Just an update. I’ve built the second prototype board (I skipped over the first design), and it’s powering my tbolt right now.

The design calls for 15v in (though it would also work with 13.8v). The +12 output comes from a D2PAK 7812. For +5, there is an AP1509 buck converter to make around 6.5 volts, then a DPAK 7805. For -12, there is an MC34063 configured as an inverter to make around -13.75 volts and then a DPAK 7912.

Steady-state, the system appears to be working just fine. The AP1509’s inductor and the D2PAK 7812 are just warm to the touch.

I checked for noise and ripple on the outputs and it’s somewhere around ±2 mV or so generally. From what I can see on the scope, there’s no ripple - it’s all high frequency noise. I am not absolutely certain that the noise measurement represents real noise or the limits of my measuring ability. I’m just using the scope probes the scope came with, and 2 mV/div is its lowest range.

I haven’t compared the noise with the ex laptop supply that I was using before, but I’d have to believe it’s cleaner. I don’t really have a way to check the oscillator’s before and after ADEV. My only other reference is an FE5680A, and I think the thunderbolt’s going to be far better at lower tau (where this all matters).

I know also that ±2 mV is still one and perhaps two orders of magnitude higher than some have called for. But before I attempt to reduce the noise further, I’d like to know that there are real gains to be had. Would someone with a Thunderbolt and better output noise measuring wherewithal be willing to take a prototype and compare it with something that does have µV levels of noise and ripple so I can get an idea of what there is to gain? If you like, you can make such comparisons public - no secrets here.

I'm sure I have some 7805s lying around, maybe a 7812/7912. I'm interested
to see the 1/f noise of a classic regulator, what load current do you
expect? I can bias a 7805 for the same load and measure the 0.1 to 10 Hz
noise.

Also if you have a digital scope without a very good builtin FFT, octave
would be one solution.

On Tue, Oct 18, 2016 at 10:46 AM, Nick Sayer via time-nuts <
time-nuts@febo.com> wrote:

Nick had mention that the -12V rail on the thunderbolt has the poorest PSRR
with respect to frequency output, so I first took a look at the venerable
7912.

The first data-set was taken with a -13.5 VDC input. Attached is the 0.1 Hz
to 10 Hz noise of an essentially quiescently loaded 7912, only a 10k
resistor was added as load for preliminary evaluation. With a 60 dB preamp
the scale of the scope plot is 20 uV/div. The 0.1Hz to 10Hz band noise is
15 uVrms, which is about 1.3 ppm rms of the DC mean.

In allan deviation terms, a quiescently loaded 7912 has a spot noise of 7
uV/rtHz at 1 Hz (on the 1/f slope), normalized that's 580 ppb/rtHz.
Equivalently speaking, the flicker noise floor of an allan deviation plot
would be sqrt(2*ln(2)) that figure to be 6.8E-7.

Assuming a thunderbolt should be achieving 1/f floor of around 1E-12, it
would need a PSRR of at least 1 ppm/V. I'm sure someone has gone to the
trouble of actually measuring it.

So from a 0.1 Hz to 10 Hz noise standpoint, the 7912 isn't terrible
with 1.3 ppm rms noise, considering an LM399 is about 0.1 ppm rms, only one
order of magnitude off.

The bad side of a 7912 is in long-term stability and tempCo, the sample I
tested had at least a 150 ppm/degC tempCo, which is going to put a serious
lump/bump in the 10s tau to gps crossover point on an allan deviation plot.

On Tue, Oct 18, 2016 at 3:05 PM, Scott Stobbe scott.j.stobbe@gmail.com
wrote:

On Fri, 21 Oct 2016 00:20:43 -0400
Scott Stobbe scott.j.stobbe@gmail.com wrote:

If the Thunderbolt ist most sensitive to the -12V input, why not use
something like the LT3090? Its temperature coefficient is quite low
in the order of a few ppm/°C around room temperature. Using a metal
film resistor that should keep the output variations low as well.
As added bonus, you get a very low output noise.

And while you are at it, use three LT3090 for the positive supplies :-)

		Attila Kinali

--
Malek's Law:
Any simple idea will be worded in the most complicated way.

On Fri, 21 Oct 2016 10:59:59 +0200
Attila Kinali attila@kinali.ch wrote:

Ermm... LT3045

		Attila Kinali

--
Malek's Law:
Any simple idea will be worded in the most complicated way.

Hi

The EFC on the OCXO is less sensitive than 1 ppm / V. It is a slam dunk
to say that the whole OCXO is less sensitive than the EFC by a good margin ….

Bob

Well, because it's easily an order of magnitude more expensive than a 7912. $5 instead of 50¢ (Q:1).

If it matters, then fine, but I am sensitive to cost efficiency in addition to efficacy.

If you put the board in a box in a stable temperature environment (which I'd kind of assume you'd do if you cared about temperature stability generally), then how far do you really have to go?

Sent from my iPhone

Hi

At least in terms of voltage regulation (as opposed to noise), the -12V input on the TBolt is
the least sensitive input on the TBolt. It’s issue is only in terms of PSRR. The internals of
the unit take care of any drift or really low frequency stuff on the -12 input.

Bob

On Fri, 21 Oct 2016 06:14:03 -0700
Nick Sayer via time-nuts time-nuts@febo.com wrote:

Sorry, I've been hunting ps level delay uncertainties in the last weeks
and kind of forgot that not everyone does things to that extreme.

If you want to go for cheap LDOs, I would recommend going for the
modern variants available from Ti and Linear. Those have much less
noise and drift than the 79xx and 78xx families and are also zero
load impedance stable (ie can be used with ceramic capacitors).
But these are already all in the 1-3USD range if you want to supply 800mA.
Ie they are at least a factor 2 more expensive than an 78xx.

Let's put a few numbers here:
The stability of a cheap OCXO (SCOCXOHS by Microcrystal) is specced
as <510^-7/V. An 7812 (LM7812 by Fairchild) is specced to 0.8mV/°C @ 5mA.
Assuming that using 500mA output makes the Voltage drifft worse by a factor
of 10 (i hope this conservative enough) we are at 8mV/°C. Ie. we get a
total, temperature to supply voltage induced frequency dependence of 410^-9/°C.
This is quite noticable. Compare this to the 75ppb over 0°C to 60° of
the temperatur spec of the SXOCXOHS. I.e. we are suddenly adding as much
temperature dependence to the OCXO trough the power supply as it has itself.
Even if you discard the factor 10 for the 7812, that's still 4*10^-10/°C.

Using a more expensive OCXO (AOCJY4 by Abracon) gets you to 310^-11/°C
to 310^-12/°C, still quite noticable for a GPSDO and again in the order
of its intrinsic temperature coefficient.

And these calculations are only for the temperature of the LDO,
ie they assume an otherwise stable environment. If you start accounting
for drafts of air and changes in humidity, then things become even worse.

Also keep in mind, that unless you start putting cardboard boxes around
everything, then changes of 2-3°C within 10s is pretty normal when you
have people opening/closing doors/windows.

So.. does all this matter? Maybe, maybe not. :-) It depends on what you
actually do with the LDOs. If you put them in a metal box, maybe
with an cardboard box around it, then you dampen the temperature variations
quite a bit and give the GPSDO and its control loop  a chance to compensate.
If you leave them in the open, possibly with an heatsink attached, then
you definitely want to have a lower temperature coefficient.

		Attila Kinali

--
Malek's Law:
Any simple idea will be worded in the most complicated way.

On Fri, 21 Oct 2016 10:59:59 +0200, you wrote:

Or if space is not an issue, use a discrete reference or zener,
operational amplifier, and pass transistor for better performance yet
at less cost.  If all of the supplies are to be regulated, then use a
common reference to further save cost.

Separating the reference, error amplifier, and feedback network from
the power pass transistor lowers the effects from thermal feedback.