One for the Ethernet grey beards - magnetics

[TomS_]

Hi all. I've been looking into the parts I need to build an Ethernet adapter for a Motorola 68k system I have been working on, and I'm up to the part where I need to find suitable magnetics.

Having some past experience with the Am79C90 Ethernet controller I'm going to use that. This requires an external serialiser for which I am looking to use an LXT901.

The datasheet for the LXT901 says it needs a 1:1 transformer on the RX side, and a 1:√2 transformer on the TX side.

The interesting thing about this is I have a device that also uses an LXT907 (very similar to the LXT901 with the same requirements), but this is using 1:1 transformers on both the TX and RX.

My question is, how much do the transformer winding ratios matter? And if the datasheet says it needs 1:√2, what will be the side effects of using 1:1 in its place?

Basically, I've found somewhere to buy a small quantity of each of the above chips plus the same model magnetics as used in this other device (from an eBay seller that I have had good success with in the past), but I don't want to buy the magnetics if they simply aren't going to work.

Thanks!

[Benta]

Easiest is to just buy an RJ PCB connector with integrated magnetics. No need to fiddle with the transformers yourself.

[TomS_]

I am aware of jacks with built-in magnetics, but that isnt really answering the question of what effect a 1:1 transformer has when the datasheet calls for 1:√2.

[T3sl4co1l]

Have seen that ratio listed in the usual places (DigiKey etc.), though I've never had a need to look at them and see what actual availability or other specs are like.  Assuming they're available -- what's wrong with that?

Otherwise, if you don't mind the reduced maximum range, I would expect it's fine (going with 1:1)?  Says the driver has extremely low output resistance, as CMOS output pins go (5Ω typ?), so put in whatever series termination resistors (half and half each, mind) make up the difference for your medium, and that should do.

Not sure why they specify 24.9Ω resistors when the driver should be accounting for a couple of those ohms.  And arguably it's better to be on the low side, helps account for HF losses in the media.  I'd probably go with 22Ω, myself?

Or why an "irrational transformer" caught on, an obvious impossibility; a 5:7 or whatever ratio is well defined and only off by 2% (51Ω primary side).

Note that they don't show CMCs on the transformers; I wouldn't be surprised if this is easy to get away with at a mere 10Mb.  Shouldn't hurt to include it (and the grounded (at AC, usually through a 75R + 1.5nF "Howard Johnson termination") tap), so aside from the ratio, off-the-shelf parts shouldn't have any problems.

Mind, if you're going for absolutely period-correct capability, and you're going to be making design changes like these -- you'll probably have to get the contemporary standards, or diagnostic tools or whatever, to check it all out.  If you're only interfacing with modern hardware (sensitive receiver, lots of line compensation tricks), and don't need long runs, I doubt you'll notice anything different, hell, run it over some dumbass audio transformers for almost all that matters, 10BASE-T is terribly tolerant of poor conditions.

Tim

[TomS_]

Thanks Tim.

I suppose I am curious why a big name brand networking company (in this case, Cisco) would go against the datasheet recommendations. Analog is not my strong point, and this falls well within that realm. 

But its also highly likely that "they know something that I dont", or had some people with the brains and equipment to figure out what would work. There are a bunch of resistors and capacitors between the transceiver and the transformers, so maybe they are doing something there to "fix things up", although the topology and values look very much the same as the datasheet examples but with a couple of capacitors sprinkled around.

There is also a CMC between the transformers and the jack.

The device is an old ethernet router, its probably something that would be somewhat unlikely to hang off the end of a very long run of network cable. It has both 8P8C 10bT and AUI connectors, so presumably if it was going to sit on the end of a long cable run you would have used an optical AUI adapter instead of copper.

But if what youre saying is correct - that you just lose a little bit of range by using 1:1 on the TX side - and that 10bT is quite tolerant of poor conditions - then maybe they determined that it would be "good enough".

I would simply be inclined to use the same part because it is in a known working configuration (and although it is a surface mount part, it has 0.1" pin pitch which is relatively friendly).

EDIT: incorrect diagram/schematic, see later post

[capt bullshot]

That 1:1.414 TX transformer ratio has been quite common in ye olde days for 10MBit Ethernet. 100MBit continued to use it for some time. I guess it's been appropriate for older technology that could easier made for lower drive impedance than higher amplitude (what would be required for 1:1 ration). The 1:1.141 would match the lower driver impedance to the cable (and give some higher amplitude). Most probably this evolved to a de facto standard for PHY and magnetics manufacturers.

At some point Auto-MDIX appeared. This requires the RX / TX pins swappable in the PHY chip, so it wouldn't work with 1:1 / 1.414 transformers. This was the point when the 1:1 RX / TX transformers took over.
 If you match an old 1.414 ratio PHY to a modern 1:1 transformer, you'd lose some amplitude, but IMO that wouldn't matter too much.

[TomS_]

Ah, good point about Auto-MDIX. This device doesnt have it, so that makes a lot of sense. 

[T3sl4co1l]

Oh neat, the transformer datasheet is still around too,
https://static6.arrow.com/aropdfconversion/1a4bfe6a510553961104554f1f42ead995b9c7a5/sf1012.pdf
that's a fancy filter!

Indeed, it seems they didn't care much about source mismatch, it's off by 2:1 even.

The CMC is also around,
https://www.pca.com/datasheets/ept7003.pdf
which is evidently 4 lines 1 core.

The direct connections to the pairs look... just wrong.  They could be using it as a balun, but the impedance is still wrong (with equal windings, that's a 4:1 impedance ratio), and it would have 100% crosstalk between RX/TX being on the same core.  Anyway the inductance is too small (5uH).

I can't reason out any obvious value to the circuit as shown... maybe the parts aren't what they say they are, maybe it's worth running a sim.

Doesn't seem to be any ESD protection ("Y" cap and termination resistor) (unless that's on the other side of the connector... which would make about as much sense).

Tim

[TomS_]

Nothing on the underside of the board except what look like test points. Its a single sided load.

Probably worth ignoring the original diagrams. I definitely had something wrong with them. Looking into it a bit more, and looking visually from different angles and backlighting the board (yay for lack of ground plane) has revealed there definitely are no direct connections between the jack and the transformer - everything goes through the CMC first. It looks much more like the attached diagram where the middle taps of the line side transformers are NC.

I had tried measuring resistances between pads with my multimeter, and it seemed like it was connected that way given the results I was getting.

I believe the operation of this was meant to be in half duplex (although the ethernet controller in the CPU is capable of full duplex), so could that explain why they have gone with a single core for the CMC?

[T3sl4co1l]

Ah! That looks much better 

The outboard CMC, just takes the edge off (or, quite a bit of "edge" given it's a kohm or so; though, comparatively speaking, probably the internal CMCs are a similar value), and acts on all pairs because, well, they're all cabled together, won't make much difference to separate them.  Does invite some opportunity for crosstalk (depending on how the four lines are arranged internally).  It may be more effective (as well as the internal ones) with the CT bypassed (R+C to GND), and it still seems suspicious that they don't have that (because ESD). :shruggie:

And the filtering in that transformer, ought to do a fine job of dealing with interference, even if the PHY doesn't.  Which I'm guessing not so well, being relatively primitive; hence why it was so well filtered in the first place.  Whereas modern PHYs likely have filtering built into their analog or converter section (a necessity for 100/1G operation, as applicable), so, unnecessary externally.

Tim

[bson]

Why make life so hard for yourself?  The CS8900A for example is a much simpler device to work with than Lance.  It has on-board buffer memory, so doesn't require access to external memory, has the PHY built in, and is super easy to interface and program.  I worked at Sun Microsystems in the 90s (fixing an endless parade of functional and reliability problems with AT&T's streams-based TCP implementation as this was a major barrier to SunOS 5 adoption; I must have fixed 200+ TCP and socket related bugs and issues between Solaris 2.5 and 2.6) - and the Lance controllers presented tons of headaches.  To be honest I don't remember what the problems with Lance were, specifically (it's been 25+ years after all), but there were some serious ones.  Lance would be my absolutely last choice!

[TomS_]

Well I'm not going to lie, I've spent basically zero time looking into other options.

I'm choosing the LANCE because I have experience with it from a previous project, and I know that I can get them in PLCC packages so they can be socketed easily.

That's about all I can say. 🙂