Design of De-Emphasis and Equalization Circuits for Gigabit Serial ...
Design of De-Emphasis and Equalization Circuits for Gigabit Serial ...
Design of De-Emphasis and Equalization Circuits for Gigabit Serial ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong><strong>De</strong>sign</strong> <strong>of</strong> <strong>De</strong>-<strong>Emphasis</strong> <strong>and</strong><br />
<strong>Equalization</strong> <strong>Circuits</strong> <strong>for</strong> <strong>Gigabit</strong> <strong>Serial</strong><br />
Interconnects<br />
Girish Ramesh<br />
Ans<strong>of</strong>t Corporation
Agenda<br />
� High Speed Interconnects<br />
� Driving <strong>for</strong>ces<br />
� Challenges<br />
� <strong>Equalization</strong> Techniques<br />
� <strong>De</strong>-emphasis Circuitry compliant with FB-DIMM<br />
st<strong>and</strong>ard<br />
� Circuit schematics<br />
� Nexxim Simulation Results<br />
� Example <strong>of</strong> Passive <strong>Equalization</strong> at Receive side<br />
� Example <strong>of</strong> combination <strong>of</strong> transmit <strong>and</strong> receive<br />
equalization
Why High Speed <strong>Serial</strong><br />
Interconnects ?<br />
� Modern systems like servers, storage, PCs,<br />
telecom systems have necessitated large<br />
volume <strong>of</strong> data transmission<br />
� Shared parallel buses get denser <strong>and</strong> board<br />
design becomes complicated <strong>and</strong> costly<br />
� Solution is to move to gigabit serial links<br />
with more complicated IC design since<br />
Silicon manufacturing cost is shrinking
Challenges <strong>of</strong> High Speed <strong>Serial</strong><br />
� Signal Integrity<br />
Transmission<br />
� Channel Loss hampers board design<br />
� Skin Loss ∝ √(freq)<br />
� Dielectric Loss freq<br />
� Impedance Mismatch<br />
� Lossy Connectors<br />
� Verification <strong>and</strong> testing <strong>of</strong> high speed designs<br />
� Inaccurate modeling <strong>and</strong> simulation <strong>of</strong><br />
channel, vias, interconnect<br />
� Accurate transistor models<br />
� New Simulation Tools <strong>and</strong> Diagnostics<br />
∝
FR4 loss<br />
•Channel loss per<strong>for</strong>ms low pass<br />
filtering <strong>of</strong> input signal<br />
•High Frequency content <strong>and</strong> Low<br />
frequency content must be made<br />
equal<br />
•Boost High Frequency content or<br />
decrease low frequency content
Potential Solutions<br />
<strong>Equalization</strong> must be employed to deal with<br />
channel loss<br />
� Transmit <strong>Equalization</strong><br />
� Pre-<strong>Emphasis</strong>: Increase in voltage swing <strong>for</strong><br />
transition bits (Boost High Frequency content)<br />
� <strong>De</strong>-<strong>Emphasis</strong>: Reduce voltage swing <strong>for</strong> nontransition<br />
bits (Attenuate Low Frequency content)<br />
� Receive <strong>Equalization</strong><br />
� Analog Techniques: High Pass Filter to reverse<br />
channel effect<br />
� Digital Techniques: <strong>De</strong>cision Feedback<br />
<strong>Equalization</strong>
Transmit vs. Receive<br />
<strong>Equalization</strong><br />
� Which method to adopt?<br />
� Pre-<strong>Emphasis</strong>/<strong>De</strong>-<strong>Emphasis</strong> used if noise is a<br />
big factor, receive equalization amplifies<br />
transmitted noise<br />
� Receive <strong>Equalization</strong> used if cross talk/radiated<br />
emissions is a factor<br />
� Receive <strong>Equalization</strong> can be implemented with<br />
a passive network without a system clock<br />
� Many systems adopt a combination <strong>of</strong> the<br />
two schemes
Signaling Features<br />
FB-DIMM St<strong>and</strong>ard<br />
� DC coupled output<br />
� Transmitter with de-emphasis (3.5dB &<br />
6dB)<br />
� Reference clock at 1/24 <strong>of</strong> data rate<br />
� Data transmission at 3.2 Gb/s, 4.0 Gb/s <strong>and</strong><br />
4.8 Gb/s<br />
� 50 Ω single ended termination
<strong>De</strong>-<strong>Emphasis</strong> Architecture<br />
Use edge detection circuitry to detect transition bits<br />
� <strong>De</strong>lay data by one clock cycle using a positive edge triggered flipflop.<br />
� XOR data with flip-flop output to generate a pulse (width=clock<br />
period) <strong>for</strong> every transition.<br />
� Use this pulse to control a high speed current mirror that changes<br />
the tail current <strong>of</strong> the output driver to reduce voltage <strong>of</strong> non-transition<br />
bits.<br />
� Block level implementation<br />
IN<br />
D<br />
Driver RX<br />
D FlipFlop<br />
CK<br />
XOR
Control Wave<strong>for</strong>ms<br />
� Timing Diagram to show signals that used by the Edge <strong>De</strong>tection<br />
Circuit<br />
� Output <strong>of</strong> XOR gate is high after every transition
Transmitter Top Level Implementation<br />
� Adjustable <strong>De</strong>-emphasis through Vcontrol1,2,3 (digital control)<br />
� Four levels <strong>of</strong> <strong>De</strong>-emphasis possible 6.5dB, 3.5dB, 1.75 dB <strong>and</strong> 0dB<br />
� All <strong>Circuits</strong> realized using UMC’s 0.13um CMOS process<br />
� <strong>Circuits</strong> designed using Ans<strong>of</strong>t <strong><strong>De</strong>sign</strong>erRF <strong>and</strong> Nexxim<br />
vee<br />
i400u5<br />
i400u6<br />
i800u3<br />
i400u5<br />
i400u6<br />
i400u7<br />
i800u4<br />
i800u3<br />
Vss<br />
Vdd<br />
U3<br />
CurrentSource8<br />
Vdd<br />
DIS<br />
vee vee<br />
ckl<br />
spinh<br />
spinl<br />
Vdd<br />
Vdd<br />
U54<br />
Driver5<br />
vee<br />
i400u4 i400u2<br />
i20m<br />
i800u1 i800u1<br />
i400u1 i400u1<br />
i400u2 i400u2<br />
i400u3 i400u3<br />
vee<br />
i400u4 i400u4<br />
i800u1 i800u1<br />
i800u2 i800u2 Vdd<br />
ckh<br />
dl<br />
dh<br />
dh<br />
dl<br />
ckh<br />
ckl<br />
Vss<br />
vee<br />
Vdd<br />
i400u1<br />
i400u2<br />
inh<br />
U31<br />
DFlipflop4<br />
i400u1 i400u2<br />
qh<br />
ql<br />
Vdd<br />
i20m<br />
inl outl<br />
Vss<br />
Output Driver<br />
D Flip-flop XOR Gate<br />
outh<br />
ah<br />
al<br />
bh<br />
bl<br />
i400u3<br />
i400u<br />
XORgate6<br />
Vss<br />
vee<br />
outh<br />
Vdd<br />
Vdd<br />
U6<br />
outh<br />
outl<br />
outl<br />
inh<br />
inl<br />
inh<br />
inl<br />
i400u5 i400u2<br />
i800u2 i800u1<br />
inh<br />
inl<br />
i400u6 i400u2<br />
i800u3 i800u1<br />
Vdd<br />
U21<br />
<strong>De</strong>mphcurrsrc4<br />
Vdd Vss<br />
Vdd Vss<br />
vee<br />
Vdd Vss<br />
vee<br />
Adjustable Current<br />
Source<br />
Idemph<br />
DIS<br />
U65<br />
<strong>De</strong>mphcurrsrc4<br />
Idemph<br />
DIS<br />
U75<br />
<strong>De</strong>mphcurrsrc4<br />
Idemph<br />
DIS<br />
VCONTROL1<br />
VCONTROL2<br />
VCONTROL3<br />
User<br />
Programmed<br />
<strong>De</strong>-emphasis<br />
(0dB, 1.8dB,<br />
3.5dB,6.2dB)
�CML Output Driver<br />
�50 Ohm load<br />
Vdd<br />
Output Driver<br />
INH<br />
inh M1558<br />
M1559<br />
inl<br />
P_12_MML130E<br />
P_12_MML130E<br />
Vss<br />
L=0.12u<br />
W=80u<br />
M=2<br />
l=6u<br />
w=1.8u<br />
m=63<br />
rnhr_rf<br />
r_zbt_m=0.05k<br />
i20m<br />
outl outh<br />
OUT<br />
Vdd Vdd<br />
Vss Vss<br />
l=6u<br />
w=1.8u<br />
m=63<br />
L=0.12u<br />
W=80u<br />
M=2<br />
rnhr_rf<br />
r_zbt_m=0.05k<br />
INL
dl<br />
dh<br />
ckh ckl<br />
Vss<br />
Flip Flop<br />
� D Flip-Flop realized by two D latches in series using complementary clocks<br />
� Optimized <strong>for</strong> data rates <strong>of</strong> up to 5Gbps<br />
ckl<br />
ckh<br />
U33<br />
Dlatch8<br />
dh<br />
dl<br />
ckh<br />
ckl<br />
Vss<br />
Vdd<br />
i400u<br />
qh<br />
ql<br />
Vdd<br />
i400u1<br />
U32<br />
Dlatch8<br />
i400u2<br />
ckh<br />
ckl<br />
dh<br />
dl<br />
ckh<br />
ckl<br />
Vss<br />
Vdd<br />
i400u<br />
qh<br />
ql<br />
qh<br />
ql
Vss<br />
Vdd<br />
m=30<br />
w=1u<br />
l=10u<br />
ql<br />
D Latch<br />
� Current Steering D latch with positive feedback<br />
Data<br />
Vdd<br />
Vdd<br />
dh<br />
n_bpw_12_rf<br />
n_bpw_12_rf<br />
dl<br />
Clock<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=1<br />
wt=4u<br />
qh<br />
Vdd<br />
Vdd<br />
ckh<br />
n_bpw_12_rf<br />
n_bpw_12_rf<br />
ckl<br />
Vss<br />
Vdd<br />
i400u<br />
n_bpw_12_rf<br />
r_zbt_m=0.336k<br />
rnhr_rf<br />
Vdd Vdd<br />
Vss Vss<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=1<br />
wt=4u<br />
Vss Vss<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=4<br />
wt=4u<br />
Vss<br />
m=30<br />
w=1u<br />
l=10u<br />
r_zbt_m=0.336k<br />
rnhr_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=1<br />
wt=4u<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=4<br />
wt=4u<br />
Vdd<br />
Vss<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=4<br />
wt=4u<br />
Vdd<br />
Vss<br />
Vss<br />
Vss<br />
Vdd<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=4<br />
wt=4u<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=4<br />
wt=4u
XOR Gate<br />
� Gilbert Cell Variation<br />
bh<br />
ah al<br />
Vdd<br />
outh<br />
outl<br />
i400u<br />
Vss<br />
bl<br />
l=10u<br />
w=1u<br />
m=24<br />
rnhr_rf<br />
r_zbt_m=0.42k<br />
l=10u<br />
w=1u<br />
m=24<br />
rnhr_rf<br />
r_zbt_m=0.42k<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=2<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=2<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=2<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=2<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=2<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=2<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=4<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=1<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=1<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=1<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=1<br />
wt=4u<br />
n_bpw_12_rf<br />
nf=4<br />
lf=0.12u<br />
wf=1u<br />
M=1<br />
wt=4u<br />
Vdd Vdd<br />
Vdd Vdd<br />
Vdd<br />
Vdd<br />
Vdd<br />
Vdd<br />
Vdd<br />
Vdd<br />
Vss<br />
Vss<br />
Vss Vss<br />
Vss Vss<br />
Vss Vss<br />
Vss Vss<br />
Vdd<br />
Vss<br />
Vdd<br />
Vss<br />
Vdd<br />
Vss<br />
Vdd<br />
Vss<br />
A<br />
B
High Speed Current mirror switch<br />
� Changes bias current in driver to incorporate de-emphasis<br />
Vss<br />
i400u2<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=1<br />
wt=4u<br />
inl<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=1<br />
wt=4u<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=1<br />
wt=4u<br />
Vdd<br />
inh<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=1<br />
wt=4u<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=1<br />
wt=4u<br />
M1348<br />
P_12_MML130E<br />
L=0.12u<br />
W=4u<br />
M=6<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=1<br />
wt=4u<br />
i800u1<br />
M1349<br />
P_12_MML130E<br />
L=0.12u<br />
W=4u<br />
M=6<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=1<br />
wt=4u<br />
Vdd<br />
Vss<br />
Mirror turns on to pull current<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=11<br />
wt=4u<br />
Idemph<br />
n_bpw_12_rf<br />
lf=0.12u<br />
wf=1u<br />
nf=4<br />
M=5<br />
wt=4u<br />
DIS
Receiver Input Stage<br />
� CML input stage with internal 50 Ohm resistors<br />
Vss<br />
inh M1633<br />
M1634<br />
inl<br />
P_12_MML130E<br />
P_12_MML130E<br />
Vdd<br />
l=6u<br />
w=1.8u<br />
m=63<br />
rnhr_rf<br />
r_zbt_m=0.05k<br />
L=0.12u<br />
W=3u<br />
M=4<br />
l=6u<br />
w=1.8u<br />
m=15<br />
rnhr_rf<br />
r_zbt_m=0.211k<br />
1m<br />
outl outh<br />
l=6u<br />
w=1.8u<br />
m=15<br />
L=0.12u<br />
W=3u<br />
M=4<br />
Vdd<br />
Vdd<br />
Vss Vdd<br />
Vdd<br />
Vss Vss<br />
Vss<br />
rnhr_rf<br />
r_zbt_m=0.211k<br />
l=6u<br />
w=1.8u<br />
m=63<br />
rnhr_rf<br />
r_zbt_m=0.05k
� <strong>De</strong>-emphasis disabled<br />
Output without <strong>De</strong>-emphasis
Output with <strong>De</strong>-emphasis<br />
� Every transition bit is 6.5dB above subsequent bits
Memory<br />
Bridge<br />
FB-DIMM Configurations<br />
Read Data<br />
(Primary North)<br />
14<br />
10<br />
Write Data<br />
(Primary South)<br />
DRAM<br />
DRAM<br />
DRAM<br />
DRAM<br />
AMB<br />
DRAM<br />
DRAM<br />
DRAM<br />
DRAM<br />
Secondary<br />
North<br />
Secondary<br />
South<br />
DRAM<br />
DRAM<br />
DRAM<br />
DRAM<br />
AMB<br />
DRAM<br />
DRAM<br />
DRAM<br />
DRAM<br />
*Third party marks <strong>and</strong> br<strong>and</strong>s are the property <strong>of</strong> their respective owners<br />
� The FB-DIMM architecture replaces the existing parallel<br />
memory bus with a point-to-point serial signaling similar to<br />
PCI Express. The DRAM is behind a buffer (Advance<br />
Memory Buffer - AMB). The AMB is a high-speed (3.2-<br />
5GB/s) unidirectional link with the host memory bridge. Ten<br />
channels are provide <strong>for</strong> writes on the South Bridge <strong>and</strong> 14<br />
<strong>for</strong> reads on the North Bridge. The AMB also has a<br />
secondary channel <strong>for</strong> read/writes that acts a pass-through<br />
to the next module. The secondary channel is an additional<br />
point-to-point serial link. This results in a total <strong>of</strong> 48 highspeed<br />
serial channels.<br />
� The primary channels are routed on the top side <strong>of</strong> the<br />
memory module using microstrip. Trace lengths vary from<br />
approximately 2-7in. The secondary channels are routed<br />
on the backside <strong>of</strong> the module <strong>and</strong> transition first into a<br />
stripline layer <strong>and</strong> then to the top <strong>of</strong> the module where they<br />
interface with the AMB BGA.<br />
� For our investigation we will focus on the first channel <strong>of</strong> the<br />
Primary South <strong>and</strong> Secondary South channels. We will<br />
show the end-to-end channels <strong>for</strong> two configurations: host<br />
memory bridge to AMB <strong>and</strong> AMB to AMB.<br />
� Passive channels will be extracted <strong>for</strong> the packages, the<br />
motherboard, the FB-DIMM connector, <strong>and</strong> module routing.<br />
� Transistor level active drivers were developed with deemphasis<br />
to meet the FB-DIMM spec.<br />
� The following slide shows a block diagram <strong>of</strong> the system to<br />
be extracted, simulated, <strong>and</strong> validated along with the<br />
applicable Ans<strong>of</strong>t simulation tools.
Vss<br />
Package<br />
inh M1558<br />
M1559<br />
inl<br />
P_12_MML130E<br />
P_12_MML130E<br />
L=0.12u<br />
W=80u<br />
M=2<br />
rnhr_rf<br />
r_zbt_m=0.05k<br />
l=6u<br />
w=1.8u<br />
m=63<br />
i20m<br />
outl outh<br />
Vdd Vdd<br />
Vss Vss<br />
l=6u<br />
w=1.8u<br />
m=63<br />
L=0.12u<br />
W=80u<br />
M=2<br />
rnhr_rf<br />
r_zbt_m=0.05k<br />
+<br />
-<br />
Model Extraction<br />
Backplane Connector Daughter Card<br />
MS SL SL<br />
Via Via Via<br />
+<br />
-<br />
Package
<strong>De</strong>-emphasis at work!<br />
� FB-DIMM board (Primary South Side)
<strong>De</strong>-emphasis at work!<br />
� Xilinx Virtex II Pro test board
<strong>De</strong>-emphasis at work!<br />
� Two Xilinx Virtex II Pro test boards in series<br />
� Increased <strong>De</strong>-emphasis to improve per<strong>for</strong>mance
Receive <strong>Equalization</strong><br />
� Passive Network<br />
� Adaptive <strong>Equalization</strong> can incorporated<br />
Incorporating passive network into reciever to maintain 50 Ohm matching<br />
Simple Resistive Divider with Frequency <strong>De</strong>pendent Gain<br />
50 Out<br />
50<br />
IN+ IN-<br />
IN Out<br />
0<br />
50<br />
0
Receive Passive <strong>Equalization</strong><br />
� Xilinx Virtex II Pro test board<br />
spinh<br />
spinl<br />
50<br />
1<br />
50 25<br />
2<br />
4<br />
ref<br />
0<br />
3<br />
12p<br />
12p<br />
25<br />
vee<br />
DIS<br />
Vdd<br />
Vss<br />
U2<br />
rxcurrsrc13<br />
vee<br />
1m<br />
inh<br />
inl outl<br />
Vss<br />
U1<br />
rxip13<br />
Vdd<br />
1m<br />
vee<br />
outh
Combination <strong>of</strong> <strong>De</strong>-emphasis <strong>and</strong> Receive<br />
<strong>Equalization</strong><br />
� Two Xilinx Virtex II Pro test boards in series<br />
� Some Channels require higher order high pass<br />
filter <strong>for</strong> greater eye opening
Summary<br />
� Outlined challenges faced in high speed serial interconnects<br />
� Discussed <strong>Equalization</strong> techniques<br />
� Presented a <strong>De</strong>-emphasis circuit that is compliant with FB-DIMM<br />
st<strong>and</strong>ard<br />
� Simulations (Nexxim) show the benefits <strong>of</strong> using <strong>De</strong>-emphasis<br />
� Example <strong>of</strong> a simple Receive equalization technique<br />
� Example <strong>of</strong> a combination <strong>of</strong> <strong>De</strong>-emphasis <strong>and</strong> Receive<br />
equalization<br />
� <strong>De</strong>monstrated diminishing returns from simple low order<br />
equalization circuits <strong>for</strong> a channel with higher order properties<br />
� Ans<strong>of</strong>t has come out with a well integrated design kit that includes<br />
� Channel Modeling: Trace/Backplane, Vias, Connectors<br />
� Transceiver design with equalization techniques to enhance<br />
high speed data transmission over Copper<br />
� Work with board designers to solve their signal integrity issues to<br />
ensure high signal quality <strong>for</strong> data recovery