differential stripline question

Oh crap, it's "Nelco" not Melco...

Yup. Just taper in the obvious way. Check that the halfway point is still 100 ohms, just for fun. But at your speeds, a sudden jump wouldn't make much difference.

A tenth of an inch should be fine. 100 ps is a pretty slow edge.

Tapered lines are common in some situations. A slow taper, many many edges long, makes a nice impedance transformer. Since you're not trying to change impedances, your transition could be most any length.

I'm assuming the skinnier traces kill off higher frequency

Yes. Skinny traces like yours will be very lossy for 100 ps edges.

We do 1mm BGAs with 6 mil traces, transitioning to 8 or 10 as soon as we get clear of the balls. Try to keep the fast stuff on layer 1, on the outer rows of balls. But mostly, keep these runs short.

Thicker traces are better, but that implies a thicker dielectric, lower Er, and/or wider differential trace spacings.

Two examples in my collection, both 50 ohms CPW:

You can see the trend. Skinnier traces will be worse.

There's not as much debate as with bypassing, as everything can be measured, and fewer people do this fast stuff. There is some debate about "reference planes" and "return currents", including a great deal of nonsense.

"a great deal of nonsense"

National and Xilinx have appnotes. Look for "equalized lvds receivers" which should lead to some pcb loss info.

How far do those 100 ps diff signals have to go? What's the data rate?

John

Thanks John. What specifically don't you like about the Black Magic books?

If you search the archives you can find several discussions about this...

I think those "Black Magic" books are about 100x better than the "nothing" most people have. John points out, though, that some of what they tell you is based on various assumptions that don't always apply, and you'll sooner or later probably have to figure out what those assumptions were. (The classic example is the whole "a cut in the ground plane forces current to go around the cut!" -- well, yeah, it does, at least at some frequencies, but eventually the capacitance across that cut is high enough that the signal will just skip right through it. Figuring out the frequency at which you have, say, half the current going down each path is not particularly trivial, though. John has a fancy TDR setup where he can test this sort of thing experimentally, though...)

Note that formal education is usually taught the same way -- in physics class you begin with those frictionless pulleys and massless springs, after all, and in electronics class you start with ideal resistors (that are anything but at high enough frequencies or voltages) and Ebers-Moll transistor models (also falls apart at high frequencies). In those cases it's rather more obvious that simplifying assumptions are being made than in the case of Howard's book there... (but maybe not always... plenty of people see no reason you can't use a 1/4W carbon comp. resistor at 1kV assuming the current is small enough, and look at all the people surprised that a 120VAC fuse is not a 120VDC fuse...)

Half of what he says is good, and half is nonsense. If you can tell which is which, you don't need the book.

And 90% of his writing is _really_ annoying.

John

The 100ohm differential impedance is usually not important at all, it's the

Not having to fight the 50ohm SE/100ohm DIFF non-linear trade-off balance can give you a lot extra routing space.

Dave.

Yeah, a pair of traces could transition from closely coupled to not coupled at all, as needed. Maybe A7 is using 4 mil traces so that he can keep them arguably differential, when it might be better to separate them, at least until he's cleared the bga. I think losses are generally lower if the traces are not closely coupled.

But if he's terminating differentially, it's the odd mode impedance (2x, actually) that matters.

Here's a "100 ohm differential" run, actually not very differential.

ftp://66.117.156.8/T860A.jpg

(and, as of this afternoon, it works!)

John