Smart-Pixel Cellular Neural Networks in Analog Current-Mode CMOS Technology

IEEE JOURNAL OF SOLID-STATE CIRCUITS. VOL. 29, NO. 8, AUGUST 1994 895 Sma -Pixel Cellula Neu al Ne wo ks in Analog Cu en -Mode CMOS Technology S. Espejo, A. Roddguez-Vazquez, Membe , ZEEE, R. Domhguez-Cas o, J. L. Hue as, and E. Shnchez-Sinencio Abs ac -This pape p esen s a sys ema ic app oach o design CMOS chips wi h concu en pic u e acquisi ion and p ocess- ing capabili ies. These chips consis o egula a angemen s o elemen a y uni s, called sma pixels. Ligh de ec ion is made wi h e ical CMOS-BJT’s connec ed in a Da ling on s uc u e. Pixel sma ness is achie ed by exploi ing he Cellula Neu al Ne wo k pa adigm [l], [2], inco po a ing a each pixel loca ion an analog compu ing cell which in e ac s wi h hose o nea by pixels. We p opose a cu en -mode implemen a ion echnique and gi e measu emen s om wo 16 x 16 p o o ypes in a single- poly double-me al CMOS n-well 1.6-pm echnology. In addi ion o he senso y and p ocessing ci cui y, bo h chips inco po a e ligh -adap a ion ci cui y o au oma ic con as adjus men . They ob ain sma -pixel densi ies up o 89 uni s/mm2, wi h a powe consump ion down o 105 pW/uni and image p ocessing imes below 2 ps. I. INTRODUCTION OMMON a chi ec u es o image-p ocessing sys ems use C a on -end senso y plane wi h digi al-encoding o he pixel alues, and se ial ansmission o hese digi al da a o subsequen p ocessing using ei he ASIC’s o gene al-pu pose compu e s. Con a y o his app oach, sma -pixel chips [3] inco po a e an analog compu ing cell a each senso y poin , achie ing high speed and low a ea occupa ion in he combined senso ylp ocessing unc ions by ully exploi ing pa allelism. The combined spa ial dis ibu ion o senso y and p ocessing ci cui y elimina es he ime equi ed o da a ansmission om he senso y o he p ocessing plane du ing he image acquisi ion p ocess. In addi ion, in some image-p ocessing applica ions, he ele an in o ma ion con ained in he ou pu image can be desc ibed by a educed numbe o a iables, allowing a as downloading o he esul s o subsequen e alua ion. CMOS echnologies o e unique ea u es o he design o sma -pixel chips. On one hand, MOS ansis o ope a ion unde no mal biasing in s ong in e sion is no d as ically a ec ed by inciden ligh ; on he o he , pho osensi i e CMOS de ices can be buil by exploi ing he many junc ion de ices a ailable in CMOS echnologies [4]. Howe e , p e ious ap- p oaches o CMOS design o sma -pixel chips lack gene ali y, as hey ely on implemen a ion me hods sui able o speci ic Manusc ip ecei ed Decembe 1993; e ised Ap il 6, 1994. S. Espejo, A. Rod iguez-VBzquez, R. Domlnguez-Cas o, and J. L. Hue as a e wi h he Cen o Nacional de Mic oelec 6nica-Uni e sidad de Se illa, Edi icio CICA, Cma ia sn, 41012-Se illa, Spain. E. Sinchez-Sinencio is wi h he Depa men o Elec ical Enginee ing, Texas A&M Uni e si . College S a ion. TX 77843 USA. applica ions. In some cases, he p ocessing- ask pe o med a each pixel does no imply collec i e compu a ion [3], while mos o he app oaches o “pixel-sma ness’’ a e based on ac i e implemen a ions o esis i e-g id ne wo ks [5], [6]. The pa adigm o Cellula Neu al Ne wo ks (CNN) [l], [2] is a e y sui able amewo k o sys ema ic design o pa allel senso y-p ocessing chips. On one hand, CNN’s consis o egula a angemen s o cells- - opologically iden ical o sma - pixel chips. On he o he , hei cells a e only locally connec ed, and hus, equi e simple ou ing. Also, he as body o li e a- u e on CNN heo y and applica ions demons a es ou s anding ea u es o his pa adigm o a ay-p ocessing [7]. In pa icula , esis i e g ids ha e ecen ly been demons a ed as a pa icula CNN class [8]. No expe imen al sma -pixel CNN chips ha e been epo ed o da e. This pape ou lines a design app oach using Da ling on pho o ansis o s and cu en -mode p ocessing ci cui y. I is based on a modi ied e sion o he o iginal CNN model which enables op imum speedlpowe and a ea occupa ion in VLSI design [9], [IO]. The senso s include an au oma ic adjus men ci cui y which ensu es p ope beha io unde di e en illu- mina ion condi ions. Ou p oposals a e demons a ed ia wo wo king sma -pixel chips, in a single-poly, 1.6-pm, n-well CMOS echnology. In addi ion o hei op ical inpu , hese chips exhibi much be e a ea and speed/powe igu es han p e ious CNN implemen a ions [ 111, [ 121. Sec ion I1 desc ibes some gene al aspec s o sma -pixel chips, and Sec ion I11 ou lines he p oposed compu a ion algo- i hm. Sec ions IV and V discuss he senso y and p ocessing ci cui y, espec i ely, and he expe imen al p o o ypes a e desc ibed in Sec ion VI. 11. SMART-PIXEL CHIPS In his pape , pixel deno es he elemen a y senso y uni used o de ec poin wise ligh signals. These senso y uni s a e ealized in CMOS echnology using any compa ible junc ion de ice o gene a e a cu en whose alue is an inc easing unc ion o he ligh in ensi y [3], [ 131, [ 141. The acquisi ion o wo-dimensional scenes equi es pixels a anged on o egula g ids, as shown in Fig. 1. Each pixel in his senso y plane gene a es a cu en I, which codi ies a co esponding poin o he inpu image, whe e he index c (i.,j) indica es he pixel a he i h ow and j h column on he g id and a ies o e he whole g id domain GD(c E GD). Thus, he whole image is IEEE Log Numbe 4402 142; cap u ed in o a ma ix o cu en s [Ic]. 0018-9200/94$04.00 0 1994 IEEE Au ho ized licensed use limi ed o: Uni e sidad de Se illa. Downloaded on Ma ch 20,2020 a 15:33:44 UTC om IEEE Xplo e. Res ic ions apply. 896 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 29. NO. 8, AUGUST 1994 om o he cells J Fig. 1. Illus a ing he co e a chi ec u e o sma -pixel chips. Fig. 1 illus a es he a chi ec u e o sma -pixel chips: each uni (also called sma -pixel o cell ) senses a poin o he inpu image and in e ac s wi h he o he uni s in he a angemen o pe o m pa allel-p ocessing asks on he inpu cu en ma ix Sma -pixel chips a e o s ong p ac ical in e es o pa e n [IC]. ecogni ion p oblems, o de ec ea u es o he inpu image. Fo example, Fig. 2 illus a es he ask o de ec ion o connec ed componen s (DCC), which consis s o coun ing he numbe o Fig. 2. Connec ed componen de ec ion in ou di e en di ec ions. - connec ed pieces encoun e ed by scanning an inpu image in a gi en di ec ion [15]. Pa e n ecogni ion can be ealized by p ocessing he da a ob ained a e pe o ming his ask in he di ec ions shown in Fig. 2 [16], [17]. This da a is con ained in a ew ows and columns a he g id bo de s. In addi ion o hei usage o p ep ocessing asks, sma -pixel chips a e also use ul as s and-alone uni s o nonin ensi e compu a ion asks such as hal oning [ 181, mo ion de ec ion [ 191-[21], ange- inding [3], e c. 111. THE CNN PARALLEL PROCESSING PARADIGM As Fig. 1 illus a es, sma -pixel CNN chips consis o egula a angemen s o iden ical uni s, each including a pho osenso and an analog compu ing cell. Such an en i y ans o ms he inpu image [I,] in o an ou pu ma ix [y,] ia a dynamic p ocess o in e ac ions among he compu ing cells. The dis inc i e ea u e o he CNN pa adigm is ha hese in e ac ions a e local, limi ed o each cell o a educed se o neighbo s, loca ed wi hin a dis ance T in he g id. In pa icula , he e is a wide ca alog o image p ocessing asks a ailable o ne wo ks whe e pa ame e T (called neighbo hood adius) is uni y- e y appealing o VLSI implemen a ions because connec ion among uni s is made by abu men , equi ing no ex a ou ing. The dynamic compu a ion p ocess o CNN's, as p oposed in [1], in ol es h ee a iables pe cell: (a) cell s a e: xc( ), which con eys cell ene gy in o ma ion as a unc ion o ime; (b) cell ou pu : yc( ), ob ained om he cell s a e ia a so - limi e piecewise-linea ans o ma ion, (1) Yc = (G) = ;(I.c + 11 - 15, - 11) d awn in Fig. 3(a); and (c) cell ex e nal-inpu : U,. P ocessing i sel is go e ned by a se o coupled nonlinea di e en ial equa ions, one pe cell. We use equa ions ha di e om hose o iginally p oposed by Chua-Yang [ 13, and which enable he op imiza ion o he speedlpowe a io and a ea occupa ion o VLSI CNN chips. The p oposed equa ions a e gi en by [9], [lo]: dz, d - =- + {A,dYd( ) + &dud) dE A', (c) Vc E GD (2) whe e g( .) is a nonlinea dissipa i e e m de ined as, m(5, + 1) - 1 2, < -1 g(z,) = .c o he wise (3) { m(5, - 1) + 1 5, > 1 whe e m > 1 is a pa ame e o he model. Func ion g(.) is d awn in Fig. 3(b). Summa ions in (2) ex end o e he neigh- bo hood o he c h cell, deno ed by NT(c), which con ains adjacen cells loca ed wi hin a dis ance T in he g id, and includes cell c i sel . P ocessing asks pe o med by CNN's a e de e mined by he con e gence o (2) o bina y (y, = *l,Vc E GD) equilib ium s a es ollowing he ansien ini ialized by [xc( O)], d i en by [U,], and unde he bounda y condi ions imposed by cells a he ne bo de . Depending on he applica ion, he cu en 1, gene a ed a each cell's pho osenso is used as ini ial alue o he s a e a iable zc(0) o as ex e nal inpu Au ho ized licensed use limi ed o: Uni e sidad de Se illa. Downloaded on Ma ch 20,2020 a 15:33:44 UTC om IEEE Xplo e. Res ic ions apply. ESPEJO e al.: SMART-PIXEL CELLULAR NEURAL NETWORKS IN ANALOG CURRENT-MODE CMOS TECHNOLOGY (b) Fig. 3. CNN cell nonlinea i ies. (a) Ou pu nonlinea i y; (b) dissipa i e e m. U,. In he la e case, he ini ial s a es a e usually se o a cons an alue. The ou come o he ask depends on pa ame e s Bcd,Acd and D, o (2), called con ol, eedback, and o se pa ame e s, espec i ely, and on he bounda y condi ions. The con ol and eedback pa ame e s can be a anged in o ma ices, which p o ide a pic o ial iew o he in e ac ions wi hin each cell’s neighbo hood. Fo uni o m ne wo ks hese ma ices a e in a ian h oughou he g id domain- hey a e empla es. The unc ionali y o uni o m CNN’s is de e mined by i s con ol, B, and eedback, A, empla e ma ices, and i s o se pa ame e , D. Fo illus a ion pu poses, Table I sum- ma izes he empla es used o some signi ican p ep ocessing asks. To gua an ee co ec ope a ion o sma -pixel CNN chips, an impo an ma hema ical issue is o de e mine condi ions o he empla e pa ame e s ha yield con e gence o he ou pu ma ix [yC] o bina y s a es o any inpu . Such a ma hema ical analysis o he model p oposed in his pape , gi en by (2) and (3), is ou o his pape ’s scope and has been epo ed elsewhe e [9] o any m 2 1. Ou ci cui s use he pa icula ~ 891 TABLE I SOME CNN TEMPLATES Applica ion A B D Noise Fil e ing [! 1 PI [! ! ?I 0 PI 000 010 Hole Filling [28] Con ex Come s Ex ac ion [2] Bo de s Ex ac ion [2] Connec ed Componen De ec ion [15] Shadow C ea ion 1291 -11.1 -11.1 -114 2 -114 -11-2 -114 2 -114 -114 -114 -114 -1 -11.1 -3 -1141 -1 14 -1 1.1 -2 0 0 case m -+ 00, in which he nonlinea dissipa i e e m o ces he s a e a iable zc o emain wi hin he in e al [-l> 11. Consequen ly, zc( ) = yc( ), and he implemen a ion o he nonlinea ope a o in (1) is no equi ed. IV. SENSORY CIRCUITRY A. Pho osenso s The simples pho osensi i e de ices o CMOS n-well ech- nologies a e e e se-biased pho odiodes, o med ei he di ec ly be ween n+-di usion and subs a e [3] o be ween well and subs a e [13]. Cu en le el o bo h de ices is an inc easing unc ion o he junc ion a ea. In pa icula , we ha e measu ed cu en s up o 20 nA o well-subs a e pho odiodes wi h well a ea o 100 x 100pm2, in a 1.6-pm single-poly echnology, unde en i onmen al labo a o y ligh ing. This cu en le el inc eases signi ican ly using a e ical CMOS-BJT as pho o- senso . Fig. 4(a) shows a concep ual layou and c oss-sec ion o his de ice, whose cu en is app oxima ely p opo ional o he a ea o he wellhubs a e junc ion, A, in he igu e. Cu en gene a ed by his de ice is p + 1 imes la ge han ha o a pho odiode wi h he same well a ea whe e IT deno es he pho o ansis o cu en , Il is he co esponding cu en o he well-diode, and is he ansis o cu en -gain; measu ed in o his echnology is 37.7 0.8, basically independen o ansis o geome y [22]. We ha e measu ed cu en s up o 430 30 nA (unde no mal labo a o y illumina ion) o pho o ansis o s wi h passi a ed well a ea o 60 x 60pm2. Consequen ly, and since cu en and well a ea a e app oxima ely linea ly ela ed, i ex apola es cu en le els o 20 nA o minimum a ea de ices (13.6 x 13.6 pm2)-needed o inc eased densi y sma -pixel chips. Howe e , o some c i ical asks [7] hese le els p o ided by minimum pho osenso s may no be la ge enough o gua an ee he ma ching le el equi ed by he signal p ocessing ci cui y, Au ho ized licensed use limi ed o: Uni e sidad de Se illa. Downloaded on Ma ch 20,2020 a 15:33:44 UTC om IEEE Xplo e. Res ic ions apply. 898 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 29, NO. 8, AUGUST 1994 I' I i 0 0 Fig. 4. igu a ion o wo e ical p-n-p ansis o s. (a) CMOS compa ible e ical p-n-p an sis o . (1 J) Da ling on con- hus equi ing some ampli ica ion. Simples s a egies use ei he la ge wells o cascaded cu en ampli ie s- e y cos ly in e ms o a ea occupa ion and, o he la e , inaccu a e. In- s ead, we use an addi ional e ical BJT o achie e Da ling on ampli ica ion by a ac o o p + 1, wi h p ac ically no a ea o e head. Fig. 4(b) shows he concep ual layou and c oss- sec ion o his Da ling on pho o ansis o . Cu en o his de ice is 1, N (P + 1)1T N (P + 1)21w. (5) while i s a ea occupa ion is sca cely inc eased by ha o a minimum-size e ical BJT. Measu emen s wi h A, in Fig. 4(b) equal o 60 x 60pm2 esul in cu en s up o 18 * 2pA. Fig. 5(a) shows he ou pu cha ac e is ic measu ed om a Da ling on pho o ansis o wi h A, = 60 x 60pm2 unde cons an en i onmen al illumina ion (b igh -cu en ), while Fig. 5(b) shows he esul ob ained when he en i onmen ligh is g adually educed o comple e da kness. Da k-cu en was 215+ 10 PA, which means ha he b igh - o-da k cu en - ange is close o 100 dB o en i onmen al labo a o y illumina ion. The same ange is obse ed o a single p-n-p de ice, while simple pho odiodes yield abou 80 dB. Al hough hese esul s a e op imis ic in he sense ha in eal images he e will be (b) Fig. 5. Measu ed ou pu cha ac e is ics o a Da ling on pho o ansis o wi h A + = GO x GO pm2: (a) unde cons an en i onmen illumina ion; (b) e ec o g adual educ ion o illumina ion du ing he sweep o I ;.E. no comple ely da k a eas, he b igh - o-da k cu en - a ios measu ed p o ide a wide enough ange o da a acquisi- ion. The ampli ica ion o he Da ling on s uc u e p o ides a su icien cu en le el e en i de ice a ea is subs an ially dec eased. B. Au oze o S a egy Al hough pho osenso s p oduce unidi ec ional cu en low, double- ail signals a e easily ob ained by bias-shi ing, as shown in Fig. 6(a). Cu en sou ce ITH se s he ze o-le el o he double- ail signal. To gua an ee good con as , i s alue should be se somewhe e be ween he maximum and he minimum ligh -induced cu en s among all pho osenso s. I ligh ing condi ions o all possible inpu scenes a e uni o m and known a p io i, ITH can be se o a ixed alue. In a mo e gene al case whe e he chip mus handle scenes wi h di e en ligh ing condi ions, some kind o au o-ze o s a egy mus be de ised o gene a e ITH app oxima ely equal o he a e age o he pho osenso cu en s o e he whole a ay. Au ho ized licensed use limi ed o: Uni e sidad de Se illa. Downloaded on Ma ch 20,2020 a 15:33:44 UTC om IEEE Xplo e. Res ic ions apply. ESPEJO e al.: SMART-PIXEL CELLULAR NEURAL NETWORKS IN ANALOG CURRENT-MODE CMOS TECHNOLOGY 899 (b) Fig. 6. cu en ou pu ; (b) au o-ze o ci cui y. Th eshold ci cui y o pho osenso s. (a) Bias-shi ing o double- ail A simple, ye con enien , au o-ze o s a egy uses ou ex a ansis o s a each senso . Fig. 6(b) shows he schema ic o a senso including he au o-ze o ci cui y. All p-channel ansis o s ha e equal size; he same applies o n-channel ansis o s. The low-impedance node labelled SUM is a global node, common o all pixels. No e ha he cu en I, a he c h pho osenso is eplica ed wice. One o he eplicas in e aces he p ocessing ci cui y, while he o he is oo ed o he global- node SUM, and agg ega ed o he emaining senso cu en s. Thus, calcula ion o he cu en ITH h ough ansis o MTH ob ains he ollowing whe e gOp is he ou pu conduc ance o he p-channel ansis o , gmn is he ansconduc ance o he n-channel ansis o , and N he numbe o pixels. Fo simplici y (6) assumes equal ansconduc ances and conduc ances o all pixels. The i s ac o in (6) e lec s he cu en di ision pe o med a node SUM, while he second co esponds o he gain o he mi o o med by he pa allel combina ion o he Msuh ansis o s and MTH. Assuming gmn >> go,, (6) gi es ITH equal o he a e age o he pho osenso cu en s, and he ligh - h eshold is au oma ically adjus ed o he a e age illumina ion. Fig. 7. sma pixel. Concep ual block diag am o he p ocessing ci cui y o a CNN V. PROCESSING CIRCUITRY A. Basic Ci cui Building Blocks Fig. 7 is a block diag am o he p ocessing ci cui y o he c h uni in a sma -pixel CNN chip, acco ding o (2). This igu e shows a co e in eg a o wi h nonlinea losses and an ou pu s uc u e o gene a e weigh ed eplicas o he c h inpu U, and s a e x,, o ansmission o he neighbo sma -pixels. The in eg a o is d i en by weigh ed eplicas o he inpu and s a e signals o he sma pixels in he neighbo hood N,(c), plus an o se e m, ob aining he ollowing signal o d i e he co e in eg a o Jc( ) = D, + {Acdxd( ) + Bcd~d}. (7) Cu en -mode p o ides a con enien choice o ealize he p ocessing ci cui y o sma -pixel CNN's. On one hand, i enables di ec in e ace wi h he senso s, whose ou pu s a e cu en s. On he o he , cu en summa ion a he in eg a o inpu node is di ec ly achie ed by ou ing signals o a common node. Finally, analog ope a o s in ol ed in Fig. 7 (weigh ed- eplica ion, in eg a ion, and limi a ion) a e ealized by simple cu en mi o ci cui s. Fig. 8(a) ealizes he co e in eg a o . Inpu cu en J,*( ) is an unno malized e sion o Jc( ) in (7), wi h no maliza ion ac o IQ:J:(~) = IQJc( ). Ou pu cu en z ( ) is he co esponding unno malized e sion o x, ( ). The pa allel combina ion o he diode-connec ed inpu ansis o A41 and capaci o C yields a ime cons an = GIs,, whe e gm is he ansconduc ance pa ame e o MI. On he o he hand, no e ha cu en x: canno swing beyond he alues o he cu en sou ces which d i e he common ou pu node o ansis o s and M3-meaning ha IxZI < IQ. Thus, analysis o his ci cui esul s in: dE A', (c) as equi ed o ealize (2), and whe e g(.) is he unc ion de ined in (3) wi h m i CO. In p ac ice 7 does no emain cons an , bu a ies wi h inpu cu en le el. Howe e , mos p ocessing asks ole a e his a ia ion wi h no deg ada ion o he ne wo k unc ionali y [9]. Fig. 8(b) shows a ci cui o ealize he ou pu s uc u e o Fig. 7 om ol age V,, and using he basic cu en mi o Au ho ized licensed use limi ed o: Uni e sidad de Se illa. Downloaded on Ma ch 20,2020 a 15:33:44 UTC om IEEE Xplo e. Res ic ions apply. 900 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 29, NO. 8, AUGUST 1994 ""'I6 ;d( 6 Cd'IQ I 1 (b) Fig. 8. sma pixels. (a) Co e in eg a o ; (b) ou pu s uc u e. Cu en -mode ci cui blocks o he p ocessing ci cui y o CNN p inciple o weigh ed eplica ion [23]. No e ha Fig. 8(b) con ains wo di e en subs uc u es o co e each possible sign o he weigh Acd. Posi i e weigh s a e ob ained using a single ou pu ansis o whose geome y ac o is lAcdl imes ha o ansis o MI. Thus, a cu en A,,x; is sou ced o he ou pu node. Nega i e weigh s equi e an addi ional cu en mi o wi h uni y weigh o sign in e sion. B. Some Ci cui Design Issues The ollowing is a b ie commen o dominan nonideali ies encoun e ed in he p ac ical implemen a ion o sma -pixel CNN chips and associa ed ci cui s. I) Cu en Gain E o : A majo sou ce o e o is he ini e a io o he inpu conduc ance yi, o he ou pu conduc ance yo o he cu en mi o s, which causes cu en gain e o due o spu ious cu en di ision. I is especially signi ican a he inpu node o he in eg a o , whe e he gain e o , is gi en app oxinia ely by [9]: egula ed mi o s, o a combina ion o bo h mus be used [23]. In pa icula , analysis shows ha he cascode mi o o Fig. 9(a) ob ains alues o g0/gi L se e al o de s o magni ude lowe han ha o single mi o s, wi h smalle a ea occupa- ion. Chips epo ed in Sec ion VI a e ealized using hese mi o s, and sized o handle he whole inpu cu en ange wi h minimum dis o ion and smalles possible de ices. Fo mi o s biased by a cu en IQ, we ob ain he ollowing sizing equa ions whe e k, = pC0,/2. V,, is he h eshold ol age, and Vc.4~ is he cascode ol age, which can be gene a ed as shown in Fig. 9(b). We assume he same geome ies W, and Ln o all n-channel ansis o s in he cascode mi o . W alues o la ge cu en s (associa ed o weigh ed eplica ion) a e calcula ed by imposing he cons ain ha all ansis o s ha e equal cu en densi y. Al e na i ely, o a gi en aspec a io W,/L,, (10) es ablishes a bound o he maximum bias cu en o he de ice. 2) Misma ch e o ; a ea, powe and eliabili y: T ansis o geome y a ios, s a ic gain e o due o nonnull go/gzn, and powe dissipa ion inc ease wi h IQ. Hence, a bias cu en as small as possible should be chosen. The issue is o iden i y he minimum easible ail cu en alue. A lowes limi is ce ainly es ablished by leakage, which in ou case is inc eased by ligh e ec s. Howe e , a mo e es ic i e bound exis s due o MOS ansis o misma ch and Ea ly ol age (V,) deg ada ion wi h channel leng h. Misma ch is p oduced mainly by a ia ions o V, and /3 = pCo,W/L, whose s anda d de ia ions ~(VT) and o(@)/p o de ices wi h equal layou show a componen in e sely p opo ional o he squa e oo o he channel a ea, and ano he p opo ional o he dis ance be ween de ices [24]. Howe e , in he echnology used and o ansis o pai s close han abou 2.5 mm, he dis ance-dependen componen is negligible o de ices wi h channel a ea o less han 100 piii2 [24]-la ge han he alues ob ained using (10) o bias cu en below -50 pA and channel leng hs o 3.2 /mi. Lowe channel leng hs ha e no been conside ed o se e al easons, like sho -channel e ec s, ea ly- ol age deg ada ion, and in- c eased misma ch e ec s due o he associa ed low channel a eas. In addi ion, lowe ansis o geome ies do no esul in app eciable a ea educ ions due o he minimum con ac size (4 pm wi h su ounding me al and di usion in he echnology used). Ano he impo an conside a ion is ha o a gi en o(V'+) and g(/?)//?, he a io a(I)/I in MOS ansis o s ope a ing in s ong in e sion and a e pinch-o has an in e se dependency wi h iigs - V,. This means ha once geome ies ha e been se o achie e accep able misma ch le els, bias cu en canno be EM (N+ I)* (9) Sin whe e N deno es he numbe o mi o s d i ing his node-up o 18 o empla es wi h no ze o en ies on a ec angula g id ne wi h uni y neighbo hood pa ame e . Fo imp o ed go/gin igu es wi h sho channel de ices, cascode mi o s, dec eased oo a below he bound gi en by (lo), since his would p oduce a low ugs ol age a he bias poin , wi h he co esponding la ge o(I)/I. Hence, misma ch conside a ions es ablish bounds o bo h minimum a ea and powe ends. 3) Ligh e ec s on he p ocessing ci cui y: Op ical image acquisi ion o ces he p ocessing ci cui y o be exposed o Au ho ized licensed use limi ed o: Uni e sidad de Se illa. Downloaded on Ma ch 20,2020 a 15:33:44 UTC om IEEE Xplo e. Res ic ions apply. ESPEJO e al.: SMART-PIXEL CELLULAR NEURAL NETWORKS IN ANALOG CURRENT-MODE CMOS TECHNOLOGY YO1 - W L - ..... . . . .... lin io + W L - - W L - W’ L W L - ‘W - L W L - (b) Fig. 9. and e e ence ci cui y, (b) cascode ol age gene a ion. CMOS bias-shi ed mi o s and biasing de ices. (a) Cascode mi o ligh , which esul s in an inc ease o he leakage cu en s a he e e se-biased subs a e-di usion junc ions. Uni a y bias cu en s mus be su icien ly la ge in o de o neglec his e ec . Also, MOS h eshold ol age depends on ligh in ensi y, inc easing he misma ch e ec on cu en mi o s and sou ces. Cu en mi o ansis o s a e commonly placed nea by, and hence ligh -in ensi y g adien s ha e a educed e ec . On he con a y, cu en sou ces in di e en cells, biased by common global ol ages, can exhibi la ge dispe sions. The ole ance o a pa icula applica ion o a ia ions in he uni a y bias cu en mus be e alua ed in gene al, and local e e ences should be used when equi ed. Fig. 10. Mic opho og aph o he DCC p o o ype. VI. EXPERIMENTAL RESULTS A. 16 x 16 DCC P o o ype The ollowing measu emen s we e aken om a 16 x 16 sma -pixel CNN chip in ended o ho izon al connec ed componen de ec ion [15] (see Fig. 2)-a basic p ep ocessing s ep o pa e n ecogni ion. Fig. 10 shows a mic opho og aph o he p o o ype, which in addi ion o he sma -pixel a - ay con ains bounda y cells, ou pu bu e s, bias s ages, and some digi al con ol ci cui y o he ou pu image download- ing p ocess. The dimensions o he co e a ay a e 1890 x 1530 pm2, and i s powe dissipa ion is 27 mW. The o al chip dimensions, including he bonding pads, a e 2480 x 2500 p n2, wi h a o al powe dissipa ion o 42 mW and a o al o 24 pins. Fig. 11 shows he schema ic and layou o one elemen a y uni . Uni dimensions a e 118 x 96pm2, which include he senso and associa ed egula ion ci cui y (-30% o he a ea), he p ocessing ci cui y, an addi ional cu en eplica ion o ou pu e alua ion, and all equi ed ou ing (cells a e connec ed o each o he by abu men ). Senso is ealized wi h wo minimum-size p-n-p de ices in a Da ling on con igu a ion, o p oduce a b igh -cu en unde labo a o y ligh ing o abou lpA, la ge enough o he ma ching equi emen s o his applica ion. Cascode s uc u es a e used o bo h cu en sou ces and mi o s, and excep o con ol swi ches, all MOS ansis o s ha e W = 4p n and L = 3.2pm. Powe dissipa ion wi h a 5 V supply and unde en i onmen al ligh in he labo a o y is lO5bW pe cell, uni a y cu en being We ha e ob ained 100% success (ou o 30 ials) o ull de ice le el Mon eca lo simula ion o his chip. These Mon eca lo simula ions a e based on he expec ed a ia ions o he h eshold ol ages VTo and he la ge signal anscon- duc ance p (body e ec pa ame e y in luences only he cascode ansis o s). Global biasing ol ages a e used o cu en e e ence gene a ion, and bias s ages a e included in he simula ion. Dispe sion due o misma ch among ansis o s o di e en cu en sou ces did no p oduce c i ical esul s. Thus, global biasing is a ai app oach o his applica ion. Fig. 12(a) illus a es he chip measu emen se up and Fig. 12(b) shows i e inpu images (le column) and he IQ = 2pA. Au ho ized licensed use limi ed o: Uni e sidad de Se illa. Downloaded on Ma ch 20,2020 a 15:33:44 UTC om IEEE Xplo e. Res ic ions apply. 902 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 29, NO. 8, AUGUST 1994 SET! VPB!VCP!VCN! Objec D.U.T. (Film Nega i e) (Chip) Wo k S a ion Digi al De ice Tes ing Equipmen Regulable Focusing Ligh Sou ce (a) VTH! START! (a) io 6 /ai* 9 .*'( a 111" (b) Fig. 11. elemen a y uni o DCC p o o ype. (a) Schema ic ( e e o ex o dimensions) and (b) layou o measu ed ou pu images ( igh column). The p o o ype was exhaus i ely es ed wi h 1200 inpu images. Fig. 12(c) con ains he ou pu wa e o ms obse ed om he cells in a pa icula ow o he a ay du ing a p ocessing example. The inpu pixels a e displayed a he le side column, while he ou pu ones a e a he igh . The signals display he measu ed ansien e olu ion o he ou pu o he cells in he ow. Measu ed con e gence ime is 1.6~~. Ou pu image downloading equi es 8ps, using a 2 MHz digi al clock equency o he se ial downloading p ocess. Ci cui ope a ion emains co ec , wi h no speed deg ada ion, i he ol age supply is educed om he nominal 5 V down o 2.7 V. This is ano he posi i e consequence o using cu en -mode echniques. 1w I I B. 16 x 16 Radon T ans o m P o o ype This p o o ype pe o ms he Radon T ans o m [25] o 16 x 16 pixels inpu images. This chip accep s elec ical, as well as op ical, inpu . The p ocessing ci cui y is based on a modi ied e sion o (2), whe e ime has been disc e ized, and he nonlinea i y is ha d x,(n + 1) ime 1, o D, + {Acd~d(n) + Bcdud} > 0. (C) d€N,(c) Fig. 12. (a) Measu emen se up, (b) i e inpu images and he ou pu measu ed om DCC, and (c) measu ed ansien esponse on a ow o cells. - 1, o he wise (1 1) Au ho ized licensed use limi ed o: Uni e sidad de Se illa. Downloaded on Ma ch 20,2020 a 15:33:44 UTC om IEEE Xplo e. Res ic ions apply. ESPEJO e al.: SMART-PIXEL CELLULAR NEURAL NETWORKS IN ANALOG CURRENT-MODE CMOS TECHNOLOGY 903 FOT!LOAD!LOAD! VTH! CI! c2! Fig. 14. Radon T ans o m p o o ype. Fi e inpu images and he co esponding ou pu measu ed om he (b) Fig. 13. Radon T ans o m p o o ype. (a) Schema ic o elemen a y uni and (b) mic opho og aph o he Also, his applica ion equi es signal-dependen weigh s [26]. In pa icula , he weigh s o he con ibu ions going om a pa icula cell c o i s neighbo s depend on 5,. The comple e se o CNN coe icien s can be desc ibed using unidimensional empla es as ollows i (z, 2 0) A = { [l 0 01, i (zc < 0) [0 0 11, B=[OOO] D=O (12) which e lec he scaling ac o s applied o he con ibu ions o a pa icula cell o i s neighbo s. Fig. 13(a) shows a simpli ied schema ic o a cell, which uses pass ansis o s o ealize he delay equi ed in (1 1) and a high- esolu ion cu en compa a o [27] o he ha d nonlinea i y. The design echnique and he algo i hm used in his ci cui y is desc ibed in de ail in [9]. Fig. 13(b) shows a mic opho og aph o he p o o ype. Cell dimensions a e 121 pm x 124 pm, and he powe dissipa ed by each cell is 1 mW-signi ican ly la ge han o he DCC due o he ci cui y used o implemen he ha d nonlinea i y. The sys em con ains a numbe o blocks loca ed in he pe iphe y o he cell a ay, like ou pu bu e s, bias s ages, and digi al con ol ci cui y dedica ed o he uploading and downloading p ocesses. This addi ional ci cui y, oge he wi h he bonding pads, esul in a o al sys em a ea o 2670pm x 2680 pm, and a o al sys em dissipa ion o 330 mW. The chip equi es a o al o 43 pins. This numbe is signi ican ly highe han ha o he p e ious p o o ype due o 16 inpu pads used o elec ical inpu image uploading. Using a 2 MHz digi al clock equency, image p ocessing ime is 8 ps. The se ial downloading p ocess also equi es 8 ps. As an example, Fig. 14 shows i e inpu images (le column) and he co esponding ou pu images measu ed om he chip ( igh column). The comple e es o he p o o ype in ol ed 1200 images. VII. CONCLUSIONS Summa izing, his pape has ou lined a basic model and some design issues ela ed o a me hodology o design CNN sma -pixel chips in digi al CMOS p ocesses, and has p e- sen ed measu emen s om wo wo king p o o ypes in a 1.6- pm n-well CMOS echnology. One calcula es he numbe o connec ed pieces (DCC) o an inpu image in he ho izon al di ec ion, and he o he e alua es he Radon T ans o m o an inpu image. The DCC chip ob ains a densi y o -89 sma -pixels pe mm2 (each including senso y, egula ion and p ocessing ci cui y), wi h a powe consump ion o 105 pLw pe sma pixel and image p ocessing imes below 2ps. A ea and speed igu es o he RT chip a e simila . Al hough powe dissipa ion is la ge o his p o o ype, his can be co ec ed wi h a ca e ul design o he cu en compa a o [27]. As compa ed o p e ious CNN implemen a ions, he p oposed echnique makes he equi ed syne gy be ween sensing and p ocessing, and signi ican ly imp o es a ea and speed powe igu es. In pa icula , when compa ed o p e ious Au ho ized licensed use limi ed o: Uni e sidad de Se illa. Downloaded on Ma ch 20,2020 a 15:33:44 UTC om IEEE Xplo e. Res ic ions apply.