Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Gigabit signals via uncontrolled-impedance connectors

Recently, there has been a lot of discussion of the difficulties in getting
10-gigabit or 2.5-gigabit signals through card-edge connectors and the
like.  I would like to offer a possible way to handle this class of

In August 1998, this same problem came up in an unrelated context, and a
dodge then occurred to me, a way to control the reflections from
uncontrolled-impedance connectors.  This trick actually came from the RF
world, but should work for digital signals as well, although I have not
myself tried it.  This approach parallels the use of special compensation
networks to increase the signalling rate allowed on Fibre Channel 150-ohm
balanced electrical link cables.

Differential logic signals can tolerate some attenuation, if the waveforms
are clean, so it's OK to make the signal paths somewhat lossy, and we can
use this attenuation to swamp any impedance mismatches and variations,
especially if the lossy connector is physically small.

At each connector, make the following changes:  Build a 3-db balanced "pi"
attenuator, split it into two halves (by splitting the two in-line
resistors in half), and connect the halves back together using two pins
from the uncontrolled connector.  This takes a total of six surface-mount
chip resistors, three on each side.  Because the resistors will have values
from a few ohms to a few hundred ohms, carry very little power, and the
tolerance is quite loose, it should be easy to obtain suitable tiny
low-inductance surface-mount components.  Because the attenuator itself can
be made small relative to a wavelength (3 centimeters) at the signaling
rate (10 gigabit), no significant reflections will occur within the pad.
The net result is that very good waveform fidelity is possible, despite the
imperfections of the connector.

The effect of this is to isolate the connector (and thus its impedance
bump) from the rest of the transmission lines, preserving signal
waveshapes, at the expense of some added loss.  Noise from the transmitter
side of the lossy connector will be attenuated in the same proportion as
the digital waveform, so only receiver-side noise can cause a SNR
degradation due to the reduced signal amplitude at the receiver, at least
to the first order.

A 3-db pad with equal in and out impedances will by definition pass 71% of
the input voltage swing, and this should pose no problem.   For instance,
the minimum ECL differential output swing is 600 mV, while the required
minimum input differential swing is 400 mV, which is 400/600= 67%, and so
71% will work.  Using typical values, the output swing is closer to 1,000
mV, and the required input swing is 200 mV, which is 200/1000= 20%, or -14
dB, so one could at least in theory cascade four such 3-db pads in a
laboratory setup, where it's OK to hand-select parts.  In production
designs, one would instead regenerate the signal after each transmission
through the attenuated connector, so that end-to-end losses would not grow
with the number of lossy connectors in series.

Actual resistor values are tabulated, given transmission-line impedance and
desired loss, in various books, such as the old ITT Radio Engineers
Handbook: "Reference Data for Radio Engineers", fifth edition, Sams 1968,
Chapter 10 "Attenuators".  (Later editions don't have nearly as good a
treatment of attenuators as the 1968 edition cited above.)   Table 2 on
page 10-6, covering the "symmetrical-Pi" attenuator design, covers the 3-dB
pad described in the following paragraph.  Note that Table 2 assumes
500-ohm input and output impedances, so the values given in the table have
been divided by ten to yield the following 50-ohm values.

For 50-ohm input and output impedances and 3-decibel loss, the two shunt
resistors (R1 and R2 in Table 2) are both 292.4 ohms, and the four series
resistors (R3 divided by 4) are each 4.40 ohms.  For transmission-line
impedances other than 50 ohms, scale the values just given in linear
proportion to the actual line impedance.  For instance, for 75-ohm
transmission lines, the shunt resistors are (292.4)(75/50)= 438.6 ohms
each, and the series resistors are (4.40)(75/50)= 6.6 ohms each.

Such resistive attenuators can also help to control the destructive connect
and disconnect spikes seen when circuits are opened or closed, possibly
permitting hot insertion and removal without damage to the receiver input
structures, an added attraction.

Other logic families will have their own voltage levels, but in the context
of designing 10GbE, we ought to be able to allow for some attenuation, if
it solves some signal integrity problems.

Although the above description assumes a differential signal, there is no
reason this approach won't work for unbalanced signals as well.  Slightly
more complex attenuators can also compensate for excess stray capacitance
in the connector, if needed, but at the expense of larger added

Likewise, one can make stripline on FR4 better-behaved if one is willing to
suffer some added attenuation.  A little bit of loss can greatly reduce the
effect of reflections, and if one knows in advance the exact path loss to
expect and so can design for it.