for some hint of such behavior in the data sheet, but I've not seen clear indications when perusing such.

The phase noise specs/graph are a pretty good indicator of funny business.

Various cheap XOs vary geatly, like 10:1 or so, in actual phase noise.

Some TCXOs are great at constant temperature, but have lots of phase noise in real life. The thermal tau of the quartz is different from that of the temp sensor.

That would suggest that the temperature sensor was in the wrong place.

The thermal time constants are the weighed sums of the thermal masses aroun d the crystal and around the temperature sensor (weighed according to the t hermal resistance between the sensor or the crystal and the particular ther mal mass being added in). If you put the temperature sensor close to the qu artz crystal they should be much the same, but in practice it pays to put t he temperature sensor closer to the heat source.

Intelligent design would wrap the heat source around the crystal and the se nsor, but this isn't always practical and a lot of "designs" evolve without getting too much intelligent attention.

Yes. I see no reason for that price, given its specs.

SiTime does this with their mems products.

Most xtal oscillators have analog TC compensation. Some might have digital post comp correcting for residual errors.

-- Kevin Aylward

The time constant of the temp sensor is totally independent of the time constant of the xtal. They are different physical objects made from different materials, and so there is no reason why their time constants should be the same. Of course, connecting them together creates a combined thermal circuit which may well have different effective time constants, much like a cascaded RC network has different poles and zeros than the circuits do separately.

The best place place to put the sensor is as close to the xtal as possible. For high performance TCXOs temperature tracking is the dominant error source, the sensor needs to track the xtal temperature within around 10mK.

-- Kevin Aylward

The temp sensor is usually a surface-mount part on a PCB or ceramic substrate, well coupled to ambient, whereas the crystal is delicately suspended with more mass and higher thermal resistance.

Even the crystal itself, bulk cut for low TC, will have interesting transient temperature gradients.

Putting a simple plastic cover over the TCXO (or over an OCXO) will block air currents and greatly improve low frequency phase noise. Less pushing of the various time constants inside the can.

There are, historically, some crystal oscillators that have stationary parts (GC cut comes to mind) where a sensor could be cemeted (and, of course, tuning-fork crystals have a stationary end, too, but bad aging).

The AT-cut crystal structure (usually a disc on delicate wires) isn't a good starting point for the best oscillators (though it was nicely mass-producible for color TVs). My stash of old slab/spring/box quartz crystals of 194x production might actually be good candidates for temperature compensation, and they've certainly aged into the flat part of the bathtub... (just kidding, they'd pick up vibrations something horrible).

he

to

or

ce

It's stuck in close contact with the crystal, so it isn't going to be "tota lly independent".

uch

If the sensor is close to the outside world than the crystal, if may over-c ompensate for external fluctuations, producing unnecessary temperature chan ges at the crystal. With two temperature sensors, with the second one a lot closer to the outside world, you could do better.

e.

That's the best place if you can live with a slow temperature compensation loop.

Putting the sensor closer to the heater means that you can start rolling of f the high frequency gain around the control loop at a rather higher frequ ency.

In reality the crystal, the heat source, the temperature sensor and the the rmal insulation that ought to be wrapped around all of them need to be desi gned together, as an integrated unit.

Manufacturing tends not to like this kind of solution.

But it should be in a place where there aren't much in the way of temperature gradients.

That's the first approximation to improving the thermal isolation between the module and the outside world. A simple plastic cover that blocks air current is good, but it can't get too big.

The Rayleigh number goes up as the cube of the linear dimensions and convection with the box gets going at about 10mm (which you can stop with aerogel and other low mass fillers).

MEMS oscillators can put the resonator and temperature sensor on the same piece of silicon. SiTime claim that their "super TXCO" MEMS oscillators have close in phase noise superior to that of quartz crystal TXCOs because of the better temperature tracking. They market them as OCXO replacements.

I didn't notice any statement of helium tolerance though :) (SiTime MEMS oscs (of a cheaper and lower powered family) were responsible for the iPhone crashes in helium rich atmospheres.)

Allan

whats the issue with helium?

does it temporarily move the frequency or permanently damage the part?

mark

onsdag den 2. januar 2019 kl. 17.46.22 UTC+1 skrev snipped-for-privacy@yahoo.com:

temporarily moves frequency enough to make it stop working for days

.> >it pays to put the temperature sensor closer to the heat source.

Ahmmm...

Indeed. With two sensors one can do something rather special, which I am not currently at liberty to divulge.

There is no heater nor temperature loop in a TCXO. That would be an OCXO.

There are two standard ways to place the sensor in a standard TCXO, today.

First, they all use ASICS.

The cheapest is to use a temperature sensor on the ASIC itself. Typically a diode. This has the advantage of being quite linear with temperature, the disadvantage is that it it is harder to get close to the xtal. An NTC thermistor is easier to place closer to the xtal. It has a larger sensitivity, which can reduce final phase noise, but its non-linearity is another issue to deal with.

The sensor time constant is not really significant compared to the other time constants in the system. Its ~ mH for the main thermal bit. Pure delay is a bit of problem though.

Indeed. It is a key design problem in OCXOs. Having a full time dedicated modelling engineer using tools such as COMSOL is indispensible. Modelling allows for the optimum package design and the creation of an electrical thermal model for simulations.

At Rakon, we spend vast amounts of effort in optimising the control electronics in the oscillator asics I design, and the mechanical structures in order to get < 1ppb over -40 deg to 85 degs with -130 dBc @ 1Hz phase noise.

-- Kevin Aylward

Ahmmm....That certainly does not represent the TCXOs and OCXOs ASICS I design at Rakon. For starters there is no DIP within the main package. Its bare die.

The temperature angle cut depends on the application. For a watch xtal, where temperature is approximately constant on the wrist at ~30deg, its low over a narrow range, but much higher over the full range.

For wide ranging -40 to 85 deg TCXO, the cut gives an approximate 3rd order chebychev response, which has a lower peak to peak variation than the AT cut.

Covers do help.

-- Kevin Aylward

With even one temp sensor, one could tweak the thermal compensation transient response to second-guess the dynamic crystal temperature. That would need a uP to do really right.

The quartz slab is relatively big, and has to be in air or vacuum except for a couple of supports at vibration nodes. That gives it slow and complex thermal time constants, with transient gradients. The gradients could mess up the inherent cut-based tempco of the quartz.

I've seen OCXOs with such bad thermal design, not naming any names, where there was so much lag from the heater to the sensor that the loop gain had to be low and just proportional. They used two temp sensors, one inside and one closer to ambient, where some feed-forward compensation helped.

But is probably blindingly obvious to those skilled in the art, who won't include the lawyers who tell you what you can publish.

True. My mistake.

Not a big deal. The Steinhart-Hart equation is messy, but good to about a thousandth of a a degree Kelvin. Practical applications tend to rely on linear interpolation between closely spaced points on a look-up table.

Old-fashioned glass encapsulated thermistors tend to have about a second of thermal lag in the package. That looks very like a pure delay. Thin-film surface mount thermistors may do better, but you do have to keep the local environment benign.

Not having worked for anybody for whom oscillator performance was the headline selling parameter, I've had to put up with less complaisant people on the shop floor.

Yes, it is, and well used in other types of products.

Interestingly, some of the most useful information is stuff that is available, can't possible be a trade secret, but is not generally known, so you try and keep it too yourself anyway.

For example... RTS/Popcorn noise on bipolar transistor on BiCMOS processes. We got burnt on the IBM7WL process a few years back. Totally unusable because of it. Nothing in the documentation said it had a problem. Cost us a lot of time, money and effort, had to scrap the asic.

We searched and found Dongbu HiTek. They claimed a wonderful high performance BiCMOS (isolated npn/pnp, great SiCr resistors)process. We built a chip with an amp to test for RTS. Again, completely worthless bipolars.

We finally found another vendor with a BiCMOS, made a test chip again, and finally something that worked.

You can't trust the fab vendors. They just wont tell you about RTS noise. The vast majority of BiCmos process have RTS, making their bipolars useless for precision designs. So, why save a competitor a year in time finding a process that actually works.

For general TCXOs/OCXOs (non mems) its all calibrated out using continuous voltage/time analog function generators. You don't want the frequency steps.

Key problem is frequency voltage control. The transfer function has to be precisely linear. Otherwise the compensation at one control voltage is not the same as another, meaning the compensation goes off. One is trying to compensate say, 20 ppm down to 50 ppb. This is 400:1. 1% non-linearity is too much.

The mass increase in coms over the last 20 years has lead to huge demand for really tight phase noise and temperature stability. Only a dedicated company has the infostructure do it. Every single device is individually calibrated. This requires ovens/coolers and ramp up and downs several times, for all millions of devices sold.

The oscillator is the heart of any comm/navigation system. Nothing works without it.

Typically the asics I design have around 10,000 analog transistors. A one transistor Colpitts on its own don't cut it. :-)

Actually, I finally came up with an oscillator topology that uses several transistors that has around 12dB better close in (1Hz) phase noise than anything else out there. Subtle, but obvious after the fact. Around 4 separate features not used on standard designs all combine up to get the improvement.

-- Kevin Aylward

Getting sinusoids out of essentially exponential transistors is a neat tric k.

Barry Gilbert had few examples of that kind of thing.

Analog Devices had a very nice bipolar process that seemed to be able to gi ve you really good fast npn and pnp bipolar transistors - have you talked t o them at all?

I was a bit surprised when their AD797 blew the LT1028 away - it wasn't any quieter, but the output was a lot better behaved. I'd recommended it as a potential alternative to the LT1028 to somebody who was having the usual tr oubles with the LT1028, and was surprised by the difference it seemed to ma ke.