Electrical Characterization of the High Speed I/O Data BusUsing Cross Correlation Method

Jimmy Huat Since Huang, Zuri Shaameri Ahmad, Teong Guan Yew'University of Technology Malaysia, Penang, Malaysia

2Intel Microelectronic (M) Sdn Bhd, FTZ, 11900 Bayan Lepas, Penang, Malaysiajimmy.huat.since.huanggintel.com, ahmadzsWyahoo.com, teong.guan.yewWintel.com

AbstractHigh speed data transportation between the CPU and

peripherals on the PC motherboard is needed to support heavydata traffic such as multimedia, games and broadbandnetworks. At multi Gbits/sec high speed, impedance mismatchbetween the CPU and peripherals becomes critical and limitsthe possible maximum throughput. The I/0 transportation buscan be modeled as a linear time invariant system. The outputsignal at the receiver is the convolution function of thetransfer function and transmitter signal. Due to the complexityof the motherboard ingredient, it is desired to model the I/0bus in black box behavior model. Instead of using traditionalpassive measurement method such as Time DomainReflectometry (TDR) and Scattering parameter measurement,cross correlation method is used to find out the impulseresponse transfer function when the I/0 Bus is active. Byusing MATLAB and SPICE tools, the method is simulated tounderstand its accuracy and robustness under noisyenvironment.

IntroductionComputer system organization has three main components,

the CPU, the memory subsystem and the I/0 subsystem. Busis a terminology used to describe the interconnecting betweenthe components in the architecture organization. Physically, abus is a set of wire to send the information from onecomponent to another; the source output the components ontothe bus. The destination component then inputs this data fromthe bus. Due to the increasing complexity of computerarchitecture, the bus system is much more efficient in higherspeed, less power consumption; less space and fewer pinroutes.

Moore's Law drives transistor scaling by 2x for every 21months. Advanced computer system benefits from thetransistor scaling allowing more processing capabilities canbe achieved. Higher data bandwidth is needed to support theincreasing processing power. The performance degrades if thecomputer spends most of its time waiting for the data. Theneeds of data bandwidth is even critical with the introductionof parallel processing, distribution computing system, multicore CPU and more efficient pipeline architecture.

From communication view of point, the transmission buschannel is composed of transmitter block, bus channel blockand receiver channel block as illustrated in Figure 1. Thesource output is x(t), channel impulse response is h(t),receiver output is r(t) and n(t) is Additive White Gaussiannoise. The receiver output is the convolution of h(t) and x(t)plus summation of n(t).

r(t) = x(t) 0 h(t) + n(t)

Figure 1: Transmission line can be modeled intocommunication blocksets

The transmission bus channel behaves as a low passfilter because the channel loss increases with the frequency. Itis due to skin effect loss and dielectric loss. Instead ofamplitude loss, the channel also suffers from phase distortionas illustrated in Figure 2.

.......................................................................

Mani d >

Figure 2. Frequency response of transmission linebehaves like a low pass filter

Equalizer is implemented to improve the signal quality byintroducing the inverse transfer function of the transmissionbus channel so that it can cancel the channel loss effect andretrieve back the original transmitting data. The equalizer,C(f) = H- (f), can be placed at the transmitter, receiver orboth side. The equalizer can be an analog passive high passfilter or discrete FIR filter. No matter which equalizer is used,the H(f) must be characterized to obtain the inverse transferfunction coefficients. .

BackgroundThere are active components and passive components in

I/0 Bus channel. The active elements include I/0 buffers, op-amp, PLL and data pattern. Passive elements include wire,printed circuit board, IC package, connector and others. Atlow speed, the passive elements are just part of the product'spackaging but at higher speed they behave like a transmissionline and affects electrical performance directly.

1-4244-1392-3/07/$25.00 (©2007 IEEE

Figure 3: Interconnect diagram of chip, package andprinted circuit board

Due to the complexity and overheads in generating theimpulse or step response to the transmission bus channelduring active condition, auto and cross correlation method isproposed because the

Rxx(T) E[X(t)X * (t -)], (1)

R. (r) Po(I-I Tr I' (2)Tb

where the input is pseudo random sequence, Tb is thepulse period and Po is the average power of the randomsequence as Figure 4

RX4r> uf8(r)X (3)Given o-2 is the variance of the inputs, if the pseudo

random has a long and non-repetitive sequence1CI / (4)SXx (f) = j{ rim T fR (t1, t1 -r)dt1 }exp(-j2rf-r)dr(

T< 0 TTT2 (5)

Sxx (f) RXX (r) exp(-j2fr)dr (6)

For a impulse response of R(r) = o2-(i-)2

SX ) x (7)By using additive White Gaussian Pseudo Random

Sequence as the inputs,

RYX ( E[Y(t)X(t -)]

=E[ fh(A)X(t - A)dAX(t -T)(8

= H(f)Sx (f)= H(f)o=2 (13)Data packet pattern behaves as a pseudo random sequence

generator (PRSG). It itself is the inputs to this equation.

MethodologyThere are some complex motherboard components that

require detail electrical modeling such as package, breakoutmotherboard trace, vias, motherboard trace and connector.These interconnect segments are modeled in 2D or 3D modelsusing appropriate simulation tools. The board traces aremodeled using Ansoft Q2D to obtain the transmission lineRLGC model. AnsoftLinkTM is a tool used to extract 3Dmodel with actual dimension from package layout. The same3D model is then imported, modeled and analyzed in AnsoftQ3D to generate parasitic RLC values.

VI TFigure 5. Physical dimension ofQ3D interconnect model

(wirebond, via model).The I/0 bus system as Figure 6 is simulated to obtain the

transient response using step or impulse function. Then crosscorrelation method is used and the output is compared withthe traditional method. It succeeds if both them match. Then itwill be tested again under noisy environment. Figure 7illustrates the methodology chart.

Figure 4: AutoCorrelation function of Pseudo randomsequence

By putting the expectation operator within the integral, thecross correlation function obtained is

00

R,x(r) = fh(L)E[X(t - A)X * (t - r)d2]

Since X(t) is stationary randomautocorrelation ofX(t) derived,

E[X(t1)X * (t1 - (r - 2))] = Rxx ( - A)

R "(rfh(2)Rxj 2dRX(/T) = JhX) , (-r - A)dA-00

h(r) * Rxx (r)The cross power spectrum is

Syx(f) fRyx(r) exp(-j27rfr)dr

Syx(f)= fJh(A)Rxx(r -A)exp(-j2lfr)d2drO) O)11

(9)

process, the

(10)

Figure 6.0 Full EM Modeling Flatform Data Bus topologysystem

(1 1)

(12)

1-4244-1392-3/07/$25.00 (©2007 IEEE

Figure 7. Methodologyflow chart

Simulation from single transmission modelCross correlation method was used on simple transmission

line model. In HSPICE, A pulse wave signal was injected atthe transmitter to obtain impulse response. The HSPICE resultwas used as the bench mark for comparison.

Figure 8. 1HSPICE impulse response analysis. Red line isthe pulse wave at transmitter while the magenta line is outputat receiver

Figure 9. Rxy from Cross Correlation method by running10, 000 sample random data

From the pulse response result, the overshoot amplitudewas 0.784 and the delay time was 3.4ns. Comparing withcross correlation method in Figure 9 by using 10,000 samplesin HSPICE, the overshoot was 0.761, which was close to theimpulse response analysis. Both of them had same delay time,3.4ns, and very similar waveform shape. Thus it wasconcluded that cross correlation method using 10,000 sampledata producing convincing h(t) .

The correlation method was also tested under noisyenvironment. By injecting a random uniform 10% noise and10% DC gain. The result in Figure 10 still captured h(t)overshoot, delay time and waveform shape accurately. Thedifference was the DC gain only. Correlation method wasproven to be effective under noisy environment.l~~~~~~~~~~~~~~~~~~~~~~~~~~~~--iil.i--l------

'', X: 5.01 e+005--07 ' '{'Y:0.7874

-- T-- r- --1-- -- n-- --- T-- - r-- - -r-- --

04 _ -,X: 5.01 O+005 X 5.02e+0055

-----------~~~~, Y:0.24 ------- -------- Y: 0.23|92 --------|--------Y:02245 Y 29

02 l

5017 50175 5013 50ol8 5.019 50195 502 50205 5021

11rrelatiun Ro 8 J65 IFigure 10. Rxy under uniform 10% random noise

condition

Simulation using complex full EM modeling platformdata bus topology

The cross correlation method was applied on complicateddata bus system as shown in Figure 6. Again, using HSPICEsimulation, single pulse response analysis was used as thebench mark for comparison. From the result, the overshootamplitude was 0.64 and the delay time was 1. Ins. Comparingwith cross correlation method using 10,000 samples, theovershoot was 0.553, which was close to the impulseresponse analysis. Both of them had same delay time, 1.Ins,and very similar waveform shape as shown in Figure 11. Thusit was concluded that correlation method using 10,000 sampledata produce convincing h(t) .

Figure I l. Comparison between impulse response andcorrelation method on complicated data bus. Both waveformshad very similar waveform shape

By using Blackman Tukey Window, the power spectrum,Sxy was round -ldB/GHz while S21, the result was -0.875dB/GHz as illustrated in Figure 12. They were close inside 4GHzfrequency band Thus it was concluded that cross correlation

1-4244-1392-3/07/$25.00 (©2007 IEEE

method producing Sxy but the accuracy was needed to beimproved by using better Window technique.

I~~~~~~~

a US ~ 1 1S5 2.5 3 3.6 4Power Spectirumnx (Hz)

x

View Insert Tools DeAftio Window Help

l~~~~~~~~~neto losseX f , ~~~~~~~~~~~~~~~~~~~~~~~< 0Figure2. Comarisonbetwee SyandS2

Thgue corrlatonmethiodbewasn aSoy testd unernos

environment. By injecting a distributed random uniform 1000noise, the result still captured h(t) overshoot, delay time andwaveform shape accurately. The Sxy also didn't show anydifference.

By injecting SOps jitter noise, the result still captured h(t)overshoot, delay time and waveform shape accuratelyalthough not as accurate as random noise. The Sxy alsoshowed minor difference.

Figure 13. Injecting 10% distributed random noise. Nodifference compared with noiseless environment

The result from cross correlation method correlated withthe benchmark impulse response analysis. By simulating themodel under 10% random noise, the result had very minordifference compared to noise free simulation. The algorithmwas proven to be robust.

The algorithm was applied to HSPICE full EM modelingdata bus transmission topology system that representingactual motherboard system. The Rxy result correlated with h(t)from impulse response. The Sxy was close to insertion loss,S21 but not as close as h(t). It was due to the Windowingtechnique because Blackman Tukey produced noise overpower spectrum. It was suggested to explore betterWindowing technique for more accurate result.

The algorithm was tested under noisy environment. It wasproven to be very robust with almost no difference using 10%distributed random noise. With the presence of 5000 jitter, theRxy and Sxy result were still look good but not as close as10% noise.

AcknowledgmentsThanks to Intel Corporation Penang Design Center in

providing EDA tools, circuit behavior model, package andmotherboard electrical modeling.

References1. Maynard Falconer, "Bus Design Boot Camp", Intel Corp,

20052. Adam Norman, "Chapter 13: Peak Distortion analysis"

Bus Design Boot camp, 3/3/053. M. Cases, D. N. de Araujo, E. Matoglu, "Electrical Design

and Specification Challenges for High Speed SerialLinks", 2005 Electronics Packaging TechnologyConference

4. Howard Johnson, "High-Speed Signal Propagation",Prentice Hall, 2003

5. B. Casper, M. Haycock, R. Mooney, "An Accurate andEfficient Analysis Method for Multi-Gb/sChip-to-ChipSignaling Schemes,"Symposium on VLSI Circuits Digestof Technical Papers, 2002, pp. 54-57, June 13-15, 2002

6. Cattalen Pelard, Edward Gebara, Andrew J. Kim, MichaelG. Vrazel, Franklin Bien, Youngsik Hur, MoonkyunMaeng, Soumya Chandramouli, Carl Chun, SanjayBajekal, Stephen E. Ralph, Bruce Schmukler, Vincent M.Hietala,, "Realization of Multigigabit ChannelEqualization and Crosstalk Cancellation IntegratedCircuits", IEEE JOURNAL OF SOLID-STATECIRCUITS, VOL. 39, NO. 10, OCTOBER 2004

Figure 14. Injecting 50% normal distributed random jitternoise. Slight difference was observed compared with noiselessenvironment

Conclusions

1-4244-1392-3/07/$25.00 (©2007 IEEE