Design considerations for a differential pair
Differential pairs, such as for ethernet or a CAN bus, can be designed to be loosely coupled or tightly coupled. For a loosely coupled differential pair, the odd-mode and even-mode impedances are roughly equivalent to the single-ended impedance (impedance of one conductor relative to the common, with the other conductor tied to common).
If you designed the single-ended impedance to be 50 ohm for a loosely coupled differential pair, then the odd and even-mode impedances are both about 50 ohm, and thus the differential impedance is 100 ohm, and the common-mode impedance is about 25 ohm. For Ethernet and CAN bus I have not seen schematics that use a 3 resistor topology for terminating both the differential and common-mode impedances. Is this because the design is for a loosely coupled differential pair such that only 2 resistors are needed to provide a matched termination for both the differential and common-mode signals?
I have never seen the application note for the physical layer of Ethernet or CAN discuss the coupling of the differential pair. For example, for Ethernet, the 50 ohm terminating resistors on each line to the power rail, would provide a differential load impedance of 100 ohm and a common-mode impedance of 25 ohm, which would match a loosely coupled differential pair in which each trace of the pair has a singled-ended impedance of 50 ohm. Similarly, I have seen for CAN that each line is terminated with 60 ohm to the split reference (which has a capacitor to the common).
1 answer
As far as I know from my own experience, CAN and at least 10 Mbit twisted pair ethernet both care about the impedance between the two wires, not so much from each wire to some common ground.
This is quite obvious for ethernet since the twisted pair drives a transformer. There is no third conductor to be a ground reference. Take a look at common cable to carry such signals, like CAT-5. It contains only twisted pairs. There is no shield or common ground.
With typical CAN it's a little different, but the result is similar. CAN does have a common ground, but that's because nodes are directly coupled and need some guarantee that a maximum common mode range not be exceeded. CAN is a linear bus topology, with nodes being short taps anywhere along the line.
When inactive, these receivers have fairly high impedance both between the data lines and from the data lines to ground. When driving, they usually dump a fixed current onto the data lines, again being high impedance. The intent is to preserve the impedance of the bus as a whole.
This type of CAN bus should be implemented as a twisted pair with 120 Ω impedance, and therefore terminated with 120 Ω between the data lines at each end. The bus therefore has 60 Ω impedance at any point.
Both ethernet and CAN can have connections between the data lines and the (local in case of ethernet) to ground. This is usually to reduce the common mode signal with respect to ground, to reduce radiated emissions. Such filters need to be designed carefully to not mess up the differential impedance of the data lines.
Since ethernet is transformer-coupled, this is usually done with a common mode choke integrated into the line side of the transformer. Sometimes, especially with the low speed ethernets, there can be small capacitors to ground on each line after the common mode choke.
All the CAN busses I have seen simply put a 120 Ω resistor between the data lines at each end. I haven't seen one yet where there was any connection to ground. I suppose that's possible, but would again need to be done so as not to mess up the differential impedance between the data lines. It's the differential signal, and therefore the differential impedance, that really matters.
Response to comments
When you design a PCB for CAN or Ethernet, do you design the differential pair as loosely coupled or tightly coupled, or have you done both ways?
I'm not sure what you mean by "loosely" or "tightly" coupled. Coupled to what? In any case, I keep the lines on the PCB as short as possible so that their impedance doesn't matter. Put the transceiver (in the case of CAN) or the transformer (in the case of ethernet) as close as possible to the connector where the lines enter the board. For ethernet, I then additionally try to keep the distance from the inboard side of the transformer to the PHY as short as possible.
Are there any differential signal types that you have found, or have heard of, the need to terminate the common-mode impedance?
I can't think of any off the top of my head where you need to terminate the common mode impedance. I'm not sure that question even makes sense, because how do you know what the common mode impedance is, and therefore what termination to use? The lack of any such spec is a strong clue that such connection to some common is not intended.
And then there is the question of what exactly you would terminate to. In the case of twisted pair ethernet, all you get is the two wires.
You also have to be careful with the common mode voltage range. That can be something like 1000 V for ethernet. In one case I did add some small capacitors from the line side of the ethernet transformer to the local ground of a small device. In this case, the device was powered from POE (power over ethernet). The whole device was floating on the end of the ethernet cable. The purpose was not to "terminate" the lines in any way, but to reduce emissions conducted onto the ethernet cable by the device. The caps provided some impedance for the common-mode choke built into the transformer to work against. It helped with EMI testing a little bit. My main concern was not to get in the way of the ethernet signals, though.
I think (not sure) the reasons to terminate the common-mode are for reducing emissions and also for reducing the amount of common-mode from being converted to a differential signal due to unintentional asymmetries in the differential pair.
A common mode choke, also called a balun, is the usual method to increase the impedance of common mode signals, and make the differential signals more symmetric.
1 comment thread