JS> Dear SI-land,
JS> I'm pondering the wiring of some >1Ghz differential
JS> ECL, with wire lengths reaching perhaps 10 inches. It has
JS> become clear that this wiring could save alot of length (and
JS> other torture) if vias were used, at least one and possibly
JS> two (per side). These would be signal routing vias, not
JS> pin-escapes neatly located near a discontinuity to a package.
JS> Should I be concerned about the performance of these
JS> vias? Anecdotal history tells me to expect a loading of
JS> perhaps 1pF each, which I would not expect to be an issue. If
JS> each side of the pair passes through a via at approximately
JS> the same point in time, will the issue take care of itself?
Clearly there are a lot of issues with bringing signals through vias,
referencing them to new ground planes, etc. Let me concentrate here
on an issue that's unique to differential signals.
As you noted, it's important to keep the path lengths for the two
signals the same. If the two signals have to travel different
lengths, then you'll be converting differential signals into
common-mode, and vice versa. Having two signals jump down through
some vias and then continue on in the same direction is no problem,
because the path lengths are the same. But if they jump from an x
layer to a y-layer, making a right-angle bend, then problems can
occur.
For example: take this geometry (V = via).
x-layer
2 --------------V
1 ------------V |
| |
| |
| | y-layer
| |
1' 2'
Signal 1 has the "inside track" and a shorter delay than signal 2. When
you get to 1'-2', you'll find that there's now some common-mode component
to the signals, even if you have a pure differential drive at 1-2.
Similarly, if you have some common-mode noise at 1-2, you'll find that
there's now some differential-mode noise at 1'-2'; the bend has "converted"
the common-mode noise (invisible to a good differential receiver)
into something your differential receiver can see.
The effect of the bend is usually small, but as you add more bends to the
signal path, the effect accumulates.
JS> Alternatively, is there a simply way to manage the
JS> "hit" these vias will produce, perhaps by specific geometries
JS> or proximity to reference (gnd)? I can envision quite a 3-d
JS> solver excer- size attempting to use pad sizes versus planes
JS> versus clear- ances etc.
I haven't tried varying the problem geometries systematically, but what
you suggest definitely sounds like the way to go (short of building
quite a lot of test structures.)
I tried a single geometry in our HFSS full-wave solver a while ago.
You could point your browser at
http://www.ansoft.com/~bracken/appnotes/bendline.html
to see the results.
-- J. Eric Bracken, Ph.D. Tel: 1.412.261.3200 x135 R&D Manager, Signal Integrity Products Fax: 1.412.471.9427 Ansoft Corp., Four Station Square, Suite 660 [email protected] Pittsburgh, PA USA 15219-1119 http://www.ansoft.com