Need Guidance on Everyone's Favorite Subject: Grounding!

[ignilux]

All-

As the title says, I need some guidance on a grounding scheme for a project that I have in the works. Having recently repaired an old Datum 9390 GPS clock and timing unit, I'm building a distribution amplifier to route the 10 MHz reference to various test equipment around the lab. I'm hitting the point of "analysis paralysis" concerning how I'm going to ground the RF reference input and amplifier outputs. I've read through many threads here, consulted my copy of Ott's EMC Engineering, dug up an old IEEE Emerald Book, and browsed various distribution amplifier designs by others. Still many questions remain.

So, here's the premise. This amplifier is going to be powered by a toroidal transformer + linear supply, and housed in an aluminum chassis. This necessitates that the chassis will be connected to earth ground. The reference input comes from the GPSDO, and the chassis and coax shield are both connected to earth ground. The test equipment to which the outputs will be connected similarly has connector shields and chassis connected to earth ground. Because I am using the coax for a 10 MHz signal, the shield must be RF grounded to the respective chassis on either end of each cable. "RF ground loops" aren't much of a concern, since return currents will take the path of least impedance. To minimize loop area (and therefore inductance) this results in the entirety of the return current flowing along the inside of the shield, where it belongs. Ott and others consistently champion the idea that the connection between circuit ground and the chassis should in one place: the I/O section of the board. Considering the signal frequency, many, low inductance RF grounds are warranted. This much is clear to me (though, please correct me if I'm wrong).

Here's where it gets murkier. It seems to me that there are two approaches one could take in order to address low frequency ground loops, e.g. 50/60 Hz mains. The first would be to use isolated BNC connectors for the coax connections in order to prevent DC and low frequency common mode currents from riding along the outside of the coax shield. The coax shield could then be tied to the chassis through a (~10n) capacitor to provide a good RF ground. The second would be to not do anything about low frequency ground loops, and instead directly connect the coax shield to the chassis. Since the chassis is tied to earth ground, the hope is that the majority of the common mode current would flow along the chassis to the mains earth lead, rather than through the circuit ground.

However, this second solution presents its own issues. The amplifier outputs will each be connected to a different bit of test equipment, conceivably at different common mode voltages with respect to eachother. At the amplifier output, the path to earth ground through the chassis may have a resistance comparable to the path along an adjacent channel's coax. This will result in noise current flowing where only signal should be, creating potential issues (no pun intended...). But why should I care, when the offending noise is way outside the band of my signal? Well, that's the thing. In the interest of minimizing phase noise, the signal is not filtered upon entering the amplifier, except for an AC coupling cap (~100n). It then runs through two consecutive opamp stages with noninverting 2x gain. So sure, maybe amplifying the noise by a factor of four isn't a big deal, but reviewing other designs makes me think that it may be. In particular, transformers at the inputs and/or outputs seem to be popular.

Take, for instance, the SRS FS710 distribution amplifier. The schematic (see below) states that the "Input BNC is isolated from chassis ground", and uses a transformer for coupling the signal into the circuit. Curiously, they only do this on the input. All of the outputs have their shields connected to the chassis. Any speculation on the reasoning here? For another example, TAPR's TADD-1 (also below) uses transformers on both the input and the output connections (see page 19). The use of a transformer would provide some filtering of noise both below and above the passband of the particular part that I use, and would have the bonus of providing some isolation between channels. But then, if the coax connectors are bolted directly to the chassis, the isolating properties are lost. If isolated BNC connectors are used, then we're back at the original solution to preventing ground loops (capacitively coupled shields).

Finally, after thinking about it for a good long while I can't decide whether or not I want my circuit ground to be connected to the chassis at all. It seemed like an obvious thing to do, but the TAPR design appears to have an isolated circuit ground.

The more I think about all of this, the further I get from a decision. Any insights, experiences, etc. are welcome and appreciated.


Edit: Added an image of the TAPR design for the lazy, so you don't have to go search a PDF.

[TimFox]

Have you considered adding external 50 ohm/50 ohm transformers (e.g., from MiniCircuits) between the amplifier outputs and the remote loads?  You may not need it, but it would be good to have them for Plan B.

[ignilux]

Why external, and why transformers at all? Why not on the input?

I appreciate the suggestion, but the whole point of building your own test equipment is learning. I could blindly follow anyone's advice (and yours may well be great advice, don't get me wrong), but then I will have learned nothing.

[TimFox]

I can't guarantee the solution, but it is probably worth your while to investigate it.  The reason I suggest it is to remove possible ground loops at 60 Hz, while allowing full-fidelity transmission of the 10 MHz.
By "external", I mean external to the circuit you posted, after the grounded coax connectors.
The expensive RF transformers from MCL have connectors, but you could easily mount the plastic-cased ones yourself.  With the DIP or similar packages, there should be no galvanic connection between primary and secondary, unless MCL specifically shows it on their diagram for that part.

[SiliconWizard]

Have you considered using a differential output?

[ignilux]

I can't guarantee the solution, but it is probably worth your while to investigate it.  The reason I suggest it is to remove possible ground loops at 60 Hz, while allowing full-fidelity transmission of the 10 MHz.
By "external", I mean external to the circuit you posted, after the grounded coax connectors.

Why not place the transformer inside the circuit, before the output connector, then use an isolated connector? Doesn't that do the exact same thing without investing in connectorized transformers?

Have you considered using a differential output?

I would love to be able to do so, because that would greatly simplify things. However, the equipment that will be connected to the output of the amplifier is commercial gear from Tek, HP, Racal, etc. The circuitry inside is expecting a single-ended signal fed from standard 50 ohm coax.

[T3sl4co1l]

You can use coax differentially, nothing wrong with that.

The considerations are:

1. Lifting GND at DC-LF should be done within limits, for safety and transient reasons.  Typically a low value resistor (10-100R?) or TVS is used to limit CM range.  Average voltage should be quite low, limited to DC power (GND return drop), ground loop if applicable (up to a few volts between multiple earthing points?), and ambient fields (mains, SMPS noise, radio).

1a. Ground must be lifted at at least one end.  You can always isolate both ends, but lifting the transmitter (unless via transformer) is likely the hardest, so I'll assume receiver instead.

2. The shield needs good, contiguous RF grounding.  Ideally a coax feedthru capacitor would be used to provide this.  That's pretty boutique, so the next best thing is a connector with a wide footprint, surrounded by lots of bypass caps.

3. The signal input needs to subtract the ground voltage, up to whatever frequency the bypassing takes over at (and somewhere beyond that, to keep CMRR good).  At 10MHz, this isn't hard to do with op-amps (or any other differential, non-operational amp) or transformers.

4. Optional: increase the CM impedance of the cable by adding a ferrite bead or a few.  This reduces the Q factor of standing waves on the shield, between the end points, reducing the amount of interference (in selected (resonant) bands) that the bypassing has to deal with.


Note that transformers are balanced by nature; you won't get as much CMRR in this unbalanced configuration.  You can wind one with internal shields to help with this, so that CM is taken up by ground on either side, away from the signal.  The windup should be: primary, foil shield, foil shield, secondary.  (A foil shield takes the form of a slitted cylinder, so that inhomogeneous fields from the wires are blocked, while the overall solenoidal field is passed.  The primary and secondary don't "feel" each other as such (as turns of wire), but rather as the image currents in the one-turn shields, as if they were the windings in the first place.)  With whatever insulating tape between layers as needed.  Note that this gives a double round-wire-over-ground-plane transmission line configuration, so the characteristic impedance will be fairly high, maybe 100-150 ohms.  (With single layer windings, this is a simple transmission line transformer.)  As long as winding length is much less than 1/8 wave at 10MHz, this won't be a problem.  If desired, wire can be doubled up, to not-quite-halve the characteristic impedance, at the expense of turns per layer (basically, using a wider trace over the ground plane).  A small pot core and bobbin will probably give the best bandwidth while being easy to wind; ETD, RM or PQ should also give good results, having wider winding areas (fitting more turns in a single layer).

Note that wire size doesn't matter, as long as DCR is low enough for purposes of insertion loss and power handling.  What matters to the characteristic impedance (Zo = sqrt(LL/Cp)) is the ratio of wire size to height over ground plane.

Tim

[TimFox]

I can't guarantee the solution, but it is probably worth your while to investigate it.  The reason I suggest it is to remove possible ground loops at 60 Hz, while allowing full-fidelity transmission of the 10 MHz.
By "external", I mean external to the circuit you posted, after the grounded coax connectors.

Why not place the transformer inside the circuit, before the output connector, then use an isolated connector? Doesn't that do the exact same thing without investing in connectorized transformers?

Have you considered using a differential output?

I would love to be able to do so, because that would greatly simplify things. However, the equipment that will be connected to the output of the amplifier is commercial gear from Tek, HP, Racal, etc. The circuitry inside is expecting a single-ended signal fed from standard 50 ohm coax.

You can place the insulated transformer at either end of the 50 ohm coax to break the 60 Hz ground loop—whichever is more convenient to your connections.  The shield of the coax will be at low-frequency ground at only one end.

[ignilux]

1a. Ground must be lifted at at least one end.  You can always isolate both ends, but lifting the transmitter (unless via transformer) is likely the hardest, so I'll assume receiver instead.

To be clear, this device will be both a receiver and a transmitter. So it will be easiest to lift the ground at the amplifier itself rather than the GPSDO source or the test equipment destination(s).

2. The shield needs good, contiguous RF grounding.  Ideally a coax feedthru capacitor would be used to provide this.  That's pretty boutique, so the next best thing is a connector with a wide footprint, surrounded by lots of bypass caps.

I agree that a coax feedthrough cap would be ideal here. There is a picture in Ott's book of an XLR connector with a ring of radially connected ceramic caps that I find fascinating. I'm not sure I understand how your alternative solution provides similar function though. The shield would have to be bypassed to the chassis, but how would that connection be made without considerable inductance from component leads?

3. The signal input needs to subtract the ground voltage, up to whatever frequency the bypassing takes over at (and somewhere beyond that, to keep CMRR good).  At 10MHz, this isn't hard to do with op-amps (or any other differential, non-operational amp) or transformers.

I hadn't really thought about transformers like this in the past, but it is neat that (for the 1:1 turns ratio case) they really do just act like a unity gain differential amplifier. The voltage difference across the primary winding appears across the secondary. Connect one end of the secondary to ground, and you have a differential to single-ended conversion.

Note that transformers are balanced by nature; you won't get as much CMRR in this unbalanced configuration.  You can wind one with internal shields to help with this, so that CM is taken up by ground on either side, away from the signal.

I am definitely not up to the task of winding an electrostatically shielded transformer for the input and each of the outputs. Is the degradation in CMRR due solely to the interwinding capacitance in the transformer?

[cdev]

here is a design for a 10 MHz distribution amplifier on the NIST site, it uses transformers to isolate its inputs from CM signals. I think.

[T3sl4co1l]

1a. Ground must be lifted at at least one end.  You can always isolate both ends, but lifting the transmitter (unless via transformer) is likely the hardest, so I'll assume receiver instead.

To be clear, this device will be both a receiver and a transmitter. So it will be easiest to lift the ground at the amplifier itself rather than the GPSDO source or the test equipment destination(s).

Also, to be clear, by "lift" I mean anything other than hard earth grounding between the cable and circuit ground.  As noted, the "lift" may be conditional e.g. over a modest voltage range, or with some resistance, and in any case resuming a low impedance at AC.

The box itself should be grounded, for reasons as given, and also to set a ground on the transmitting sides.  I don't mean "lift" in the sense of physically cutting a ground pin, as it often means in unfortunately-too-common parlance!

Unless you mean something bidirectional in which case you may need a more involved solution, but I understand a "distribution amplifier" to have one input and multiple outputs.  In which case, that one input can have all the signal conditioning needed, and the rest is hard grounded, easy peasy.

I agree that a coax feedthrough cap would be ideal here. There is a picture in Ott's book of an XLR connector with a ring of radially connected ceramic caps that I find fascinating. I'm not sure I understand how your alternative solution provides similar function though. The shield would have to be bypassed to the chassis, but how would that connection be made without considerable inductance from component leads?

Yes, like that.

You do it without component leads, obviously!

Or enough in parallel that it's enough not to care.

Ceramic chips are cheap and affordable, and seam connections can be made between panels and PCBs with EMI gaskets.

It also doesn't necessarily have to be coronal, it just needs to surround the signal trace well enough to provide shielding.  For example, consider a connector at the end of a rectangular PCB.  The center pin transitions to a trace which goes along the board (preferably on an inner layer, surrounded by planes).  The whole end of the board can be slotted across (except for the trace), with bypass caps stacked across the ground slot, top and bottom (preferably) to provide the AC bypass with DC isolation.  Then the ground return (for EMI purposes) path mates to earth-ground somewhere just past the bypass caps, while the signal proceeds over, now being surrounded by the local ground potential.

Descriptions are probably ineffective, and anyway Ott provides diagrams; they should match this description, no?  Or do you have any particular concerns with them, or just need help explaining like, equivalent circuits of geometries or whatever..?

I hadn't really thought about transformers like this in the past, but it is neat that (for the 1:1 turns ratio case) they really do just act like a unity gain differential amplifier. The voltage difference across the primary winding appears across the secondary. Connect one end of the secondary to ground, and you have a differential to single-ended conversion.

Yes.  Well, it's not amplifying, there's no isolation, you need a terminator on the load side and you want to avoid weird impedances, nonlinearities or backfeeding up the line -- there are of course always two waves on the line, at all times, one in each direction -- more it's an extension of the line itself; which is what's particularly emphasized with transmission line transformers, as having predictable characteristics and relatively easy design.

It's also not perfectly differential, that's what the shields are for of course; coax isn't balanced so we need to take care of that.  If it were twisted pair, we'd use a single winding with no shield (or one shield for an unbalanced output), and indeed that's what Ethernet for example does.  They also add a CMC (just a bunch of twisted pair done up on a choke core) for better immunity.

A key insight is that a twisted-pair transformer is an isolation transformer either way you wire it.

Obviously, as a CMC, it doesn't isolate down to DC, but over a given bandwidth, it has at least some specified CM impedance, and so allows a degree of freedom with relatively little current flow.  The upper bandwidth however can be essentially infinite, limited by the spacing of the turns of transmission line, core characteristics, and the transmission line itself (as a finite width TL has some cutoff frequency, above which non-TEM00 modes appear).  Note that CM bandwidth is limited, while signal bandwidth is "infinite".

As a transformer proper, it isolates down to DC (CM), but signal bandwidth is limited.  In particular, in a similar way that the CMC upper cutoff is limited by geometry and losses, the signal bandwidth here is limited by winding length.  (We can model the transmission line as two ideal ports with some CM impedance between them; in a given instant, the ports have TL impedance, so each port acts in series between the source and load.  The equivalent circuit is one loop: source, Rs, port 1, RL, load, port 2.  After one propagation delay, the ports receive each others' signals, and interference occurs; this causes a notch at the 1/2 wave frequency and harmonics.  It stands to reason that, if we had a lossy transmission line, i.e. that was terminated in the middle, we could get unlimited bandwidth, if we also don't mind 6dB insertion loss.)  The difference in relevant parasitics is due to the even/odd mode the signal is applied in: one case the transmission line is simply inline between source and load, the other it's a 1/2 wave stub across them.

Even more generally, we can model the transformer as a 4-port where each pin is a port with respect to some reference plane, and by considering pairs and phases of those ports, we can understand more deeply the two cases above, and which parasitics dominate which modes.

I am definitely not up to the task of winding an electrostatically shielded transformer for the input and each of the outputs. Is the degradation in CMRR due solely to the interwinding capacitance in the transformer?

As above, it's a bit more complicated than that, but that is the LF asymptotic case, yes.  At transitional frequencies, you also have to worry about the delayed common mode.  That is: drive both primary pins high; instantly, the secondary pins both go up, then one delay later, the propagated waves bring them back down.  For an ideal terminated TL, the step response is a rect pulse.  So, the impulse response is a complementary pair of deltas, so the frequency response is a sine wave starting at zero (so, rising ~proportionally from there, hence ~capacitive CM response), peaking at the resonant frequency, then valleys and peaks at harmonics.

Tim

[ignilux]

First of all, thanks for your helpful and insightful responses. Not just here, but other threads over the years.

Unless you mean something bidirectional in which case you may need a more involved solution, but I understand a "distribution amplifier" to have one input and multiple outputs.  In which case, that one input can have all the signal conditioning needed, and the rest is hard grounded, easy peasy.

That topology is correct. Why don't I need to isolate the outputs in a similar manner? Is it simply due to the fact that the input is where there is gain in the signal + noise path? Or am I trusting the fact that professionally designed test equipment will have the requisite interface circuitry on their end?

Ceramic chips are cheap and affordable, and seam connections can be made between panels and PCBs with EMI gaskets.

Aha, I hadn't considered EMI gaskets. That seems obvious, now.

It also doesn't necessarily have to be coronal, it just needs to surround the signal trace well enough to provide shielding.  For example, consider a connector at the end of a rectangular PCB.  The center pin transitions to a trace which goes along the board (preferably on an inner layer, surrounded by planes).  The whole end of the board can be slotted across (except for the trace), with bypass caps stacked across the ground slot, top and bottom (preferably) to provide the AC bypass with DC isolation.  Then the ground return (for EMI purposes) path mates to earth-ground somewhere just past the bypass caps, while the signal proceeds over, now being surrounded by the local ground potential.

Descriptions are probably ineffective, and anyway Ott provides diagrams; they should match this description, no?  Or do you have any particular concerns with them, or just need help explaining like, equivalent circuits of geometries or whatever..?

That configuration makes sense. I'm not sure I understand the ground return, though. Naturally the signal trace needs a return path for the current, but why doesn't it cross the ground slot through one of the bridging capacitors? I may not be visualizing what you're saying correctly. Ott indeed does provide diagrams, but it can sometimes be difficult to find the right one for the task at hand. They also tend to leave out some dumb detail that shouldn't matter, but ultimately prevents me from really "getting it".

A key insight is that a twisted-pair transformer is an isolation transformer either way you wire it.

I hadn't thought about that before. It's pretty elegant how the two configurations give reciprocal behavior, too.

What do you think about something like Fig. 5-13 in Ott (see below) as an alternative to a hand wound electrostatically shielded transformer?

[TimFox]

Pre-made transformers at MCL:  many cover your frequency range.
https://www.minicircuits.com/WebStore/Transformers.html
The simple ones at the top of the list, e.g. ADT1-1+, with no inter-winding shield but galvanically isolated for 50/50 ohms, max RF power 250 mW = +24 dBm, run about $3.00 US in factory quantities of 20, but should be available from distributors.
I like MCL since they publish detailed specifications on their parts.

[T3sl4co1l]

First of all, thanks for your helpful and insightful responses. Not just here, but other threads over the years.

You're welcome!  I appreciate it.

That topology is correct. Why don't I need to isolate the outputs in a similar manner? Is it simply due to the fact that the input is where there is gain in the signal + noise path? Or am I trusting the fact that professionally designed test equipment will have the requisite interface circuitry on their end?

At least one end should have it.  If you can't guarantee the other end, then yeah, it's a good idea for the outputs as well.

A whack of transformers is starting to sound very attractive.  Give or take if they're hand wound or not... 

That configuration makes sense. I'm not sure I understand the ground return, though. Naturally the signal trace needs a return path for the current, but why doesn't it cross the ground slot through one of the bridging capacitors? I may not be visualizing what you're saying correctly. Ott indeed does provide diagrams, but it can sometimes be difficult to find the right one for the task at hand. They also tend to leave out some dumb detail that shouldn't matter, but ultimately prevents me from really "getting it".

Because you can't drill through capacitors with a pick-and-place?   The trace might well cross underneath a cap, or flanked by several, whatever.  The important part is to minimize AC voltage across the gap, because that same voltage will be impressed on the signal.  And the easiest way to do that is to use multiple caps in parallel, easily keeping the ESL under 1nH for example.

Note that you're still making a reference connection to the far side, so two traces actually cross the gap, one is just low bandwidth so it can be pretty much wherever.  Then they go to your differential receiver.

Speaking of quantities, any bandpassing you can do, will further reject interference.  I'm assuming you want the signal as clean as possible, from the source, with as little outside interference as possible.  So that the signal is suitable for any kind of reference input, filtered or PLL'd or otherwise.  So high CMRR is valuable, and some conditioning (filtering) may be desirable too.

For reference, an AC ground loop of 1nH is -40dB from 50 ohms (i.e. 0.5 ohm) at 80MHz.  So we might expect about as much susceptibility at the same point.  If your lab is near a commercial FM transmitter, you may well have 10s of mV at ~100MHz, so you might expect 100s of µV onboard.  Is that good enough?  If not, additional shielding/bypassing and filtering may be desired.

What do you think about something like Fig. 5-13 in Ott (see below) as an alternative to a hand wound electrostatically shielded transformer?

I don't get it, it's shorting the grounds together.  Was the diagram indicating AC offset only?  Or did you forget some coupling capacitors?

A differential pair of coupling capacitors, serves the same purpose as the ground-ground bypass in the unbalanced case.  Note that they should be closely matched values, to give good CMRR through the transition band; or the values should be so large that the transition band falls well outside the signal bandwidth, so the CM-DM mode conversion can be filtered away.

Tim

[ignilux]

I've always found it's beneficial to have something concrete to point at when discussing technical things. Words are only so useful in this industry.

To that end, I created a sample circuit containing one representative connection. It uses a single, isolated BNC connector whose shield pin is heavily bypassed to the chassis ground fill (on top and bottom layers) with 10n, 0603 ceramic caps. There is exposed copper along the front edge (top and bottom) to facilitate a good connection to an EMI gasket, which will be compressed against the front wall of the enclosure. There is also a low capacitance TVS diode between the shield and chassis ground pour to limit the DC voltage between shield and chassis. The split between chassis and circuit ground pours is woven with the same 10n caps on top and bottom layers. All capacitors have multiple vias between the two chassis ground planes, and there is plenty of via stitching throughout. The shield trace itself leaves the footprint pad and immediately passes through a 47n, 0805 feedthrough cap. After that there are two more 10n capacitors, before finally hitting the split in the ground planes. To cross the split, the center pin and shield connect to a 50R transformer (PWB1010-1L). This provides some bandlimiting as well, though admittedly not much. Coilcraft specifies a 3dB bandwidth of 30 kHz to 250 MHz for the transformer in question. I chose that part because I have a bunch on hand from a previous project. There will be ferrites beads over the cable jackets on both the amplifier and source/receiver ends of all cables. If additional bandwidth limiting is necessary, I might add something like a SCLF-10+ 10 MHz LPF on the circuit ground side of the split.

So, how'd I do? 

[T3sl4co1l]

Hmm, interesting.  I don't think that what you've drawn is bad, but it's not what I expected or described.

To clarify, the discussion of transformers is somewhat separate -- their powers of isolation are much stronger, so we don't need to go through such hoops to get good signal quality.  For example, adding CMC(s) in front of an unshielded transformer to improve its CMRR.  (Which I think I hinted at with the Ethernet example, but didn't motivate/develop; and which can of course be derived from the transformer-as-a-4-port, but you'd need to go and do all the work to figure that out... or have the instant insight realizing what the heck all of that means, and I can't say I've ever felt quite that smart, myself.  [It's one thing to eventually come to understand something in high level terms; it's quite another to dive right into that...])

A couple of concerns:
- That's one of those... shitty BNCs, right?  The two welded pins at the rear, plastic housing, no shell ground?  So there's a length of parallel line, completely unshielded, between the connector and PCB.

We cannot use these connectors in a low-impedance context.  Any CM current flow, drops voltage across that generous, what, ~10nH of ground pin, and boom, so goes your signal quality.

It's not useless, as we can get quite good isolation with suitable transformers -- but we certainly cannot do the hard-grounded solution that's under discussion.

- I don't get the passthru cap for GND.  The point of heavy bypassing is to keep GND at GND, for AC; so what's it passing, just filtering DC?  But for the transformer?... so, it's weird.  And it's asymmetrical, the same isn't done to the signal side (for, somewhat more obvious reasons ).  I mean, it's unbalanced so there should be some asymmetry, but it's... well it's just not the right one.

- I was expecting ground connection (EMI gasket) on the other side -- "past the bypass caps".  I mean, putting it by the connector is obviously more convenient [for that type of connector], but we can't ground at that point because of the DC offset; we need bypass caps inbetween them.  Which I think you understand, which is why you've lifted connector ground from it, but you weren't sure what else to do, and also whatever about the transformer.

I'm also assuming it's normal and fine for local circuit ground to also be tied with chassis ground.  Which if not, then maybe another row of bypasses is in order (or whatever is low enough impedance, like one or two caps in the corners at the mounting screws).


This is what I was thinking:





(That's "with vert. con.")

Not sure exactly how clear that is, I know I crammed labels a bit once or twice. Who needs pencils and erasers...

Tim

[Terry Bites]

I'm thinking dirt cheap baluns made from ethernet transformers and STP network cable.

[T3sl4co1l]

This is, after all, precisely what they're made for.

Tim

[David Hess]

Transformer isolation for each output removes a lot of possible grounding issues, and if Ethernet transformers are suitable which they should be, then inexpensive.

[TimFox]

I don't know how much the Coilcraft transformers mentioned above cost, but the MCL RF transformers I cited are $3.00 from the factory (minimum order 20 pcs).  Both vendors' parts are designed for RF use in 50 ohm systems.

[ignilux]

Hmm, interesting.  I don't think that what you've drawn is bad, but it's not what I expected or described.

I had a feeling that we were talking about two separate things. Hence the visuals added on my part. Thanks for replying in kind!

- That's one of those... shitty BNCs, right?  The two welded pins at the rear, plastic housing, no shell ground?  So there's a length of parallel line, completely unshielded, between the connector and PCB.

Yes, it is! I've been skeptical of them in the past for many reasons, but I'll admit that I hadn't thought about the unshielded parallel line in an EMI context. The problem that I'm having is that an isolated, panel-mountable connector of the type you reference in your drawings doesn't seem to exist. I could forego the panel mounting if I use the "vertical" scheme, but that presents a whole different set of challenges when looking for an enclosure, etc.

- I was expecting ground connection (EMI gasket) on the other side -- "past the bypass caps".  I mean, putting it by the connector is obviously more convenient [for that type of connector], but we can't ground at that point because of the DC offset; we need bypass caps inbetween them.  Which I think you understand, which is why you've lifted connector ground from it, but you weren't sure what else to do, and also whatever about the transformer.

I'm also assuming it's normal and fine for local circuit ground to also be tied with chassis ground.  Which if not, then maybe another row of bypasses is in order (or whatever is low enough impedance, like one or two caps in the corners at the mounting screws).

Aha! That's a great example of the power of pictures and drawings. "Past the bypass caps" could mean quite literally anything. My connection to the chassis is past the bypass caps if your perspective starts at the hypothetical power supply at the other end of the board. Re: Circuit ground at earth ground potential, I was really wondering how that was going to be done. I do want circuit ground connected to the chassis to avoid a giant dipole between the circuit ground pour and enclosure (and for safety), and it makes much more sense seeing the way that you've done it.

I'm thinking dirt cheap baluns made from ethernet transformers and STP network cable.

This is, after all, precisely what they're made for.

Transformer isolation for each output removes a lot of possible grounding issues, and if Ethernet transformers are suitable which they should be, then inexpensive.

Using a transformer as the differential receiver / driver sounds like the way to go. Looking through datasheets for pulse transformers shows the anticipated CMC / "current balun" in series with a DC isolating transformer / "voltage balun". Assuming that the signal is fed into the choke end, is it correct to think of it as the CMC providing CM rejection and better balancing the signal so that the transformer "sees" what it wants (a balanced signal), and in turn provides better performance? Then, because the rest of my circuit is single ended, I can tie one end of the transformer to ground and take the signal from the other terminal? Finally, am I correct that a balanced 100 ohm telecom/pulse/ethernet transformer provides a 50 ohm input/output impedance to single ended signals?

[T3sl4co1l]

Yes, it is! I've been skeptical of them in the past for many reasons, but I'll admit that I hadn't thought about the unshielded parallel line in an EMI context. The problem that I'm having is that an isolated, panel-mountable connector of the type you reference in your drawings doesn't seem to exist. I could forego the panel mounting if I use the "vertical" scheme, but that presents a whole different set of challenges when looking for an enclosure, etc.

Hm, may have to resort to shoulder washers then.  It's not at all an impossible feat; my signal generator is stuffed full of them (since it can float 30V to ground).  McMaster 91145A271 (or friends), 93920A170 or 90097A230 are probably a good idea (most of them look kind of long, may have to trim them -- PITA for projects, NG for production), or whatever other sizes fit a given part, and also check metric I suppose.

Using a transformer as the differential receiver / driver sounds like the way to go. Looking through datasheets for pulse transformers shows the anticipated CMC / "current balun" in series with a DC isolating transformer / "voltage balun". Assuming that the signal is fed into the choke end, is it correct to think of it as the CMC providing CM rejection and better balancing the signal so that the transformer "sees" what it wants (a balanced signal), and in turn provides better performance? Then, because the rest of my circuit is single ended, I can tie one end of the transformer to ground and take the signal from the other terminal? Finally, am I correct that a balanced 100 ohm telecom/pulse/ethernet transformer provides a 50 ohm input/output impedance to single ended signals?

They're made for 100 ohm differential, and probably have a characteristic impedance close to that.

Transformers aren't very critical of impedance matching.  As long as they're electrically short compared to the frequencies of interest, LL and Cp are adequate descriptions; and so the bandwidth shrinks proportional to the mismatch.  Uh, sort of; LL and Cp both work against a mismatch, while Lm works in favor of lower Z, at least until insertion loss (DCR) is too high to be practical.  Arguably, in terms of absolute bandwidth, the HF cutoff dominates (100s MHz vs. 10s kHz), but YMMV.

They often have centertaps too, which may be handy.  The media side ("HJ" termination) tap shouldn't be much help as it doesn't give a ground reference; or if you wire it unbalanced, you're using half the winding = quarter the impedance (25 ohms), which, I mean, it's the same magnitude mismatch as the full winding (50/100 = 25/50) so maybe there's something there anyway.  Leaving the other half of the winding open seems peculiar; it shouldn't matter at signal frequencies, and I wonder if it helps to have it terminated with a little R+C at high frequencies, or if it's fine open?  (Since it forms a stub, which will resonate at some 100s MHz.)

The PHY side CT can be grounded for proper balanced output, not that you'll necessarily be using it as such (but a differential receiver will do better; or maybe another CMC, or using a "3 core" transformer that integrates same).  The reason is, the path from cable CM, through media CMC, through Cp, induces a voltage across the PHY side windings, also common mode again, so that a CMC / diff rx. does a bit better.



So like these, in the top case the coax goes right into the transformer with minimal length -- keep Cp (between shield and GND) to a minimum.  PHY side can be single ended (CT GND, take signal from one side, optionally R+C the opposite stub), differential (take both sides, no L1, C1, R2), or CMC'd as shown (note the single ended output, the R+C just dampens the CMC so it's not flapping around; YMMV depending on receiver input impedance and all that).

In the bottom case, the transformer is wired in half, and the opposite windings do still serve some purpose, balancing the electric field in the transformer -- it's a shitty shield substitute but should still be better than a plain twisted-pair transformer.  The R+Cs are just to terminate the stubs at high frequency, and may be optional.

And L2/T1A and L3/T1B are probably integrated in one component, I've shown them for completeness.

Tim

[ignilux]

Thanks for once again appeasing the visual learner in me! 

Taking into consideration the fact that my original plan was to have a single opamp at the input to buffer the incoming signal, I think it makes a lot of sense to replace that with a full fledged differential receiver. Something like the attached schematic (see below), where the connector connects directly to the CMC, the secondary center tap is connected to circuit ground, and the line is terminated differentially at the input to the AD8130.

Now I'm trying to decide how the use of a transformer will change the input signal conditioning. It seems to me that, because the choke specifically common mode currents, not only will it reduce common mode signals at the transformer primary, but all the way up the transmission line and outside the enclosure. Because of that, there is less rejection that has to happen elsewhere in the circuit. Of course, the common mode current that remains still wants to be shunted to the chassis. So my thoughts are that the topologies shown in your initial two drawings are still important to adhere to. However, I'm assuming that something like the current loop between the chassis and connector ground plane is now less critical thanks to the CMC + transformer. Is that all reasonably accurate?

Edit: Words is hard.

[David Hess]

Transformer isolation for each output removes a lot of possible grounding issues, and if Ethernet transformers are suitable which they should be, then inexpensive.

Using a transformer as the differential receiver / driver sounds like the way to go. Looking through datasheets for pulse transformers shows the anticipated CMC / "current balun" in series with a DC isolating transformer / "voltage balun". Assuming that the signal is fed into the choke end, is it correct to think of it as the CMC providing CM rejection and better balancing the signal so that the transformer "sees" what it wants (a balanced signal), and in turn provides better performance?

Ethernet has strict requirements for balance to meet its specifications.

Then, because the rest of my circuit is single ended, I can tie one end of the transformer to ground and take the signal from the other terminal? Finally, am I correct that a balanced 100 ohm telecom/pulse/ethernet transformer provides a 50 ohm input/output impedance to single ended signals?

10 MHz is so far below the transformer's maximum operating frequency that I do not think the design impedance will matter.  If it was an issue, then I would try using two in parallel.

[T3sl4co1l]

Yes, a shield-earth TVS or R or R+C or whatever is still prudent.  A ferrite bead on the cable may therefore still make some difference.

Yeah, 50 ohm termination, and the transformer won't mind, its winding lengths are electrically short, it looks like a little inductance here.

Ethernet has strict requirements for balance to meet its specifications.

I wouldn't exactly say "strict"; the spec is around 40dB in pretty much every figure, so only 1% balance.

The annoying thing is, they never tell you the actual performance.  They only tell you the spec it meets/exceeds.  Same as any damn thing based on an industry standard, they tell you what it meets, not what it is.  Which, since gigE transformers have precisely the same numbers, you'd think there's absolutely no point to the different winding/CMC design and why bother.

So.  It could be that real parts do very well indeed, but how much better, you'll have to measure that yourself.  (Prime opportunity for a test jig and NanoVNA? )

Tim

[ignilux]

So.  It could be that real parts do very well indeed, but how much better, you'll have to measure that yourself.  (Prime opportunity for a test jig and NanoVNA? )

I don't think this was originally directed at me, but you inspired me to dig through the junk bin to find an old router that I was never going to use again. As luck would have it, one of the transformer modules on there was perfect for a quick and dirty proof of concept. It is an old Delta DIP-16 package, with four independent cores, each with a single primary/secondary in 1:1 ratio. Each winding measured in the neighborhood of 260 uH at 100 kHz, which appears to be typical of ethernet transformers. Leakage inductance was under 200 nH at 100 kHz for the pair that I measured. Interwinding capacitance was 8.13 pF, again at 100 kHz.

I hotglued it dead bug style onto a piece of copper clad, and connected three of the windings in the CMC --> Center tapped transformer arrangement shown in the schematic of my last post. Just to make the EMC engineers in the audience shudder, I soldered the coax center conductor and shield pigtail across the transformer instead of bothering with a proper connector. The output was loaded with two paralleled 100R, 1%, quarter watt, THT, metal film resistors. You can see the test fixture in all of its horrible glory below.

I ran two individual tests, both using an original NanoVNA and a span of 1 MHz to 20 MHz. With a 50 ohm load, return loss in the area of 10 MHz was around 16 dB. As predicted, the reflection coefficient shows the load impedance with a bit of extra inductance and resistance. The second test was performed by cutting one of the 100R resistors out of the load. Here the return loss drops to around 9 dB, but the response is much flatter across frequency. I didn't bother to measure S21, since the load is well characterized and mostly resistive. I can't imagine more than 1 or 2 dB of loss was present.

[T3sl4co1l]

Nice!

Can also check out what CMRR is like, or PHY side balance (CT GND, DM or CM balun(s) from winding ends to port 2)?

Tim

[ignilux]

Can also check out what CMRR is like, or PHY side balance (CT GND, DM or CM balun(s) from winding ends to port 2)?

I'm trying to decide how to do this. By DM/CM baluns, I assume you're referring to using a voltage balun (transformer, DC isolating) or current balun (common mode choke, not DC isolating) to convert balanced output back to unbalanced, then feeding that to port 2?

[T3sl4co1l]

Exactly.

The tricky part perhaps is having baluns that are good enough (potentially better than DUT?) to see it clearly.  Or if nothing else, reading both ports independently and doing it in software, if the measurement is accurate enough to take the difference of.

Note port 1 will be terminated, it will just be a resistor in parallel with the primary windings shorted together.  Well, probably slightly better to not short them completely, but use matched (<1%?) resistors to terminate them as well, so that equal resistances go from each pin to the source.  Probably, it would suffice to do, like, port 1: shunt R (50 ohm terminator), tied to primary CT; 50 ohm from CT to start winding; 50 ohm from CT to end winding.

Tim

[ignilux]

The tricky part perhaps is having baluns that are good enough (potentially better than DUT?) to see it clearly.

Right, that was my concern as well. Perhaps time to invest in some tape-wound nanocrystalline cores 

So, I spent a long time today consulting various texts and reading and watching media concerning CMRR. Combined with your suggestion to terminate the center tap of the primary, I have come up with the simulation below. Is this what you had in mind? I'm not sure how much the series resistors on the primary really matter, since we're making a relative measurement between the left and right circuit. So long as they are reasonably close to the characteristic impedance of the system it ought to come out in the wash, no?

The one thing that I learned in my research today is that you can ask 5 people how to test CMRR, and they will give you 5 different ways. 3 of them will be outright wrong. The other two will each agree that eachother's data seems reasonable, but they will fight to the death over a small detail in the procedure. One of the original three will then claim he corrected his mistake and now gets the same data as one of the last two. Then someone from Marki or Wenzel will come along and demonstrably prove everyone wrong, but refuse to explain the method. The war of attrition continues.

[T3sl4co1l]

Like this:

Oh, I suppose the "X" should be optional -- if the line is not included in the primary side CMC, it's probably better not to connect it.  Should be able to measure that too.

Maybe some other arrangement of resistors would do, with the same Thevenin resistance presented across the primary, and to the source, but whatever.

The fact that it's two 25 ohms in series between source and transformer, increases the CM source by 12.5 ohms total.  That could cause a tiny roll-off at HF, but these are low impedances either way -- likely it would amount to a dB or so near a GHz.  And the primary side CMC is way more impedance.  So who cares.  Like, ideally you'd use a balun, but again, where are you going to find one more ideal than the DUT, and that's also better than plain ass resistors at 1GHz?...

Alternately, the source doesn't really care about reflections does it?  I suppose the shunt termination (R3) isn't really needed (get an extra +6dB on your output for free).

Tim

[ignilux]

I'm finally able to show some results after a few days spent trying to create the world's best balun. I had some pretty decent results, but ultimately I decided to reevaluate my strategy. I scrapped the original transformer jig, and instead dug another module out of the junk bin. This one has four sets of three cores when viewed from the bottom, and 12 pins per side. This leads me to believe that it's exactly what I was looking for, i.e. a center tapped transformer with CMCs on either side. It was in an unusually wide SOIC package, though, so I had to get creative and cut a SOIC DIP adapter board in half to be able to easily prototype with it. I hot glued the whole mess to a 0.7mm solid copper shim, and added a couple of SMA connectors. The result is shown below. (The solder looks like sheeit, but that's just where molten solder hit some runny hot glue. All of the joints are happy and healthy.)

As for the data, it looks to me about what I'd expect. S11 floats around just below 0 dB, and S21 is ridiculously low at the 1 MHz start of the sweep, gradually rising as the frequency increases. At the target frequency of 10 MHz, S21 is down around -42 dB. Between the rejection from a ferrite over the coax jacket and the "typical" 70 dB CMRR of the AD8130, it looks like this will be a great solution.

I'm thinking of making the whole signal path be symmetrical: Input coax --> CMC --> Transformer --> AD8130 (differential receiver) --> HIgh speed opamp buffer (OPA890?) --> AD8132 (differential cable driver) --> Transformer --> CMC --> Output coax. Due to the nature of a fully differential amplifier, the isolation between individual channels is compromised when driven from a common source (since the inverting output is fed back to the noninverting input, unlike a traditional opamp). The opamp buffer helps restore this.