IEEE Solid-State Circuits Magazine - Fall 2015 - 49

20
30
40

ICR (dB)

Insertion Loss (dB)

0
10

Backplane_A
Backplane_B
Backplane_C
IL_KR_mask

50
60
70
80

10
6
8
Frequency (GHz)

14 15

and DFE, this types of channels can
be equalized without much trouble. In
the mask, the curve turns to a steeper
drop beyond Nyquist. The steep drop,
mainly from the nonideal impedance
discontinuity such as vias, connectors, copper surface roughness, etc,
means the channel can be band limited beyond that point, such as shown
in the third channel-Backplane_C. It
shows a large bumpy behavior that
is worse than the mask. This is the
so called legacy channel, which was
designed for a much lower speed (in
this case, 3 Gb/s), and therefore the
channel has a lot of impedance discontinuity due to the nonback-drilled
vias, among other factors. In this case,
DFE should be used [4] more than the
analog CTLE as the former can resolve
the channel reflection within a certain
reach while the latter does not.
Figure 4 shows the ICR, where
ICR = -IL + power sum crosstalk
(PSXT), all three terms in dB. Although
ICR at the Nyquist frequency is often

40
Return Loss (dB)

30
25
20
15
10
5
0
10-2

10-1
100
Frequency (GHz)

Figure 5: The 10GBase-KR RL and mask.

109
Frequency (Hz)

I/O pads, including transmit/receive
self-capacitance, electrostatic discharge, and wire capacitance. These
capacitive loads reduce the effective
impedance at high frequencies and
can cause reflections. One common
practice to reduce the capacitive
effect, other than optimizing the
device layout itself, is to use T-coil
or any on-chip inductor to equivalently resonate out the capacitance.
Figure 6 shows the ILD. The greater the impedance mismatch between
components, the higher the mismatch
losses. What makes it even more interesting is that, at some frequencies,
these signal reflections can add in
phase, and, at other frequencies, they
can add out of phase. Therefore, these
additional losses are not uniform, and
will vary. ILD is defined as the difference between the actual IL as measured on a channel -IL(f), and the IL
as determined by adding the component losses (fitted IL - IL fitted (f )) . In
other words,

cited for a system, it is actually a
function of frequency. To smooth
ICR(f), sometimes the fitted ICR is
preferred, as shown in the figure. The
source of the crosstalk can be from
board traces, vias, connectors, etc.
Since the crosstalk incident varies
among channels due to the different
routes in different systems, it is too
costly to build the on-chip cancellers.
Instead, most of the transceivers simply treat them as residual uncompensated noise.
Figure 5 shows RL. The low frequency loss is mainly determined by
the difference of the dc impedance
values for the on-chip termination
resistors and the board impedance.
It is worth noting that the trend
of the board impedance is lower
(< 90 X differential) from some system houses for higher-speed operations (25 Gb/s and beyond) because
it reduces the channel loss. The
high-frequency value is determined
by the parasitic capacitance of the

Backplane_A
Backplane_B
Backplane_C
RL_KR_mask

ICR_Backplane_A
ICRfit_Backplane_A
ICR_KR_mask

Figure 4: The 10GBase-KR ICR and mask.

Figure 3: 10GBase-KR IL and mask.

101

Insertion Loss Deviation (dB)

60
55
50
45
40
35
30
25
20
15
108

0
-5

Backplane_A
Backplane_B

-10
1

Backplane_C
ILD_KR_mask

3
4
Frequency (GHz)

Figure 6: The 10GBase-KR ILD mask.

IEEE SOLID-STATE CIRCUITS MAGAZINE

fa l l 2 0 15

Table of Contents for the Digital Edition of IEEE Solid-State Circuits Magazine - Fall 2015