The Fast-Links Project: Recent Results
Recent Fast-Links Results
This page describes some key results from our testing of the first
test chip. Many of the results will also be described in our paper at
Hot Interconnects '97.
A series of experiments were performed to demonstrate the
effectiveness of equalization in overcoming the equalization of the
transmission line. The measurements shown here were all performed
with the transmitter driving a short PC-board transmission line, 1
meter of 30 gauge twisted pair, another short PC-board line, and a
This first pair of waveforms result from isolated 1 and 0 bits
superimposed on each other to create an ``eye'' diagram. Without the
equalizer, there is no eye at all: the level of an isolated 1 bit is
below that of the isolated zero bit. With equalization, the overall
level signal level is reduced, but the data eye is clearly present.
This next pair of waveforms show a pseudorandom bit pattern, again
with equalization both off and on. Without equalization, there is no
data eye at all. Equalization recovers a recognizable data eye, but
attenuation in this 1m line is high enough that this probably
represents the limit of the technique at its present state of
For these experiments we set all of the weights (except the first) in
the filters to the same value; with only observation of eye patterns,
there was no improvement in the pseudo-random sequence eye patterns
with further adjustment of weights for non-adjacent bits. We
speculate that a 2-tap transition filter may be sufficient to obtain
most of the benefits of this equalization technique, allowing much
simpler transmitter circuitry than we used in the test chip.
In the isolated-bit waveforms of the previous figures, various effects
that arise from transmitter imperfections and package parasitics can
be seen. The ``ripple'' in an otherwise quiet series of 0s, for
example, is due to a slight mis-adjustment of On and Off clock
overlap. Some overshoot is evident following an isolated bit (this
overshoot contributes much of the amplitude noise in the
pseudo-random bit patterns); this is caused by the LC tank circuits
formed by the transmitter's output capacitance and bond-wire
inductance, as confirmed by SPICE simulations. This effect is
probably unavoidable in conventional packaging, but can be overcome
almost entirely by terminating the transmitter as well as the
receiver. Our future signaling chips will have both ends of the
transmission line terminated. A second phenomenon that contributes
significant noise is reflection off the capacitance at the receiver.
These reflections are reflected again at the (high-impedance)
transmitter, causing intersymbol interference at the receiver.
Simulations suggest that on-chip transmitter termination will very
effectively remove these reflections.
Two design ``features'' of the transmitter also contributed to signaling
problems. First, the transmitter has a clocked (On-clock) current
tail transistor (this circuit structure was the only one we could find
at the time that would deliver the necessary switching speed). When
the Off-clock turns on, well in advance of the time when the DAC cell
is supposed to transmit its data bit, charge is injected into the
output from the common source node. In our signaling system, this
effects introduces an unintentional signal onto the line at about 15%
of the amplitude of the actual signal. Second, and more seriously,
the gating function within the filter changes the strength by
completely turning off one or two of the three drivers in a DAC. The
result is that, when equalization is turned on, the output signal is
not truly differential. This has the unfortunate effect of sending
a common-mode signal down the line, where various non-linearities
transform the common-mode signal into a differential (noise) signal.
A new transmitter design will take care of both of these problems.
Because of the previously described problems with
package-parasitic-induced noise, we were unable to test the receiver
at 4Gb/s. We were, however, able to fully test its ability to acquire
the framing sequence and to operate the LFSR checker at 2.5Gb/s.
Because of various problems with the data transmitter, described
above, bit-error rates were unacceptably high (about 10-4).
Simulation suggests that the BER can be brought to acceptable levels
with a redesigned transmitter.
About the figures on this page
The waveforms shown on this page were captured with a Tektronix 11801A
sampling oscilloscope. For most of these measurements, the 'scope was
set in its infinite-persistence color-grading mode; pixels
representing more common data points are assigned "hotter" colors with
the most common points colored yellow. The oscilloscope screen was
stored in a PPM file using a serial connection and software written
for the purpose. For display on the web, the screen images have been
reduced to half of their original size and converted to JPEG format.
[Fast Links home page]
[MSL home page]
Last modified: Wed Jul 15 10:52:13 EDT 1998