xor mixer

[John Larkin]

I need a really low noise mixer, better than I can do with a diode
ring mixer. The signals are ECL square waves, around 150 MHz, so I
figure the best mixer starts with a really fast XOR gate. The result
is ECL, but I'm leery of the analog quality (noise and DC drift) of
the raw gate output, so I was thinking that a differential pair could
add some analog precision.

https://dl.dropboxusercontent.com/u/53724080/Circuits/PLLs/XOR_Mixer.JPG

The inputs are usually 90 degrees apart, so the output should average
around zero differential. I'm interested in the low frequency
component.

The three transistors should ideally be inside an IC, to be really
fast and balanced. This reminds me of an ancient MC15xx sort of part,
but they were slow. Does anybody make things like that in a modern,
fast process? All that I seem to google is not quite right.

I guess I could try making it from discretes if I can't find an IC.

A really fast 6-transistor Gilbert cell could do the whole job, if
somebody made one.



--

John Larkin Highland Technology, Inc

lunatic fringe electronics

[Tauno Voipio]

Building a phase comparator?

How about borrowing technology from radios:

Mix both signals to a lower frequency and use fast CMOS switches
to do the final mixing. Google for 'Tayloe mixer'.

The first mixing does not ned to be DC accurate.

--

-TV

[John Larkin]

That's interesting. I want really low low-frequency phase noise, which
the dual front-end mixers might trash. That's harder to think about.

I did find the HFA3101, a pure, fast Gilbert array.

[Winfield Hill]

John Larkin wrote...

>
> A really fast 6-transistor Gilbert cell could
> do the whole job, if somebody made one.

Intersil's HFA3101, with 10GHz transistors.


--
Thanks,
- Win

[Jim Thompson]

From a project, early 2015, we evaluated...

AD831
AD834
ADL5391
LT5560

...Jim Thompson
--
| James E.Thompson | mens |
| Analog Innovations | et |
| Analog/Mixed-Signal ASIC's and Discrete Systems | manus |
| San Tan Valley, AZ 85142 Skype: Contacts Only | |
| Voice:(480)460-2350 Fax: Available upon request | Brass Rat |
| E-mail Icon at http://www.analog-innovations.com | 1962 |

I'm looking for work... see my website.

[John Larkin]

One thing that I think I want to avoid is AM sensitivity, which is why
I may not want a linear mixer. The A and B signals will be fast
differential PECL, so we'd get AM rejection by using an XOR gate or an
over-driven Gilbert multiplier.

A Gilbert will have six transistors worth of base current noise,
whereas a diff pair (after an XOR gate) would have three or maybe two.

IC mixers generally don't show their internal schematic, so it's hard
to tell how they will actually behave.

Diode ring mixers are handicapped by their low output voltage and
internal impedance Johnson noise, so have noise floors around -150
dBc/Hz.

[John Larkin]

On Mon, 30 May 2016 09:34:19 -0700, John Larkin
<jjla...@highlandtechnology.com> wrote:

>

There are literally hundreds of times as many academic papers about
Gilbert cell multipliers, than there are data sheets of parts you can
actually buy.

The Gilbert topology seems to be a favorite for lectures and thesis
papers.

PMB2335 is a little weird but it does use 25 GHz transistors.

[Phil Hobbs]

Re: diode rings

Check out the Mini Circuits MPD series phase detectors. You get about 1V/radian and a nanovolt or two 1-Hz noise, so you're down in 170 dBc territory.

Cheers

Phil Hobbs

[John Larkin]

On 30 May 2016 10:01:59 -0700, Winfield Hill

<hi...@rowland.harvard.edu> wrote:

>John Larkin wrote...
>>
>> A really fast 6-transistor Gilbert cell could
>> do the whole job, if somebody made one.
>
> Intersil's HFA3101, with 10GHz transistors.

That's nice. It's almost worth shorting out all the upper transistors
and using the lower diff pair after my XOR gate.

[John Larkin]

That's impressive. Using two detectors and some correlation math is
the next step.

[whit3rd]

On Monday, May 30, 2016 at 9:34:23 AM UTC-7, John Larkin wrote:
> I need a really low noise mixer,... The signals are ECL square waves, around 150 MHz, so I

> figure the best mixer starts with a really fast XOR gate. The result
> is ECL, but I'm leery of the analog quality (noise and DC drift) of
> the raw gate output, so I was thinking that a differential pair could
> add some analog precision.

All 100k ECL uses differential-pair inputs; could you use 100EP40 ?

<http://www.onsemi.com/pub_link/Collateral/MC100EP40-D.PDF>

>I'm interested in the low frequency component.

I'd prefer a Gilbert-cell type for that, and it's gonna benefit from a
thermostat enclosure. Gate mixers get AM from power and thermal
sources, Gilbert cells are (theoretically) balanced but still have
gain/temperature modulation.

> The three transistors should ideally be inside an IC, to be really
> fast and balanced.

IC transistors aren't especially fast, I'd think that hand-matching
of discretes can accomplish 'balance' better than sorting through
arrays at IC test time. In any case, if balance fails, it won't hurt
the DC output, mainly just add some 'around 150 MHz'.

[John Larkin]

On Mon, 30 May 2016 12:41:20 -0700 (PDT), whit3rd <whi...@gmail.com>
wrote:

I was considering using a 10EP08 XOR gate, or, more extreme, an
NBSG86A. That last one has 40 ps edges.

IC diff pairs or Gilberts are available with 10 GHz transistors. Base
current shot noise will be a problem with a diff pair or a Gilbert
cell.

Maybe I could use the XOR gate ECL output directly, but that's scary.
Diode current steering maybe?

This is the sort of thing people spend a lifetime specializing in. I
have 5 weeks maybe.

[whit3rd]

On Monday, May 30, 2016 at 2:21:15 PM UTC-7, John Larkin wrote:

> I was considering using a 10EP08 XOR gate, or, more extreme, an
> NBSG86A. That last one has 40 ps edges.
>
> IC diff pairs or Gilberts are available with 10 GHz transistors. Base
> current shot noise will be a problem with a diff pair or a Gilbert
> cell.

Why will shot noise (broadband) be a problem for the narrow
bandpass your 'DC' signal occupies?

[John Larkin]

On Mon, 30 May 2016 14:43:49 -0700 (PDT), whit3rd <whi...@gmail.com>
wrote:

Noise is noise. If I have a transistor steering, say, 10 mA, its base
current might be 100 uA. The shot noise of the base current will be 6
pA/rthz. That's pretty small, about 185 dB down in 1 Hz, but it's not
zero. More transistors, like in a Gilbert cell, would be worse.

I wonder if finite Ft is equivalent to lower beta from a shot noise
standpoint. Probably.

[bill....@ieee.org]

http://www.analog.com/media/en/technical-documentation/data-sheets/HMC3716.pdf

Analog Devices seems to have quite a few specialised chips for this kind of work. The unspecialised AD835 is still around, but the ECL high and low logic levels would show up in the output from that.

--
Bill Sloman, Sydney

[Tim Williams]

"John Larkin" <jjla...@highlandtechnology.com> wrote in message
news:rsfpkb9qoc73q9hfi...@4ax.com...

> I wonder if finite Ft is equivalent to lower beta from a shot noise
> standpoint. Probably.

The definition of fT is, beta(freq) = 1. The dominant pole (at f3 = fT /
beta(DC)) would be the traditional rolloff point (or roll-up, if your
circuit is pushing harder at higher frequencies, like in a negative feedback
loop), which would give you your necessarily finite bandwidth.

If you're worried about current noise, why not use FETs? ;-)

Tim

--
Seven Transistor Labs, LLC
Electrical Engineering Consultation and Contract Design
Website: http://seventransistorlabs.com

[whit3rd]

On Monday, May 30, 2016 at 3:43:12 PM UTC-7, John Larkin wrote:
> On Mon, 30 May 2016 14:43:49 -0700 (PDT), whit3rd <whi...@gmail.com>
> wrote:
>
> >On Monday, May 30, 2016 at 2:21:15 PM UTC-7, John Larkin wrote:

> >> IC diff pairs or Gilberts are available with 10 GHz transistors. Base
> >> current shot noise will be a problem with a diff pair or a Gilbert
> >> cell.

> >Why will shot noise (broadband) be a problem for the narrow
> >bandpass your 'DC' signal occupies?

> Noise is noise. If I have a transistor steering, say, 10 mA, its base
> current might be 100 uA. The shot noise of the base current will be 6
> pA/rthz. That's pretty small, about 185 dB down in 1 Hz, but it's not
> zero. More transistors, like in a Gilbert cell, would be worse.

Six transistor cell will bump it up by a factor of 1.73 (square root of three);
not a killer, starting from -185 dB. I'm thinking your power supply
ripple is more important.

[Tim Wescott]

I always thought the noise figure of a diode ring mixer came about purely
because of the conversion loss and the thermal noise in the various
source resistances.

If you're driving it hard from low-noise sources is it really that
noisy? Where does the noise come from?

Waiting to be eddicated...

--
Tim Wescott
Control systems, embedded software and circuit design
I'm looking for work! See my website if you're interested
http://www.wescottdesign.com

[Chris Jones]

It's perhaps a pity you have ECL inputs. At least on-chip, for frequency
dividers, fast CMOS tends to provide better phase noise than bipolar,
provided your power rail is very quiet. I don't know if this is true
also of off-the-shelf logic chips though. If you get the opportunity to
compare fast CMOS against the ECL in an experiment, that would be
interesting.

Chris

[Phil Hobbs]

On 05/30/2016 10:48 PM, Tim Wescott wrote:
> On Mon, 30 May 2016 09:34:19 -0700, John Larkin wrote:
>
>> I need a really low noise mixer, better than I can do with a diode ring
>> mixer. The signals are ECL square waves, around 150 MHz, so I figure the
>> best mixer starts with a really fast XOR gate. The result is ECL, but
>> I'm leery of the analog quality (noise and DC drift) of the raw gate
>> output, so I was thinking that a differential pair could add some analog
>> precision.
>>
>> https://dl.dropboxusercontent.com/u/53724080/Circuits/PLLs/XOR_Mixer.JPG
>>
>> The inputs are usually 90 degrees apart, so the output should average
>> around zero differential. I'm interested in the low frequency component.
>>
>> The three transistors should ideally be inside an IC, to be really fast
>> and balanced. This reminds me of an ancient MC15xx sort of part, but
>> they were slow. Does anybody make things like that in a modern, fast
>> process? All that I seem to google is not quite right.
>>
>> I guess I could try making it from discretes if I can't find an IC.
>>
>> A really fast 6-transistor Gilbert cell could do the whole job, if
>> somebody made one.
>
> I always thought the noise figure of a diode ring mixer came about purely
> because of the conversion loss and the thermal noise in the various
> source resistances.
>
> If you're driving it hard from low-noise sources is it really that
> noisy? Where does the noise come from?
>
> Waiting to be eddicated...

The video resistance of a Schottky diode depends on the conduction angle
among other things. Conversion loss is the sum of reflection and
dissipation. Reflection doesn't introduce noise, but dissipation does.

Cheers

Phil Hobbs


--
Dr Philip C D Hobbs
Principal Consultant
ElectroOptical Innovations LLC
Optics, Electro-optics, Photonics, Analog Electronics

160 North State Road #203
Briarcliff Manor NY 10510

hobbs at electrooptical dot net
http://electrooptical.net

[Robert Baer]

The major problem is exactly how the XOR function is implemented.
I have seen designs that you could improve blindfolded.
...and if done "decently", then what about issues regarding
topological symmetry and balance?

Sounds like you want the equivalent of a low input VOS, low noise,
low drift op-amp.
So..if there is such a beast that has the required speed, the
solution would be that one part.

[bill....@ieee.org]

The great virtue of ECL is that it is current steering logic, so the ECL rails are way quieter than TTL/CMOS power rails. If you get enthusiastic and use balanced ECL inputs and outputs, the rails are even quieter.

For mixed signal design, ECL is very handy, even when you don't really need the speed, and it struck me - back when I was more closely involved - that the preponderance of ECL-interfaces on fast A/D adn D/A converters was driven by that.

--
Bill Sloman, Sydney

[Chris Jones]

I don't dispute that. I was just pointing out that the phase noise or
jitter performance of ECL is not that great compared to some CMOS. I
think part of it is that the slew rate of the CMOS signals can be quite
high compared to ECL. High slew rate is good, because it gives you less
picoseconds of jitter per millivolt of input-referred additive noise
exhibited by the input of the following logic stage.

I'm not sure how good off-the-shelf CMOS logic gates are - I guess it is
hard to buy anything faster/smaller than 0.5um gate length except in a
FPGA. If I remember correctly, an old Philips 74LVC04AD gave about 250ps
rise and fall times on a 5V supply, so that would be 16V/ns slew rate,
so I guess you could expect about 65fs of jitter per millivolt of noise
in the following stage within its bandwidth. It would be interesting to
measure how much jitter each stage of those gates adds, but I don't
really have the equipment to do it easily, or the time to do it not-easily.

Chris

[Winfield Hill]

John Larkin wrote...

> Winfield Hill wrote:
>> John Larkin wrote...
>>>
>>> A really fast 6-transistor Gilbert cell could
>>> do the whole job, if somebody made one.
>>
>> Intersil's HFA3101, with 10GHz transistors.
>
> That's nice. It's almost worth shorting out all
> the upper transistors and using the lower diff
> pair after my XOR gate.

The '3102 has two 10 GHz diff pairs, the '3048
has an 8 GHz pair, and the '3127 has matched
8 GHz transistors. The capacitances are also
very low, well under 1pF. But watch out for
the low betas and Early voltages. Have fun.


--
Thanks,
- Win

[Clifford Heath]

On 31/05/16 21:03, Winfield Hill wrote:
> John Larkin wrote...

>> Winfield Hill wrote:
>>> Intersil's HFA3101, with 10GHz transistors.
>> That's nice. It's almost worth shorting out all
>> the upper transistors and using the lower diff
>> pair after my XOR gate.

> The '3102 has.... But watch out for

> the low betas and Early voltages.

Ouch. I thought a low Early voltage was a significant
source of intermodulation (nonlinearity), and a bad
thing in a mixer. Why isn't that a problem in the
HFA3101?

[Gerhard Hoffmann]

The large emitter inductance of those transistor
arrays poisons the differential anplifiers because
you cannot get COMMON emitters at those speeds.
That is even worse for the space qualified flatpacks
that I had to use.

And I found it impossible to build a current
mirror that was better than a simple one with the
PNPs of the HFA-chips. By the time it was stabilized
with resistors most of the speed advantage was gone.
And the base stoppers count for Rbb, noise-wise.
I took them anyway; there were no better pnps
available.

If you can use discrete transistors, the Infineon
BFQ790 could be interesting. It has no bond wires
in the emitter; the emitter is directly soldered to
the center pin of a sot-89. One could place two of
them next to each other and have a wide highway
for the emitter currents.
The BFQ790 features a funny Early behaviour.
Any idea why?

And, since 1/f noise is proportional to current
density, the large chip may help. The corner is
probably high enough b/c of ft. :-(

When using PECL, remember that a high level carries
the full VCC noise without attenuation. With NECL
it looks somewhat better.

For low noise mixers, the diode ring is probably
still the best.

There is a publication from nist:
"Residual PM Noise Evaluation of Radio Frequency Mixers"
by C. A. Barnes, A. Hati, C. W. Nelson and D. A. Howe
filename is probably 2556.pdf

They come up with a ring mixer made from diode-
connected 2N2222 that seems to be interesting.
JL had recently an idea with diode-connected
phemts, I wonder how they would perform.
Well, above 1/f.


Is Fred Bertoli (sp?) still here? I'm currently playing
with a flock of Interfet IF3601 (really somewhat
presorted as 3602) IIRC he has done something
similar.

regards, Gerhard

[Phil Hobbs]

You can cascode them if it's a problem. To leading order in beta, it's
only the actual diff pair that matters. A good opportunity to try my
nice SiGe:C cascode trick.

If you run a BFP640 without its emitter grounded really well, you have
to put a bead in the base or else it'll oscillate at 12-15 GHz. I
usually use a Murata BLM18BB05 or BLM18BB10, which work fine if you
don't need the transistor's full speed.

On the other hand, its Early voltage is effectively infinite--the DC
collector curves actually show a negative resistance due to heating, but
at AC they're almost perfectly flat.

[John Larkin]

CMOS logic is generally single-ended, so power supply noise appears
directly on an output. And CMOS has a ghastly delay TC, many ps per
degree C, so minute temperature fluctuations get mapped into phase
noise. Agree that on-chip logic is always faster than off-chip.

Analog Devices makes the best, fastest comparators around, and they
are ECL/PECL.

We have often lamented that there are no LVDS logic gates or flops.

[John Larkin]

It's hard or maybe impossible to directly measure fs jitter. My
customer needs low jitter but they characterize performance in terms
of the single-sideband phase noise spectrum, which is hard to measure
(hence this thread!) but not impossible.

Correlation techniques are used to lower the noise floor of phase
noise measurements. I suppose the same thing could be done in time
domain, except that mixers are a lot cheaper than oscilloscopes.

[John Larkin]

NBSG86 is about the fastest XOR gate around, but it's a universal
any-function gate so probably has more paths than it strictly needs.

One nice thing about differential ECL is that gate inputs don't load
the signal much. So I could have two compete signal paths - XOR, diff
pair maybe, filter, amp, ADC - hung on the same ECL signal pair. Then
some correlation math on the ADC data can, over time, wash out the
noise of the individual signal paths.

I guess the same trick can be done with Gilbert cells.

[Gerhard Hoffmann]

> On 31/05/2016 02:34, John Larkin wrote:
>>
>> I need a really low noise mixer, better than I can do with a diode
>> ring mixer. The signals are ECL square waves, around 150 MHz, so I
>> figure the best mixer starts with a really fast XOR gate. The result
>> is ECL, but I'm leery of the analog quality (noise and DC drift) of
>> the raw gate output, so I was thinking that a differential pair could
>> add some analog precision.
>>
>> https://dl.dropboxusercontent.com/u/53724080/Circuits/PLLs/XOR_Mixer.JPG
>>
>> The inputs are usually 90 degrees apart, so the output should average
>> around zero differential. I'm interested in the low frequency
>> component.
>>
>> The three transistors should ideally be inside an IC, to be really
>> fast and balanced. This reminds me of an ancient MC15xx sort of part,
>> but they were slow. Does anybody make things like that in a modern,
>> fast process? All that I seem to google is not quite right.
>>
>> I guess I could try making it from discretes if I can't find an IC.
>>

Isn't a SY58051U CML gate exactly what you want?
There are faster versions also for 10 GHz.
Or the gate that you have and a CML driver?

regards, Gerhard

[Tim Wescott]

Got a white paper that happens to treat this in the context of finding
phase errors?

It still seems that with a good square wave on the LO input and a high-
level RF input, the noise could be made almost arbitrarily low (well,
assuming that you aren't burning up the diodes). I want to see where I'm
wrong...

--

Tim Wescott
Wescott Design Services
http://www.wescottdesign.com

I'm looking for work -- see my website!

[John Larkin]

On Mon, 30 May 2016 18:45:22 -0500, "Tim Williams"
<tiw...@seventransistorlabs.com> wrote:

>"John Larkin" <jjla...@highlandtechnology.com> wrote in message
>news:rsfpkb9qoc73q9hfi...@4ax.com...
>> I wonder if finite Ft is equivalent to lower beta from a shot noise
>> standpoint. Probably.
>
>The definition of fT is, beta(freq) = 1. The dominant pole (at f3 = fT /
>beta(DC)) would be the traditional rolloff point (or roll-up, if your
>circuit is pushing harder at higher frequencies, like in a negative feedback
>loop), which would give you your necessarily finite bandwidth.
>
>If you're worried about current noise, why not use FETs? ;-)
>
>Tim

A pair of phemts would work with ECL swings, but I don't know if the
phase noise would be any better. Phemts have horrible drain slopes
(Early voltage) and horrible 1/f corner frequencies.

--

John Larkin Highland Technology, Inc

picosecond timing precision measurement

jlarkin att highlandtechnology dott com
http://www.highlandtechnology.com

[whit3rd]

On Tuesday, May 31, 2016 at 12:07:05 PM UTC-7, Tim Wescott wrote:
> On Tue, 31 May 2016 00:18:47 -0400, Phil Hobbs wrote:

[about mixer noise]

> > The video resistance of a Schottky diode depends on the conduction angle

> > among other things...

> Got a white paper that happens to treat this in the context of finding
> phase errors?
>
> It still seems that with a good square wave on the LO input and a high-
> level RF input, the noise could be made almost arbitrarily low (well,
> assuming that you aren't burning up the diodes). I want to see where I'm
> wrong...

The noise problem with diode mixers (and XOR has similar issues) is that it
saturates; the signal only activates the mixing mechanism at edges. So,
the duty cycle of that edge transition becomes the ONLY sensitive part of the cycle,
and that implies that you have sensitivity at all of the (odd) high frequency
harmonics of the intended signal, because they are now aliased.

Even if the device doesn't, on paper, have higher-frequency gain, it DOESN'T HELP,
because noise needn't come in through an input pin; your SPICE model
fits the noise injected from a wire, but not generally the noise present
in the device.

[John Larkin]

On Tue, 31 May 2016 12:32:35 -0700 (PDT), whit3rd <whi...@gmail.com>
wrote:

>On Tuesday, May 31, 2016 at 12:07:05 PM UTC-7, Tim Wescott wrote:
>> On Tue, 31 May 2016 00:18:47 -0400, Phil Hobbs wrote:
>
>[about mixer noise]
>
>> > The video resistance of a Schottky diode depends on the conduction angle
>> > among other things...
>
>> Got a white paper that happens to treat this in the context of finding
>> phase errors?
>>
>> It still seems that with a good square wave on the LO input and a high-
>> level RF input, the noise could be made almost arbitrarily low (well,
>> assuming that you aren't burning up the diodes). I want to see where I'm
>> wrong...
>
>The noise problem with diode mixers (and XOR has similar issues) is that it
>saturates; the signal only activates the mixing mechanism at edges. So,
>the duty cycle of that edge transition becomes the ONLY sensitive part of the cycle,
>and that implies that you have sensitivity at all of the (odd) high frequency
>harmonics of the intended signal, because they are now aliased.

Saturation is good, because it makes the detected output independent
of the signal amplitudes. The Gilbert cell, driven by ECL, will do
that. So will an XOR feeding some saturating current steering thing.
The diode mixer is somewhat sensitive to the input amplitudes,
depending on how hard you drive it.

Sure, if the mixer saturates, all the information is in the edge
timing. But that's good.

A ECL XOR feeding some current-steering schottky diodes is
interesting. Diode reverse leakage current matters, but shouldn't be
too bad, way less than the base current of an NPN.

[whit3rd]

On Tuesday, May 31, 2016 at 1:18:48 PM UTC-7, John Larkin wrote:
> On Tue, 31 May 2016 12:32:35 -0700 (PDT), whit3rd <whi...@gmail.com>
> wrote:
>
> >On Tuesday, May 31, 2016 at 12:07:05 PM UTC-7, Tim Wescott wrote:
> >> On Tue, 31 May 2016 00:18:47 -0400, Phil Hobbs wrote:
> >
> >[about mixer noise]
> >

> >> > The video resistance of a Schottky diode depends on the conduction angle...

> >> It still seems that with a good square wave on the LO input and a high-

> >> level RF input...

> >The noise problem with diode mixers (and XOR has similar issues) is that it
> >saturates; the signal only activates the mixing mechanism at edges. So,
> >the duty cycle of that edge transition becomes the ONLY sensitive part of the cycle,
> >and that implies that you have sensitivity at all of the (odd) high frequency
> >harmonics of the intended signal, because they are now aliased.

> Saturation is good, because it makes the detected output independent
> of the signal amplitudes.

That's true of the LO amplitude, but not of RF input, or device internal noise sources.
Getting rid of one problem, but leaving others, and introducing new ones...

> Sure, if the mixer saturates, all the information is in the edge
> timing. But that's good.

Why? A continuous-multiplier mixer doesn't sense noise from inband
harmonics of one input signal, but a switching/XOR mixer does. It's
aliasing that noise source so you cannot later filter it away.

[John Larkin]

On Tue, 31 May 2016 14:00:23 -0700 (PDT), whit3rd <whi...@gmail.com>

I have two 150 MHz square waves and I want to evaluate the phase noise
between them. Neither is explicitly RF or LO.

Later on, I may do a PLL using a very good OCXO, but the signals will
still be square waves and the phase detector noise is still critical.

The XOR phase detector output DC level vs phase difference is a simple
triangle. All that frequency-domain aliased harmonics thinking washes
out. Sometimes what's complex in one domain is simple in the other.

[John Larkin]

That's similar to the NBSG86 with, unfortunately, a different pinout.

The '86 is maybe a bit faster, and has more (true ECL) output swing.

The Johnson noise on the 50 ohm, 400 mV CML output will be
significant. But so will the Johnson noise from the resistor at the
bases of the '86 ECL output transistors. I guess the trick is to keep
the edges really fast and drive some very fast threshold/saturating
device after the XOR gate.

This is really hard.

[Clifford Heath]

On 01/06/16 00:48, Phil Hobbs wrote:
> On 05/31/2016 06:27 AM, Clifford Heath wrote:
>> On 31/05/16 21:03, Winfield Hill wrote:
>>> John Larkin wrote...
>>>> Winfield Hill wrote:
>>>>> Intersil's HFA3101, with 10GHz transistors.
>>>> That's nice. It's almost worth shorting out all
>>>> the upper transistors and using the lower diff
>>>> pair after my XOR gate.
>>> The '3102 has.... But watch out for
>>> the low betas and Early voltages.
>> Ouch. I thought a low Early voltage was a significant
>> source of intermodulation (nonlinearity), and a bad
>> thing in a mixer. Why isn't that a problem in the
>> HFA3101?
> You can cascode them if it's a problem. To leading order in beta, it's
> only the actual diff pair that matters.

Would that work with the HFA3101 though? The collectors
crossing between the two pairs are already tied together.
Actually... perhaps it would work! That would be a cunning
trick. All the nice speed and matching of the 3101, with
much more gain and linearity, just for the price of another
couple of transistors. It would be very good on the input
of an all-band receiver.

> A good opportunity to try my nice SiGe:C cascode trick.

I'm keen to try that sometime. The dual-gate GaAsFET that
VK3YNG used in his foxhunt sniffer is now unobtainium, and
I think that this approach (with either a dual-gate FET
or a parallel pair of FETs) could give the same linearity
and extreme dynamic range that he obtained (almost 140dB
of AGC, +-70dB) in a single stage. This is necessary for
foxhunting if you want to avoid switched attenuators,
because you want accurate RSSI sensitivity from <1uV to >1V
of antenna input, without swamping or IMD. If you're hunting
a 30W TX that's well hidden, there's nothing more annoying
that to arrive first, and discover that as soon as you're
inside the 15m radius, you can't get a direction any more
because your front-end is swamped - and the signal is no
longer *somewhere*, it's now *everywhere* - but you still
can't see the TX!

[Gerhard Hoffmann]

Am 31.05.2016 um 23:20 schrieb John Larkin:

>
> I have two 150 MHz square waves and I want to evaluate the phase noise
> between them. Neither is explicitly RF or LO.
>
> Later on, I may do a PLL using a very good OCXO, but the signals will
> still be square waves and the phase detector noise is still critical.
>
> The XOR phase detector output DC level vs phase difference is a simple
> triangle. All that frequency-domain aliased harmonics thinking washes
> out. Sometimes what's complex in one domain is simple in the other.

Why don't you do it like everybody else and mix the two 150 MHz
waves down to 1 Mhz or even less with 2 ring mixers and a common
transfer oscillator? This is called DMTD, Dual mixer time
difference system. Your problems shrink by 20 log(149) dB.

Same phase changes, but now on two 1 MHz carriers.

Most of the noise of the transfer oscillator is common mode
and vanishes. A little bit stays since you de-correlate your
zero-crossings by upto 1 usec for the zoom wrt the xfer osc.
So it still pays to be not too shabby with the oscillator.

I did that to the 100 MHz outputs of a cesium and a hydrogen
maser with ordinary MCL ring mixers. After limiting, then there
is a second stage in the following FPGA with D-FFs as mixers
and the conversion clock derived from one of the signals.

regards, Gerhard

[Gerhard Hoffmann]

Look at the SY58051U also. (Micrel, or whoever owns them today.)

The Johnson noise will drown in the noise of the tail current source.
The two transistors are only similar.
And you get the noise of the 2 emitter followers that you really
do not need.

> This is really hard.

Yes.


Also interesting: < http://arxiv.org/abs/physics/0608211 >
and everything on rubiola.org .

regards, Gerhard

[Phil Hobbs]

>Why don't you do it like everybody else and mix the two 150 MHz
>waves down to 1 Mhz or even less with 2 ring mixers and a common
>transfer oscillator? This is called DMTD, Dual mixer time
>difference system. Your problems shrink by 20 log(149) dB

If Johnson noise in the mixer is the limiting problem, that method won't help, ISTM. High dV/dphi and low noise is the simpler approach.

Cheers

Phil Hobbs

[George Herold]

Hah, send them into laser diodes, mix 'em with beam splitters
and send 'em into a fast photodiode.

(someone has been asking about phase locking LD's
so I've thinking about that...)
George H.

[Gerhard Hoffmann]

But it isn't. We have nearly Volts at the input of the mixer.
There is not much available that has less noise than transformers
and Schottky diodes. The noise of the Schottkies is even only
half-thermal. The real problem is limiting the bandwidth
against aliasing. Do not drive the first mixer too hard.

There is a whole industry built on dual mixer systems.

> Cheers

Gerhard

[John Larkin]

Laser diodes are noisy!

--

John Larkin Highland Technology, Inc

lunatic fringe electronics

[Tauno Voipio]

If you roll back to the start of this thread, you'll notice
that John turned down this idea at the very beginning.

--

-TV

[Gerhard Hoffmann]

Am 01.06.2016 um 09:06 schrieb Tauno Voipio:

>
> If you roll back to the start of this thread, you'll notice
> that John turned down this idea at the very beginning.
>

oh, sorry, that was gone already locally.

Well, this then is just like High End Audio where they deny
that feedback could ever work and then claim that audio is
the hardest thing in engineering.

I mean, if it it works easily for two atomic clocks, it
should be good enough for what is probably a noisy
155 Mb/s plastic fiber optic link.


regards, Gerhard

[frank]

Clifford Heath <no....@please.net> wrote:
> that to arrive first, and discover that as soon as you're
> inside the 15m radius, you can't get a direction any more
> because your front-end is swamped - and the signal is no
> longer *somewhere*, it's now *everywhere* - but you still
> can't see the TX!

on the few occasions that I tried foxhunting, I just switched to 3rd, then 5th
harmonic of the fundamental as soon as my poor receiver was overloading with
the fundamental signal. Every TX makes some signal at odd harmonics.
Of course wide RX frontend is a no-no with this approach.
I was the winner on the first attempt (then I got bored quickly).

Frank IZ8DWF (sorry for the OT)

[Tauno Voipio]

Right - maybe John wants thick golden speaker cables.

--

Gruesse, Tauno

[Phil Hobbs]

I don't think it's quite that simple. The Fourier series for a square
wave dies off as 1/f, and has only odd harmonics, so the noise gain vs.
the sinusoidal case isn't very large. Since you're LP filtering the
output, you only see noise in narrow passbands centred on the harmonics,
and the sensitivity goes as 1/N:

noise gain = sum (n=1,2,....) 1/(2n+1)**2

= sum (n=1,2,....) (1/n**2 - 1/(2n)**2)

= 0.75 * pi**2/6 = 0.9 dB.

That's well within the range of performance differences in a single
mixer type, and certainly within the range of different types.

In the pure weak-additive-noise case, the phase noise spectrum depends
only on the carrier-to-noise ratio:

delta phi = 1/sqrt(2*CNR) .

This is true regardless of whether you've mixed down to a low IF or not.
Also, the shape of the phase noise spectrum is identical to that of
the additive noise. So in this case the LO-plus-2-mixer trick is bound
to make things worse and not better, unless there's a serious difference
in the performance of available phase detectors at high vs low frequency.

If the dominant noise source is an irreducible amount of *time* jitter
at the input of the phase detector, then you win by mixing down, because
a picosecond is 100 ppm at 100 MHz but only 100 ppb at 100 kHz. However
I don't think that's the usual case, and it doesn't sound like John's in
that position.

[Chris Jones]

Continuous-multiplier mixers based on Gilbert-cell topologies that are
not hard-swiched tend to be very noisy. It is really not easy to make a
multiplier that is as quiet as a switching mixer. It might be easier to
imagine it with the LO at, say 1Hz. When the (differential) LO signal is
going through zero, each input current is equally divided between two
transistors, and even if there were no noise in the input current coming
from the gm stage, there would be noise in each current coming out of
the current-steering stages. That problem essentially doesn't occur with
a switching mixer, except when the (differential) LO signal is going
through zero. That is why for low noise (and good linearity) you want to
drive the LO port with a nice crisp square wave, not a sine wave. Many
people don't believe this but it is fairly straightforward to measure.
(These sorts of things are hard to simulate properly without SpectreRF
or quite a bit of careful thought.)

[John Larkin]

It's not 155.52 and it's kilometers of singlemode fiber, but the
problem is essentially the same. I want to characterize the link phase
noise spectrum down to -180 dBc if possible, and I need a really good
phase detector to do that. The DMTD idea will just add more mixer
phase noise to the problem.

I'm thinking two fast XOR gates, each driving a diode current-steering
network and amp/filter, then dual ADCs and some FFT/correlation math.
Averaging washes out the mixer noise.

DMTD does a lot of averaging, too.

--

John Larkin Highland Technology, Inc

lunatic fringe electronics

[John Larkin]

That aligns with my thinking: two ECL square waves with really fast
edges over-driving a Gilbert multiplier, or an XOR over-driving a
differential pair or some current-steering diodes. The transistors
have base-current shot noise, but that's pretty far down.

Transistor Ft has to be really high, 10 GHz or so, to keep the base
currents down. That's why diode current steering is interesting.

I want to graph phase noise vs frequency, out to 10s of KHz, so a low
IF, or any IF, doesn't work. A single mixer looks best.

--

John Larkin Highland Technology, Inc

lunatic fringe electronics

[John Larkin]

Sounds right. Adding the phase noise of multiple mixers can't help.

Multiple mixers help if they are used in parallel. I could literally
parallel four mixers (XOR or Gilberts have fairly high input
impedances, so string them along differential transmission lines) and
sum their outputs, and gain 6 dB s/n. But I may as well digitize all
four outputs. Then I can sum digitally, or do correlation tricks to
get way better than 6 dB. There are commercial phase-noise measurement
systems that do correlation of two mixer outputs, so it would only
take a really smart math guy to extend that to four mixers. I have
access to really smart math guys.

--

John Larkin Highland Technology, Inc

lunatic fringe electronics

[Gerhard Hoffmann]

Am 01.06.2016 um 16:37 schrieb John Larkin:

> It's not 155.52 and it's kilometers of singlemode fiber, but the
> problem is essentially the same. I want to characterize the link phase
> noise spectrum down to -180 dBc if possible, and I need a really good
> phase detector to do that. The DMTD idea will just add more mixer
> phase noise to the problem.

Can't you lend a E5052B or an R&S FSUP if it's only for
characterization? They also run out of steam at -180dBc,
but there has already gone a lot of verfication work into them.

In this league you must have everything twice including the reference
oscillators and include everything in the three cornered hat /
cross FFT.

There are surprises like correlated thermal noise that comes in
antiphase out of a power divider and that happily averages away,
but the true result is much worse.

> I'm thinking two fast XOR gates, each driving a diode current-steering
> network and amp/filter, then dual ADCs and some FFT/correlation math.
> Averaging washes out the mixer noise.

I'm also interested in building sth. simple for this purpose,
probably in Wenzel's down mixer style, but in stereo with
2 references and using an Agilent 89441A to do the cross-FFT stuff.
For fun & self education, but on the job there is a E5052B
for reality checks.

A cheaper alternative might be a Microsemi/Symmetricom Timepod.

regards, Gerhard

[Phil Hobbs]

'tain't that hard. With N mixers, you get (N**2-N )/2 independent
cross-correlations, so asymptotically your noise power goes down by
1/N**2. That's pretty favourable compared with straight averaging,
which goes as 1/N.

That's how you measure really low noise levels as well--multiple FET
buffers, which don't contribute current noise (which wouldn't be
uncorrelated like the voltage noise).

[Gerhard Hoffmann]

Am 01.06.2016 um 17:39 schrieb John Larkin:

>
> Sounds right. Adding the phase noise of multiple mixers can't help.
>
> Multiple mixers help if they are used in parallel. I could literally
> parallel four mixers (XOR or Gilberts have fairly high input
> impedances, so string them along differential transmission lines) and
> sum their outputs, and gain 6 dB s/n. But I may as well digitize all
> four outputs. Then I can sum digitally, or do correlation tricks to

Hey, hardware multiplication is MY fetisch :-) I did it with 20 ADA4898
opamps to get to 220 pV/sqrt Hz, built but not yet verified with 8
Interfet IF3601 (should be somewhat better and avoid those huge
capacitors at the input) and already made the layout for 1:8 dividers /
combiners and ring mixers with MAcom baluns and Avago low 1/f schottkys.

> get way better than 6 dB. There are commercial phase-noise measurement
> systems that do correlation of two mixer outputs, so it would only
> take a really smart math guy to extend that to four mixers. I have
> access to really smart math guys.

Yes, that reduces the averaging time until you hit the limit.
Thermal noise is hard to beat.


Gerhard

[John Larkin]

On Wed, 1 Jun 2016 11:55:53 -0400, Phil Hobbs

Given two mixers with the same input signals, I guess you'd digitize
the output of both and do the complex FFT of each. Then at each
spectral point, do the dot product if the vectors and sum the result
into a third, scalar array. Do that lots of times, and the correlated
noise accululates in the scalar array but the random (mixer/amp/ADC)
noise doesn't.

For multipler mixers, just pick all possible pairs and sum the dot
products into that same single scalar array. The result is eventually
the phase noise spectrum. Something like that.

[George Herold]

Yeah, youd' have to lock the laser to something.. too fussy.
This is probably a stupid idea, but you could send both signals
into the same laser diode.. and mix the result in the photodiode.
(I think that works..?) Most likely horrible IMD.

George H.

[Phil Hobbs]

It isn't hard to show analytically, either. An easy-to-remember rule of
thumb is that if you feed a BJT diff pair a current with full shot
noise, the collector currents each have exactly full shot noise
regardless of delta-V_BE. (This assumes infinite beta and zero R_ee'.)
You can figure out other cases from that one.

Plus there's the noise of the base resistance. Both of those go away if
you use a quiet bias current and switch hard.

[Phil Hobbs]

The physical origin of base current is recombination, which is
Poissonian. Thus I'd expect the base current noise to be white,
irrespective of f_T. Capacitance doesn't contribute noise.

[Phil Hobbs]

Of course that assumes that the collector current is quiet, as it will
be with a quiet tail source and hard switching, and the collector load
resistance is small. Otherwise the collector shot noise and resistor
Johnson noise will get coupled into the base and make the base current
noise go up with frequency.

[Phil Hobbs]

NIST and JILA pipe atomic clock signals in fibre back and forth for a
few miles under Boulder traffic. Makes a dandy seismometer too. ;)

[Phil Hobbs]

Yup. Parseval's theorem says that the for cross-correlation,

g star h => GH*

so you can use some convenient-length FFT, pick just the coefficients
you care about, and multiply-accumulate all the cross terms into a
single array. (Personally I'd probably keep the individual cross terms
to allow sanity checking.)

[John Larkin]

On Wed, 1 Jun 2016 13:50:54 -0400, Phil Hobbs

Looks like a quad-core 3 GHz Xeon can do a 64K point single-precision
FFT in vaguely under 1 second. So we'd get some serious noise floor
reduction in an hour.

It is hard to tell how fast FFTs run on various CPUs. Execution times
are given in Mflops or somethings per nanosecond and all sorts of
units.

--

John Larkin Highland Technology, Inc

[Phil Hobbs]

>Looks like a quad-core 3 GHz Xeon can do a 64K point single-precision
>FFT in vaguely under 1 second. So we'd get some serious noise floor
>reduction in an hour.

Yikes, that's only about 20 Mflops. It should be a good thousand times faster than that. What is it, a batch file? ;)

Cheers

Phil Hobbs

[John Larkin]

Rob says a 64K point 1D complex FFT, using the Python library, will be
done before you take your finger off the key. He says a 10 megapoint
FFT won't be especially boring to wait for.

[John Larkin]

Rob just did a 16M 1D complex FFT in 4 seconds. It's the Python
library, which uses FFTW.

So 64K should take about 11 milliseconds. That's "vaguely under 1
second."

[Tim Wescott]

If Python is like Scilab, it'll be fast as long as you're doing a lot of
work in library code and not much in the "glue code" that actually
executes in the Python interpreter.

I tend to prototype that sort of thing in Scilab, then if it's not mostly
running out of library code, put all of it or chunks of it into C++.

--
Tim Wescott
Control systems, embedded software and circuit design
I'm looking for work! See my website if you're interested
http://www.wescottdesign.com

[Joe Gwinn]

In article <2cstkbh12uk1k0in5...@4ax.com>, John Larkin

Here is the latest improvement in DMTD, from a Time Nuts posting. Get
your copy of the Rev Sci Insts article now, before the one month
expires.

============================================

Message: 11
Date: Wed, 25 May 2016 16:01:51 +0000
From: "Sherman, Jeffrey A. (Fed)" <jeff.s...@nist.gov>
To: "time...@febo.com" <time...@febo.com>
Subject: [time-nuts] Commercial software defined radio for clock
metrology
Message-ID: <16A8CDC4-DBF9-49F3...@nist.gov>
Content-Type: text/plain; charset="us-ascii"

Hello,

A recently published paper might be of interest to the time-nuts
community. We studied how well an unmodified commercial software
defined radio (SDR) device/firmware could serve in comparing
high-performance oscillators and atomic clocks. Though we chose to
study the USRP platform, the discussion easily generalizes to many
other SDRs.

I understand that for one month, the journal allows for free electronic
downloads of the manuscript at:
<http://scitation.aip.org/content/aip/journal/rsi/87/5/10.1063/1.4950898>

(Review of Scientific Instruments 87, 054711 (2016))

Afterwards, a preprint will remain available at:
<http://arxiv.org/abs/1605.03505>

There are commercial instruments available with SDR architecture
under-the-hood, but they often cost many thousands of dollars per
measurement channel. In contrast, commercial general-purpose SDRs scale
horizontally and can cost <= $1k per channel. Unlike the classic
dual-mixer time-difference (DMTD) approach, SDRs are frequency agile.
The carrier-acceptance range is limited not by the sample clock rate
but by the ADC's input bandwidth (assuming one allows for aliasing),
which can be many times greater. This property is an important feature
in considering the future measurement of optical clocks, often
accomplished through a heterodyne beatnote (often at "practically any"
frequency between ~1 MHz to 500 MHz) with a femtosecond laser frequency
comb. At typical microwave clock frequencies (5 MHz, 10 MHz), we show
that a stock SDR outperforms a purpose-built DMTD instrument.

Perhaps the biggest worry about the SDR approach is that fast ADCs are
in general much noisier than the analog processing components in DMTD.
However, quantization noise is at least amenable to averaging. As you
all likely appreciate, what really limits high precision clock
comparison is instrument stability. In this regard, the SDR's digital
signal processing steps (frequency translation, sample rate decimation,
and low-pass filtering) are at least perfectly stable and can be made
sufficiently accurate.

We found that in the studied units the limiting non-stationary noise
source was likely the aperture jitter of the ADC (the instability of
the delay between an idealized sample trigger and actuation of the
sample/hold circuitry). However, the ADC's aperture jitter appears
highly common-mode in chips with a second "simultaneously-sampled"
input channel, allowing for an order-of-magnitude improvement after
channel-to-channel subtraction. For example, at 5 MHz, the SDR showed a
time deviation floor of ~20 fs after just 10 ms of averaging; the
aperture jitter specification was 150 fs. We also describe tests with
maser signals lasting several days.

Best wishes,
Jeff Sherman, Ph.D.
--------------------------------------------------------------------
National Institute of Standards & Technology
Time and Frequency Division (688)
325 Broadway / Boulder, CO 80305 / 303-497-3511

=================================================

Joe Gwinn

[George Herold]

Wonderful! Thanks Joe. I get monthly emails from AIP with the table of contents
for RSI and App. Phys. Lett. but I never have time to do more than browse the titles.

George

[bill....@ieee.org]

Nice. I've downloaded it.

--
Bill Sloman, Sydney

[John Larkin]

Hey, we met John Miles, the Timepod guy, at the MTTS show. He made the
pilgrimage by foot (about a mile) to Highland World Headquarters. He
may give us some help on this.

The pod only works to 30 MHz, and I don't think it likes square waves.
But it is a fact that a really good ADC can have a fast s/h in the
front end with fs apearture jitter. But lots of LSBs of noise.

I need to read the Timepod manual. It's 163 pages.

[whit3rd]

On Wednesday, June 1, 2016 at 7:06:10 AM UTC-7, Phil Hobbs wrote:
> On 05/31/2016 05:00 PM, whit3rd wrote:
> > On Tuesday, May 31, 2016 at 1:18:48 PM UTC-7, John Larkin wrote:
> >> On Tue, 31 May 2016 12:32:35 -0700 (PDT), whit3rd <whi...@gmail.com>
> >> wrote:
> >>
> >>> On Tuesday, May 31, 2016 at 12:07:05 PM UTC-7, Tim Wescott wrote:
> >>>> On Tue, 31 May 2016 00:18:47 -0400, Phil Hobbs wrote:
> >>>
> >>> [about mixer noise]

> >> Saturation is good, because it makes the detected output independent
> >> of the signal amplitudes.

> > That's true of the LO amplitude, but not of RF input, or device internal noise sources.
> > Getting rid of one problem, but leaving others, and introducing new ones...
> >
> >> Sure, if the mixer saturates, all the information is in the edge
> >> timing. But that's good.
> >
> > Why? A continuous-multiplier mixer doesn't sense noise from inband
> > harmonics of one input signal, but a switching/XOR mixer does. It's
> > aliasing that noise source so you cannot later filter it away.

> I don't think it's quite that simple. The Fourier series for a square
> wave dies off as 1/f, and has only odd harmonics, so the noise gain vs.
> the sinusoidal case isn't very large. Since you're LP filtering the
> output, you only see noise in narrow passbands centred on the harmonics,
> and the sensitivity goes as 1/N:
>
> noise gain = sum (n=1,2,....) 1/(2n+1)**2

True, but my concern was not with stochastic noise alone; there's also input from
any interfering signal. The wideband nature of a square-wave-handling mixer
opens it up to beating of any harmonic of the signal with any harmonic of any
oscillator nearby. I've seen this happen, and it took a LOT of time and effort
to track to the source. The system had dozens of DC/DC converters, and one
of them interacted with a ground loop... and interference showed up in a sensitive signal
path, a foot away.

I caught it by applying freeze-mist while watching the output spectrum.

Probably (because both the described signals ARE square waves) there's no
way to improve matters for the original problem, as stated. I'd hope for better
performance, though, with linear mixing and narrowband signals, just because
the out-of-band pickup stung me that once... The band of interest is small,
and out-of-band is big and scary.

[Phil Hobbs]

Awesome. Whadda ya know, an instruments paper that's really worth
something!

[Joe Gwinn]

In article <5752145C...@electrooptical.net>, Phil Hobbs

And readable too. Thanks.

And this is the crowd that will be able to implement the design of the
paper. The Time Nuts folk are all over it as well, because this is
something that can be home brewed.

Joe Gwinn

[bill....@ieee.org]

> Awesome. Whadda ya know, an instruments paper that's really worth
> something!

The instrument literature is just like the rest of the peer-reviewed scientific literature - 90% of it is rubbish. Not necessarily wrong, but mostly unhelpful.

--
Bill Sloman, Sydney