A HIGH ACCURACY HIGH THROUGHPUT JITTER TEST SOLUTION ON ATE FOR 3GBPS AND 6GBPS SERIAL-ATA

Transcription

1 A HIGH ACCURACY HIGH THROUGHPUT JITTER TEST SOLUTION ON ATE FOR 3GBPS AND 6GBPS SERIAL-ATA Yogqua Fa, Y Ca ad Zeljko Zlc LSI Corporato 0 Amerca Parkway NE, Alletow, Pesylvaa 809 Emal: y.ca@ls.com Departmet of ECE, McGll Uversty ABSTRACT Jtter test producto s otorous for ts log test tme ad the challege of accuracy verfcato. Amog varous types of jtter, Radom Jtter (RJ) s most challegg to test o Automatc Test Equpmet (ATE) because of ts radomess. To be cosdered as a favorable jtter test producto for mult-ggabt devces, the RJ eeds to be measured wth sub-pcosecod accuracy ad the whole test tme s expected to be a few tes of mllsecods. However, o kow solutos meet these crtera to our best kowledge. I ths paper, we preset a systematc soluto for multple Gga-bt-per-secod (Gbps) Trasmtter (TX) jtter testg o ATE. Our udersamplg-based soluto extracts jtter ether from edge hstograms tme doma or from the jtter spectrum frequecy doma. Both approaches provde a RJ precso better tha ±0.5ps ad are capable of fshg the whole TX test wth 00ms. We have verfed the soluto wth data rates up to 6Gbps ad appled t mass producto.. INTRODUCTION Jtter s the devato of a sgal from ts deal tmg. It s composed of both determstc ad radom cotets. Radom Jtter (RJ) s caused by radom evets ad s usually characterzed statstcally. Determstc Jtter (DJ) s caused by determstc evets. Major DJ sources clude Perodc Jtter (PJ), Duty Cycle Dstorto (DCD) ad Iter- Symbol Iterferece (ISI). PJ s caused by repettve ose sources, such as clock sgals ad oscllators. DCD s caused by a mbalace the drve crcut. ISI s caused by frequecy related losses the sgal path, such as those caused by the badwdth lmtato. Most commucato stadards, such as Seral ATA (SATA), Fber Chael ad XAUI, specfy jtter terms of DJ ad TJ as separate specfcatos. TJ s the total jtter, whch s assocated wth a certa Bt Error Rate (BER) level. Table summarzes the TX jtter specfcatos for the SATA II [], whch drectly determe the jtter test lmts we should set. The TJ of the SATA TX should ot exceed 0.3UI at 0 - BER level ad DJ should ot exceed 0.7UI. Though the SATA specfcato does ot specfy the RJ lmt, we ca get the lmt by assumg all TJ s cotrbuted by RJ. The RJ wth 0.3UI peak-to-peak value at 0 - BER level traslates to a RJ root-mea-square (RMS) value of 0.0UI, or 7.0ps at 3Gbps data rate ad 3.5ps at 6Gbps data rate. The RJ RMS value s usually used to estmate TJ at 0 - BER level, as t s mpossble to drectly measure TJ at ths BER level volume producto due to log test tme t takes a few tes of mutes eve for 6Gbps data rate. Table. TX Jtter specfcatos SATA II TJ DJ RJ (RMS)* 0.3UI 0.7UI 0.0UI or 7.0ps@3G *Deduced from TJ specfcato To ecoomcally apply these test lmts producto, we eed the jtter test to have the followg capabltes: o Separatg jtter compoets o Achevg accuracy sub-pcosecods o Havg the test doe mllsecods Ufortuately, there s curretly o soluto o ATE that meets there crtera to our best kowledge, eve though the jtter measuremet ad decomposto have bee vestgated for years [0], []. Popular jtter testg solutos clude Bt Error Rate Testers (BERT), hstogram-based Osclloscopes, ad Tme Iterval Aalyzers (TIA) []. These solutos are commoly used for desg valdato ad characterzato o bech. However, we ca ot drectly apply them for at-speed testg producto because of the low throughput. For mult-ggabt Seralzer/Deseralzer (SerDes) testg producto, the most commo practce s to loop the output of the trasmtter to the put of the recever ether teral to the devce or through the loadboard [3], [4]. Its fucto s checked by comparg the output of the recever to the expected result. Ths loopback test ca cover the major fuctoalty of SerDes devces. Because of ts smplcty ad hgh throughput, the loopback test s very popular ad Paper. INTERNATIONAL TEST CONFERENCE /07/$ IEEE

2 may cases t s also the oly wdely used test to cover a SerDes devce/block. However, the loopback test ca ot provde ay kowledge of parametrcal characterstcs, cludg jtter performace. We ca oly assume that these specfcatos are guarateed by desg. Ufortuately, ths assumpto s o loger vald whle we keep advacg the semcoductor techology ad creasg the data rate, whch results tghteg the jtter budget. The devces ca creasgly fal just because they do ot comply wth the jtter specfcatos. It s hece becomg mperatve to clude jtter test producto order to dstgush bad devces from good oes. Ths s the oly way to esure the devce qualty ad to elmate or reduce customer returs. There are o may choces rght ow that ca do multggabt devces jtter complace testg producto. Most jtter test solutos are based o lab strumets, extra o-chp crcutry, or DUT board add-o modules [3-6]. These solutos are lmted ether by ts low throughput, low accuracy ad repeatablty, or by the hgh desg complexty of the devce or the loadboard. Because of these lmtatos, pure ATE-based solutos are preferred producto because of ther hgh portablty ad hgh throughput. I recet years, mult-ggabt sgal geerators ad dgtzers/samplers are becomg avalable as fullytegrated ATE strumets. Oe example s the GgaDg o Catalyst/Tger ATE from Teradye [9]. The GgaDg s a dgtzer, capable of capturg aalog sgals wth a tme resoluto better tha ps. Wth ths kd of strumets, t has become feasble to perform mult-ggabt devces jtter test o ATE [6-8], eve though systematc jtter extracto algorthms o ATE have ot matured. The jtter performace of a SerDes devce ca be characterzed by the jtter the output of the trasmtter ad the jtter tolerace of the recever. I [6], a TX jtter test soluto s proposed based o the hgh-speed dgtal ps of the Aglet s ATE. By shftg the compare strobes the tmg axs, ths approach frst bulds a bathtub curve, ad the apples the jtter separato algorthm. However, the test ecoomy of ths soluto eeds to be mproved: t takes ear secod eve wth ps resoluto. As jtter s just oe of hudreds parameters to be tested o a average devce, oe secod spet for oe test s stll too log o the ATE evromet. I addto, the accuracy also eeds to be mproved: RJ s ear ps hgher tha the bech result. I [7], a SATA test soluto o ATE s preseted, whch cludes TX jtter testg. However, the TX jtter testg scheme [7] s ot very accurate; t reports hgher RJ (~ ps) ad hgher TJ (0ps) tha the bech equpmet does. I addto, the test parameters ths soluto ca stll be further optmzed to acheve better test ecoomy. I [8], we preseted a recever jtter tolerace test soluto o ATE, whch s capable of acceleratg jtter tolerace test by 000 tmes. I ths paper, we preset a ew TX jtter testg soluto o ATE. Wth the curret ATE strumets, we acheve sub-pcosecod accuracy ad reduce test tme to 0.s 3G ad 6G applcatos, whch o oe else has ever acheved a ATE evromet to our best kowledge. Better performace ad hgher speed applcatos are attaable usg our soluto wth the advaces the ATE strumets the future they are oly lmted by the badwdth ad tmg resoluto of the dgtzer. The accuracy of the soluto s verfed by both bech equpmet ad ATE tself, ad the test has already bee appled volume producto. The remader of the paper presets the detals of our TX jtter test soluto. I Secto, we descrbe the prcples of settg test parameters for data acqusto. Secto 3 detals the data processg how jtter s separated ad decomposed both the tme doma ad the frequecy doma. We preset the expermetal results ad lmtatos Secto 4. Secto 5 draws coclusos.. TEST SETUP FOR DATA ACQUISITION Our jtter test soluto utlzes the GgaDg that s avalable o Teradye ATE [9]. Smlar strumets are also avalable from other ATE vedors. The GgaDg s a fully tegrated ATE dgtzg strumet wth a typcal udersamplg badwdth over 9 GHz. Its put voltage rage s 64mv to.04v (the SATA TX output rage s 400mv to 700mv). Wth a Mega sample memory ad -bt dgtzg resoluto, the GgaDg s capable of testg the TX jtter for all SATA applcatos:.5gbps, 3Gbps ad 6Gbps. Usg the GgaDg, we ca perform all TX fucto ad parameter tests wth a sgle capture of the TX output. Fgure. Test setup for data acqusto The test setup s show Fgure. The ATE provdes a referece clock sgal tx_ref_clk to the trasmtter; a PLL the TX the locks the TX output rate to the referece clock. I our applcatos, the deal tx_ref_clk s 30MHz ad the TX output data rate F DATA ca be.5g, 3G or 6G. The GgaDg Paper. INTERNATIONAL TEST CONFERENCE

3 captures the TX output wth a udersamplg rate Fs betwee 5 Mega Samples per secod (MS/s) to 0 MS/s. DJ, RJ ad TJ ca be extracted from captured samples. The udersamplg techque has bee used hgh-speed testg for years [0]. Ths techque frst captures the output sgal of the DUT at a samplg rate Fs that s lower tha the output data rate F DATA, ad the shuffles the captured samples a predetermed maer. The shuffled output s a sequece of samples that would have resulted from samplg at a much hgher frequecy - effectve samplg rate F eff. Although the udersamplg prcple s smple, the challeges are how to properly set test parameters for data acqusto ad how to extract jtter formato from captured samples. To capture wth reasoable test tme the TX output waveform of a adequate resoluto for jtter decomposto, we eed to properly set these parameters: o Test patter legth N o Effectve samplg rate F eff o Number of samples o Udersamplg rate Fs To provde adequate test coverage, the test patter legth N should be at least 0 bts because the wdth of the parallel data to the TX put s 0. O the other had, the legth should be as short as possble order to save test tme ad also smplfy the data processg. For these reasos, we choose a 0-bt test patter Ths patter cludes both hgh desty ad low desty trastos, wth a total of eght edges. Whe the 0-bt test patter s used at a data rate of F DATA =3GHz, the TX output (GgaDg put) fudametal frequecy F DUT s F F DATA DUT = = 50MHz () N The requred effectve samplg rate F eff s determed by our target test accuracy. To acheve a jtter measuremet resoluto better tha ps, we eed to have the effectve samplg resoluto better tha ps, whch correspods to a effectve samplg rate F eff hgher tha 000GHz. For 3GHz data sgals, ths traslates to capturg at least 333 samples per data bt. To leave some marg ad also keep test tme short, we choose to capture 400 samples per bt, whch results Feff = 400 * FDATA = 00GHz () The requred mmum umber of samples ca be calculated based o the patter legth ad the effectve samplg rate. To buld ts edge trasto hstograms ad acqure ther statstcal propertes wth a reasoable cofdece level, we eed to capture at least 0 cycles of the test patter. Twety cycles of the 0-bt patter wth 400 samples per bt traslate to 60k samples total. Aother factor that we eed to cosder whe determg the total umber of sample s the Fast Fourer Trasformato (FFT) requremet. We eed FFT later for jtter decomposto frequecy doma. The above derved umber of samples satsfes ths requremet. The udersamplg rate Fs eeds to be calculated based o the GgaDg put fudametal frequecy F DUT ad the effectve samplg rate F eff. I order to capture samples coheret wth the put sgal, we eed to satsfy the equato = K + (3) F F F s DUT where K s the umber of cycles of F DUT slpped to the ext sample. As show Fgure, the ATE sources the referece clock tx_ref_clk to the TX. Ideally, the clock should be 30MHz ad the TX output rate would be exactly 3GHz. However, the ATE ca ot source a clock sgal exactly at 30MHz as ths clock s derved from the Optcal Referece Clock (ORC) dvded by t0_clk_dv, where ORC=50,000THz ad t0_clk_dv ca oly be a teger. For ths reaso, we eed to keep the rato of F DUT ad F eff stead of takg the deal frequeces whe applyg Equatos () ad () to Equato (3). The ATE clock dvders Fgure are programmed accordg Equato (3) ad other ATE requremets. Table lsts the test parameters derved. Table. Testg parameters for 3Gbps sgals TX Ref. Clock (MHz) Udersamplg clock (MHz) Test patter Samples per bt Total samples bts k 3. JITTER EXTRACTION I ths secto, we descrbe how the jtter s extracted based o the acqured TX waveform wth the test setup ad the parameters dscussed Secto. Fgure s a example of the captured waveform of a 3G sgal. The waveform cossts of 0 cycles of the 0-bt test patter, wth 400 samples each data bt ad 60,000 samples total. Ths capture s used to perform all TX fucto ad parameter tests. I ths paper, we oly address jtter testg. Other measuremets, such as fucto, rse/fall tme ad pre-emphass, are fast ad straghtforward oce we capture the TX output waveform. Fgure. Captured TX sgal (ut: mv) eff Paper. INTERNATIONAL TEST CONFERENCE 3

4 3. Geeratg Edge Dsplacemet Bascally, jtter s the edge dsplacemet of the actual edge trasto posto compared to ts deal posto. I our test setup as show Fgure, the referece clock tx_ref_clk determes the data rate of the TX. The TX deal edge postos ca be calculated by assumg that all the data bts are trasmtted wthout ay jtter. The actual posto of each edge trasto mght devate from ts deal posto due to jtter. Fgure 3 shows a example of two edge trastos (L to H to L) captured usg the dgtzer. As the edge trastos are ot smooth, we use a curve fttg techque to extract actual zero crossg postos. The edge dsplacemet s obtaed from the derved edge posto mus the deal edge posto. The deal posto s calculated based o the frst derved edge posto ad the deal data rate of the trasmtter. I ths way, we extract 60 samples of the edge dsplacemet data from the 60 derved edge postos. I order to perform FFT for the jtter spectrum aalyss, we eed edge dsplacemet formato for every data bt. I our mplemetato, we assume that o jtter s troduced the data bts where o data trastos occur betwee two or more bts, so we just sert the edge dsplacemet data from the prevous edge trasto to terpolate o-trasto data bts. We wll later elmate the effects that the terpolato may cause. Fgure 5 llustrates the edge dsplacemet data of all the 400 captured data bts. Wth the terpolato, t s equvalet that the edge dsplacemet data s obtaed wth a samplg rate of F DATA, where F DATA =3G for 3Gpbs sgals. Fgure 3. Actual edge trastos ad curve fttg The curve fttg s doe a wdow cetered the edge trasto perod. Accordg to the SATA specfcato [], the TX rse/fall tme (0% - 80%) of 3G sgals s betwee 0.UI (67ps) ad 0.4UI(36ps). Whe the effectve samplg resoluto s 400 samples per bt, the umber of samples durg a edge trasto (0% - 80%) perod s betwee 80 ad 64. Therefore, we choose a wdow wth 80 samples to perform the curvg fttg as show Fgure 3. The frst captured zero crossg sample a trasto edge determes the cetre of the wdow that we choose for the curve fttg. Our expermets demostrate that a fast lear curve fttg provdes a smlar accuracy compared to other more tme-cosumg curve fttg techques such as computg the best-ft le [7]. Wth our lear curve fttg algorthm, the actual zero crossg posto s calculated based o the averages of the left part ad the rght part of the wdow used for the curve fttg. Fgure 4 plots all the edge trasto postos calculated from our lear curvg fttg techque. The x-axs deotes the edge sequece, whch has 60 edges the captured 400 data bts. The edge posto y-axs s deoted by the umber of samples relatve to the frst edge. Fgure 4. Derved edge postos from curve fttg Fgure 5. Edge dsplacemet samples Oce we get the edge dsplacemet data, we ca extract the DJ ad RJ compoets based o ther dfferet propertes both the tme doma ad the frequecy doma. The TJ ca be obtaed based o DJ ad RJ. 3. Tme Doma Approach I the tme doma approach, we buld the 0-bt test patter edge hstograms to extract the RJ ad DJ formato of the devce. Fgure 6 plots oe cycle of the actually captured 0-bt test patter ad the deal waveform. There are eght edges the test patter. The hstograms are bult by foldg (overlayg) the 60 samples of the extracted edge dsplacemet data every 8 samples, startg from the frst oe. As the eght cosecutve samples of the edge dsplacemet data correspod to oe cycle of the 0-bt test patter, we defe 0-bt as the foldg legth ths case. The foldg legth drectly determes the foldg frequecy, whch s also the lowest DJ frequecy that ca be cacelled ad therefore excluded from the RJ. Ay DJ whose frequecy s lower tha the foldg frequecy wll affect the RJ measuremet accuracy. For example, we fold the test patter every 0 bts for 3GHz applcatos, so the lowest DJ frequecy that ca be excluded from the RJ s 50MHz. I our applcatos, 50MHz, the word clock frequecy, s the domat fudametal DJ frequecy, so the DJ compoets do ot leak to RJ. Paper. INTERNATIONAL TEST CONFERENCE 4

5 By jtter defto, we eed a lot of samples at each edge order to capture the radomess ts hstogram. Based o our prevous aalyss, for the 0-bt test patter, takg samples o 400 bts would be adequate to acheve good accuracy wth reasoable test tme. It s also proved by the fact that we stll get smlar results whe we crease the umber of samples. The upper part of Fgure 7 llustrates the hstograms of the eght edges. As oly 0 cycles of the test patter are captured, each edge hstogram s costructed usg 0 samples of the edge dsplacemet data ad the hstogram resoluto s ps. ad Varace δ = E[( X mx ) = E[ X ] mx where E[..] s the expectato operator, δ s the Stadard Devato (SD) ad P [ x ] s the probablty of the hstogram at x. Accordg to these deftos, we ca obta the mea ad stadard devato of the hstogram at each edge. The hstogram formato s used to extract the RJ, DJ ad TJ of the devce. 3.. RJ Extracto RJ s caused by radom evets, prmarly by thermal ose electrcal compoets. As ths kd of evets exhbts a Gaussa dstrbuto, we assume RJ s Gaussa [], characterzed by ts RMS or SD value. The RJ value of the devce s obtaed by gettg the RMS value of the SDs of the eght edge hstograms: δ RJ + δ δ N = N where N = 8 for the 0-bt test patter. Fgure 6. Oe cycle of the test patter The RJ Gaussa property s also demostrated by the actual edge hstograms bult from the captured data. As we ca see from Fgure 7 ad Fgure (dscussed later), the hstograms are very close to Gaussa dstrbutos eve though oly 0 samples are captured at each edge. Therefore, we ca represet the RJ probablty desty fucto (PDF) at a edge usg the Gaussa fucto p x e x ( ) = ( ) δ δ π where δ s the SD of the hstogram at that edge. The RJ PDF at each of the trasto edges ca lead us to get the TJ profle of the devce oce we get the DJ at each edge. 3.. DJ Extracto By defto, the mea value m of a edge hstogram would reflect the DJ at that edge, whch gves DJ = m deal _ posto where s the edge dex. Fgure 7. Hstograms ad DJ of all eght edges The most mportat propertes of a hstogram are the mea, the mea-square, ad the varace. These parameters are defed by Mea m = x P[ x ] Mea-Square x x m = E[ x ] = x P[ x ] The DJ of a devce s the maxmum value mus the mmum value of the DJ values at all edges, whch gves DJ = max( DJ, DJ,..., DJ ) m( DJ, DJ,..., DJ where s the umber of total edges ad = 8 our case. The DJ value at each edge of the 0-bt data patter s llustrated the lower part of Fgure 7. The DJ of the devce s the peak-to-peak value of the plot, whch s 3.ps (the jtter at the 3 th UI mus the jtter at the 0 th UI). ) Paper. INTERNATIONAL TEST CONFERENCE 5

6 3..3 TJ Extrapolato TJ s comprsed of DJ ad RJ. As RJ s ubouded, RJ ad TJ are always assocated wth BER. The TJ specfcato defed ay commucato stadard s actually the peakto-peak total jtter value at a certa BER level. I order to extract the TJ peak-to-peak value, we eed frst to costruct the TJ profle. As we kow the DJ ad RJ profle at each trasto edge of the data patter, we ca costruct ts TJ profle through covoluto. Table 3 lsts all the RJ ad DJ values at each of the eght edges show Fgure 7. Table 3: RJ ad DJ values Fgure 7 Posto RJ RMS(ps) DJ (ps) Notes Edge Edge Mmum DJ Edge Edge Edge Maxmum DJ Edge Edge Edge where erf( x ) deotes the error fucto, defed as erf ( x) = x e π Oce we get the TJ CDF at each edge, the TJ CDF of the devce ca be represeted by 8 TJ _ CDF( x) = TJ _ CDF ( x) (9) 8 = Accordg to equato (8), equato (9) becomes 8 x m TJ _ CDF( x) = [ * erf ( )] (0) 8 = δ * Fgure 8 plots the PDF ad CDF of the devce TJ. 0 t dt As dscussed prevously, the RJ PDF at each edge ca be characterzed by RJ _ PDF x e ( x) = ( ) δ (4) δ π where s the edge dex ad δ s the RJ RMS at that edge. As we kow the exact DJ value at each edge, the TJ profle at a edge ca be calculated by covolutg RJ ad DJ at that edge: TJ _ PDF = RJ _ PDF DJ (5) where s the edge dex, =,,,8. If we deote the DJ value at edge wth m, accordg to equatos (4) ad (5), the TJ PDF at edge s represeted by TJ _ PDF x m e ( x) = ( ) δ (6) δ π To assocate the TJ wth BER, we eed to costruct the Cumulatve Dstrbuto Fucto (CDF) of the TJ profle at each edge: TJ _ CDF = x ( x) TJ _ PDF dx (7) The TJ _ CDF ( x) represets the probablty that the jtter (edge dsplacemet) resdes wth the rage of [, x]. For a zero mea Gaussa dstrbuto, we have CDF(- )=0, CDF(0)=0.5 ad CDF( )=. Accordg to equatos (6) ad (7), we have x TJ _ CDF x x m ( ) = ( ) δ e dx δ π x m = * erf ( ) (8) δ * Fgure 8. The PDF ad CDF of the devce TJ Oce we get the TJ CDF, we ca get TJ peak-to-peak value at a certa BER level by calculatg the tme dfferece betwee t ad t : TJ peak to BER = t t where t ad t satsfy TJ _ CDF( t) = BER / TJ _ CDF( t) = BER /. For the TJ profle show Fgure 8, we have TJ = UI ( UI ) pk = 0.550UI The above calculated TJ would reflect the TJ peak-to-peak value of the devce at BER=0 -. I ths example, the data rate s 3G; the calculated TJ value s 50.8ps. As we ca see, the above TJ extracto process volves tesve computatos ad hece takes a lot of tme. I producto, we ca estmate the RJ peak-to-peak value at BER = 0 - by multplyg the RJ RMS value wth the Q-factor at ths BER level. The TJ at BER = 0 - ca the be obtaed by summg the DJ ad the RJ peak-to-peak value: TJ = DJ +4.07* RJ where 4.07 s the Q-factor value at BER = 0 - []. Paper. INTERNATIONAL TEST CONFERENCE 6

7 For the above example (edge jtter values lsted Table 3), the Q-factor based TJ estmato gves a TJ value of 49.ps. Ths value s very close to the TJ value calculated based o the TJ CDF profle (50.8ps). Therefore, t s acceptable to use the Q-factor method for TJ calculato producto. 3.3 Frequecy Doma Approach I frequecy doma, jtter compoets are extracted from the jtter spectrum. The TJ spectrum ca be obtaed by passg the edge dsplacemet data as show Fgure 5 through a FFT [7]. RJ s the ose floor whle DJ compoets are the mpulses the spectrum. Fgure 9 llustrates the TJ spectrum of the captured sgal Fgure. The spectrum s obtaed by performg FFT o the edge dsplacemet data show Fgure 5. Accordg to ths spectrum, we ca get the power at each frequecy b, whch ca be deoted by C, where s from 0 to 99. frequecy bs, 49 of them are DJ bs (all C wth mod 4 = 0). To calculate the ose floor of the TJ spectrum, we replace all the DJ bs wth the average of the RJ bs gve by C RJ _ average 99 = mod 4 0 = *( 50 C ) Fgure 0 plots the spectrum after the above replacemet. It represets the RJ spectrum of the devce. Accordg to Parseval s theorem [7], the RMS value of the RJ spectrum s the square-root-of-sum-of-power of all bs gve by 99 RJ = C k + 49* C k= k mod 4 0 RJ _ average Fgure 0. RJ spectrum Fgure 9. TJ spectrum 3.3. RJ Extracto I frequecy doma, the RJ RMS value s equvalet to the total ose power the TJ spectrum. The ose power spectrum s costructed by replacg all the DJ frequecy bs the TJ spectrum wth the average of the o-dj frequecy bs. I order to remove the DJ completely for RJ extracto, ths approach requres the DJ frequeces to be coheret [7]: all the DJ frequeces eed to be exactly multples of the FFT frequecy resoluto. A o-coheret DJ frequecy appears to cosst of may frequecy compoets the FFT frequecy bs ad hece cotamates the RJ spectrum. I our applcatos, oe DJ source s the devce referece clock, whch s 30MHz. Aother DJ source s the word clock of the devce, whch s 50MHz for 3G sgals. The word clock s used the SerDes crcutry to sychroze the parallel data. I addto, the ISI s also a DJ source. For the 0-bt data patter, the ISI frequeces would be the multples of 50MHz for 3G sgals. For these facts ad also accordg to the TJ spectrum, we kow that all the DJ frequeces our applcatos are multples of 30MHz the devce referece clock frequecy. I addto, as dscussed Secto, our test setup strctly makes the referece clock frequecy ad the output data rate coheret. I our applcatos, the FFT frequecy resoluto s 7.5MHz for 3G sgals. Therefore, all the DJ frequeces are multples of the FFT frequecy resoluto. Amog the DJ Extracto I frequecy doma, we extract the devce DJ compoets from ts TJ spectrum. Lke some commercal stad-alog jtter equpmet [] ad related patets, we adopt the followg steps for the DJ extracto (PJ ca also be extracted usg the smlar procedure): () Obtag the DJ-oly spectrum by settg to zero all bs the TJ spectrum that are attrbutable to RJ. I our case, we set to zero all the TJ bs that are ot multples of 4 (b 4 correspods to 30MHz) () Performg a verse FFT o the DJ-oly spectrum to geerate the tme-doma data. The geerated data would reflect the edge dsplacemet cotrbuted by DJ. (3) Gettg the peak-to-peak value of the data excludg locatos that actually do ot have edge trastos. The peak-to-peak value s DJ value of the devce. I step (3), we exclude the locatos that actually do ot have edge trastos whe calculatg the fal DJ value. Ths would elmate the artfacts that mght have bee troduced whe we sert the edge dsplacemet data o o-trasto edges order to perform the FFT. Due to the DJ coherece costrat, we eed to vestgate the valdty of each ew desg whe usg the spectrum approach for jtter extracto. Oe good thg s that the jtter spectrum s maly determed by the devce archtecture (such as CDR ad PLL structure) ad the test setup (the test hardware ad the test patter). Oce a desg s falzed, ts jtter spectrum costtutes are fxed. Therefore, the valdato oly eeds to be doe oce for every ew desg. Paper. INTERNATIONAL TEST CONFERENCE 7

8 3.4 Hybrd Approach As dscussed prevously, we ca extract the jtter compoets from ether the tme doma or the frequecy doma. Each approach has ts advatages ad dsadvatages. If the test patter s ot too log, such as 0- bt, we prefer the tme doma approach. The reasos are that we do ot eed to pay much atteto to the actual DJ frequeces ad that foldg 0-bt data s ot too complcated. However, some specal cases, the lmtato of the tme doma approach may arse. As we kow, the tme doma approach ca ot exclude DJ frequeces below the foldg frequecy from RJ. Whe we fold the data every 0 bts 3Gbps applcatos, the lowest DJ frequecy that ca be excluded from RJ s 50MHz. Ths s good eough our applcatos as 50MHz ad ts multples are the domat DJ frequeces as show the TJ spectrum. However, we dd observe that some specal cases (such as at low power supples slow materals), lower frequecy DJ may be troduced. Fgure captures such a case the edge hstograms that are fold at 50MHz but cota 30MHz DJ. I ths case, the RJ dstrbuto s ot Gaussa ay more due to the DJ leakage. If we stll use the SD to represet the RJ, the RJ would be exaggerated. Fgure. Hstograms wth low frequecy DJ: SD = 4.08Ps I order to exclude the 30MHz DJ frequecy tme doma, we eed a mmum foldg legth of 00 bts for 3G applcatos. Ths requres capturg at least 000-bt data. I producto, we ca ot afford the log test tme requred for the acqusto ad processg for such large amout of data. The above problem ca be solved by removg the low frequecy DJ from the edge dsplacemet data before buldg edge hstograms. We set the specfc low frequecy DJ compoets the jtter spectrum to zero ad the perform a verse FFT to get the edge dsplacemet data that does ot cota low frequecy DJ. Fgure plots the hstograms where the low frequecy DJ has bee removed. The stadard devato of the hstogram would reflect the true RJ of the devce. Ths approach s also useful whe we ca ot get accurate RJ measuremets usg the frequecy doma approach because DJ compoets are ot coheret. Fgure. Hstograms after removg low frequecy DJ: RJ = SD =.70Ps 4. EXPERIMENTAL RESULTS To evaluate a jtter test soluto used mass producto, throughput ad accuracy are the two most mportat crtera. Our expermetal results demostrate the superorty of our proposed soluto both throughput ad accuracy. I ATE evromet, every mllsecod couts supplyg the most compettve products terms of both performace ad prce. As dscussed Secto, all the test parameters (patter legth, effectve samplg rate, umber of samples, ad udersamplg rate) our soluto have bee optmzed to keep the test tme as short as possble whle stll capable of accurately capturg all the formato we eed for the trasmtter tests. For both 3Gbps ad 6Gpbs applcatos, we maaged to fsh the etre trasmtter testg wth 00 mllsecods, whch cludes the data capture, the jtter extracto ad other trasmtter tests, such as fucto ad rsg/fallg tme tests. Accuracy shows how close the measured jtter value s to ts true value. The true value s usually obtaed through a bech strumet whose accuracy has bee verfed ad s wdely accepted. Repeatablty shows whether the test gves the same or smlar result from ru to ru ad from tme to tme for the same devce whle other codtos, such as voltages ad temperature are the same. We have coducted tesve explorato of the repeatablty ad accuracy of our soluto. 4. Bech Correlato We have correlated our ATE jtter test soluto wth the commercally avalable jtter test strumet Tektrox TDS654C, whch s favored by may test/applcato egeers for ts excellet jtter extracto ablty ad accuracy. Table 4 shows the results from 3 correlato devces. The ATE data ths table records the jtter mea values from the tme doma approach wth 0 rus for each devce a 3Gbps applcato. The repeatablty of our ATE soluto s dscussed later. As we ca see, the RJ dfferece betwee bech ad ATE s wth 0.ps; the DJ dfferece s wth 3ps (DJ from ATE s Paper. INTERNATIONAL TEST CONFERENCE 8

9 cosstetly slghtly hgher tha that from the bech equpmet). As we kow, absolute correlato umbers for dfferet jtter test solutos rarely happes. Cosderg ths s doe o ATE wth a completely dfferet strumet ad setup from the bech evromet, the correlato result s very good. Table 4. Jtter measuremet betwee ATE ad bech Jtter / Devce Devce Devce 3 Devce Bech ATE Bech ATE Bech ATE RJ DJ TJ Oe reaso for the hgher DJ o ATE s that the sgal path o ATE s loger tha that o bech. The loger sgal path ca troduce more ISI, ad hece results hgher DJ o ATE. I addto, the dfferet TJ extrapolato algorthms betwee the bech ad ATE also troduce dfferece the fal TJ report. 4. ATE Correlato Betwee the two RJ Approaches The frequecy doma approach s less patter-depedet as t does ot volve buldg hstograms. Ths approach s preferred o ATE f we eed to vestgate the jtter performace wth dfferet test patters. However, the results from ths approach eed to be verfed as the extracto process volves data terpolato ad jtter compoet replacemet. These steps mght troduce errors as the assumptos for these steps may ot be vald. I addto, the frequecy doma approach requres the DJ frequecy leakage s eglgble, whch may ot be satsfed some cases. O other had, the tme doma approach s very straghtforward. It ca be used to correlate the test results from the frequecy doma. Fgure 3 shows the test results of a devce wth 0 rus o ATE, where RJ_Spectrum s the RJ value from the frequecy doma approach, ad RJ_Tmg s the RJ value from the tme doma approach. It demostrates that both approaches exhbt good repeatablty ad the correlato s also very good. RJ (ps) RJ_Spectrum RJ_Tmg Ru Sequece Fgure 3: RJ repeatablty ad correlato We also evaluated the correlato betwee the two approaches by usg parts across Process, Voltage ad Temperature (PVT) corers wth a wde varety of jtter characterstcs. Fgure 4 plots the jtter dstrbuto across the PVT corers from both approaches, where the x-axs deotes the measured RJ value ad the y-axs represets the umber of hts. The dfferece betwee the two approaches s very small: the measured jtter mea dfferece s oly 0.ps ad dstrbuto profles are very smlar. Mea:.6ps M:.8ps Max: 3.8ps Mea:.4ps M:.6ps Max: 3.7ps (a) RJ_Spectrum dstrbuto (b) RJ_Tmg dstrbuto Fgure 4. Jtter dstrbuto across PVT corers As we ca see, the tme doma ad frequecy doma approaches correlate well o ATE from ether multple rus for a sgle devce or a larger umber of devces across all codtos. These expermets demostrate the excellet accuracy ad repeatablty of our jtter test soluto. 4.3 Extedg to 6 Gbps Applcatos Although our prevous dscusso cocetrates o 3Gpbs applcatos, our soluto apples ay data rates as log as the dgtzer badwdth s eough. As the badwdth of our ATE strumet s above 9 GHz, we easly exted our jtter test soluto from 3Gpbs applcatos to 6Gpbs applcatos eve though the specfcato for 6G SATA s ot avalable yet. For 6Gbps applcatos, we oly eed to adjust the referece clock ad the udersamplg clock accordg the data acqusto prcples dscussed Secto. Fgure 5 shows part of a 6Gbps waveform captured usg our soluto. Based o the waveform, we ca extract the jtter compoets usg exactly the same scheme as dscussed for 3G sgals. Fgure 6 shows the measured jtter at 6G from oe devce wth 0 rus, where the upper part plots both DJ ad RJ ad the lower part plots RJ oly. At 6G data rates, our soluto stll provdes smlar performace to that at 3G. Fgure 5. Captured 6G waveform (oly 45 bts show) Paper. INTERNATIONAL TEST CONFERENCE 9

10 Jtter (ps) RJ (ps) RJ_Spectrum RJ_Tmg DJ RJ_Spectrum RJ_Tmg Ru Sequece Fgure 6. RJ ad DJ at 6G data rate 4.4 Lmtatos of Each Approach As already metoed, each of the two jtter extracto approaches has ts lmtato. The frequecy doma approach eeds DJ compoets to be fxed ad coheret wth the FFT frequecy resoluto. Ths requremet s satsfed our applcatos. However, some applcatos, the DJ frequeces may ot be coheret ad eve may vary from devce to devce. I ths case, the frequecy leakage may degrade the jtter test accuracy f we oly rely o the frequecy doma approach. The tme doma approach requres that the major DJ frequeces are multples of the foldg frequecy order to avod DJ bleedg to RJ. If we have low frequecy DJ, the foldg frequecy must be also low. It therefore requres capturg a larger umber of data bts ad hece eeds loger test tme. Although the hybrd approach ca save some test tme ths case, t stll requres that the low frequecy DJ compoets are cosstet. Otherwse, DJ mght worse the estmato of RJ. 5. Coclusos We have preseted a systematc hgh-accuracy hghthroughput TX jtter test soluto o ATE for 3Gbps ad 6Gbps SATA. The whole test ca be doe wth 00 mllsecods ad the accuracy s wth ±0.5ps, whch o oe else has ever acheved a ATE evromet to our best kowledge. Based o stadard ATE strumet, the proposed jtter extract algorthm s portable ad scalable. Our method has bee successfully appled to several desgs producto to qualfy the TX jtter requremet for mllos of devces shpped to customers. Whe combg the TX jtter test method wth the ATE-based jtter tolerace test scheme preseted [8], we have proposed a complete jtter complace test soluto o ATE. 6. Ackowledgemets The authors would lke to ackowledge Kev Rchter for collectg the PVT data. REFERENCES [] Seral ATA Iteratoal Orgazato: Seral ATA Revso.5 Specfcato ( Fal Specfcato ), October 7, 005 [] Y. Ca, S. Werer, G. Zhag, M. Olse, R. Brk, Jtter Testg for Mult-ggabt Backplae SerDes Techques to Decompose ad Combe Varous Types of Jtter, IEEE Iteratoal Test Coferece, p , 00 [3] T. Yamaguch, Loopback or ot, IEEE Iteratoal Test Coferece, p. 434, 004 [4] Y. Ca, B. Laqua, K. Luehma, Jtter Testg For Ggabt Seral Commucato Trascevers, IEEE Desg ad Test of Computers, pp , 00. [5] Y. Ca, Jtter Test Producto for Hgh Speed Seral Lks, IEEE Iteratoal Test Coferece, 003 [6] G. Hasel, K. Steglbauer, K. Schulze ad J. Morera, Implemetato of a Ecoomc Jtter Complace Test for a Mult-Ggabt Devce o ATE, IEEE Iteratoal Test Coferece, 004 [7] Y. Ca, A. Bhattacharyya, J. Martoe, A. Verma, W. Burchaowsk, A Comprehesve Producto Test Soluto for.5gb/s ad 3GB/S Seral-ATA, IEEE Iteratoal Test Coferece, 005. [8] Y. Fa, Y. Ca, L. Fag, A. Verma, B. Burcaowsk, Z. Zlc ad S. Kumar, A Accelerated Jtter Tolerace Test Techque o ATE fro.5gg/s ad 3GB/s Seral- ATA, IEEE Iteratoal Test Coferece ITC 006. [9] [0] W. Dalal ad D. Rosethal, Measurg jtter of Hgh Speed Data Chaels Usg Udersamplg Techques, IEEE Iteratoal Test Coferece ITC, 998. [] Aalyzg Jtter Usg a Spectrum Approach, Tektrox Applcato ote, [] M. L, J.Wlstrup, R. Resse ad D. Petrch, A New Method for Jtter Decomposto through Its Dstrbuto Tal Fttg, IEEE Iteratoal Test Coferece, 999. [3] M. Hafed, D. Watks, C. Tam ad B. Pshdad, Massvely Parallel Valdato of Hgh-speed Seral Iterfaces usg Compact Istrumet Modules, IEEE Iteratoal Test Coferece, 006 [4] B. Laqua ad Y. Ca, Testg Ggabt Multlae SerDes Iterfaces wth Passve Jtter Ijecto Flters, IEEE Iteratoal Test Coferece, 00. [5] S. Suter, A. Roy, J. Cote, A Automated, Complete, Structural Test Soluto for SERDES, IEEE Iteratoal Test Coferece, 004. [6] A. H. Cha, ad G. W. Roberts, A Jtter Characterzato System Usg a Compoet-Ivarat Verer Delay Le, IEEE Trasactos o VLSI Systems, Volume, Issue, Ja. 004 [7] M. Burs ad G. W. Roberts, A Itroducto to Mxed- Sgal IC Test ad Measuremet, Oxford Uversty Press, 00 Paper. INTERNATIONAL TEST CONFERENCE 0