sli flex cable_analysis
TRANSCRIPT
Why ledge on the CLK for SLI interface?
Compliance Engineering
June 4th , 2008
12/7/2013 1
What is SLI?
12/7/2013 2
Problem Statement
12/7/2013 3
Large clock dispersion causing failure of SLI
Page 4
DDR1-266/333 DQS exampleDQS is single-ended 90-deg phase shifted w.r.to DQ
– DQ/DQS Status--• DQ/DQS crosstalk problem: Multiple curves of DQS coz
of the DQ coupling & PDN around it (even, odd, or stay-H/L)– Sims and Lab data confirms problem
– Fix identified—more spacing for DQS and maybe guardbands—impact to brd routing constraints
DDR266 DDR333Passing
Below Threshold Marginal
12/7/2013
12/7/2013 5Page 5
WR Eye Diagram – DQ58
Vref=1.15v
VIH
VIL
Robust data window:
0.91 + 0.94 = 1.85ns
Jedec
Intel
DQ58: odd DQ58: odd DQ58: even DQ58: even
tDS+tDH=1.23+1.35=2.58ns tDS+tDH=0.91+1.51=2.42nstDS+tDH=1.50+1.13=2.63nstDS+tDH=1.60+0.94=2.54ns
Three sets of curves for DQS showing such
even/odd/stay-H/L cases on DQS signal
V:200mv H:500ps
BGRVP: FabB201#251, HiZo; MCH: A0-SS, 2.30v, 90’c; BIOS:14; DDR333
DIMM0/1= Samsung x8DS / Hyunix x8DS (2.30v, RT); S/W: MARS
DDR1-400 Example showing same three-sets of curves with even/odd/stay-H/L activity on DQ
bus
12/7/2013 6
Rs-DIMM=22-ohms
Ls-DIMM=8nH
SLI interface similarity with DDR1 interface
12/7/2013 7
Clock signal SLI interface behaves similarly to DDR1 DQS that is also single-ended and also phase-shifted w.r.to DQ signals surrounding itBoth x-talk and PDN spreads the waveform as seen in the scope capture: it is a claim
What about the layout?
12/7/2013 8
Looking @ the clock signal: it is also surrounded by DQ signals, depends on the activity on DQ: even, vs. odd, vs. stay-H/L, we see such spread of Clock signal waveform: it is both X-talk and PDN impact.
Claim is: The spread happening on the Clock signal is caused by the even/odd/stay-H/L
activity on DQ signals: How to prove it?
• Sims: Extract using ADS the clock signal along with the DQ signals surrounding it, then simulate the clock phase-shifted and prove the spread of the clock waveforms as in the lab capture
• Lab: using MODS, control the data pattern and active-bits around the clock to see the impact of even/odd/stay-H/L activity on the clock signal
12/7/2013 9
If the Claim is true: what is the proposed soln space?
• Use Guard-band traces around the clock to support reduced x-talk as well as improved PD isolation for the clock
• Require GND-stitching every 1” of the guard-band trace.
12/7/2013 10
Post-Layout Sims for testing the claim that x-talk+PDN are causing
such spread in waveform of CLK for SLI interface
12/7/2013 11
SLI sims on P729-B02
Simulated nets:)P 729-B02
• odd-ports @ connector side, even-ports @ chip side
– MIOB_CLKOUT: P1-P2 – D_13: P3-P4– D_14: P5-P6– D_3: P7-P8– D_4: P9-P10– D_5: P11-P12– D_6: P13-P14– D_7: P15-P16– D_8: P17-P18
Insertion Loss
-3dB @ 970MHz
NEXT w.r.to CLKOUT
-15dB @ 1GHz of NEXT w.r.to CLKOUT
FEXT w.r.to CLKOUT
-17dB for FEXT @ 1GHz
Why NEXT is max for P2-to-P4?
P2
P4
Two vias coupling maximize the x-talk between P2-to-P4We should move the GND via in between signal vias
Why NEXT is max for P2-to-P4?, Cntd.
Two vias coupling maximize the x-talk between P2-to-P4We should move the GND via in between signal vias
Why NEXT is max for P2-to-P16?
Wrong-clock pad causes extra coupling with P16
Why NEXT is max for P2-to-P16?
Wrong-clock pad causes extra coupling with P16
Even on P977 Wrong Clock-pad!!!
Wrong-clock pad causes extra coupling with P16
Summary and Proposed Solution
Moment-Method CorrelationTransistor model
> 250mV improvement
lLab Capture
Guard-band
Adding Guard-band to clock on Flex causes successful operation of the SLI interface