Session 12Innovations from Outside the (ISSCC’s) Box.pdf

《Session 12Innovations from Outside the (ISSCC’s) Box.pdf》由會員分享，可在線閱讀，更多相關《Session 12Innovations from Outside the (ISSCC’s) Box.pdf（14頁珍藏版）》請在三個皮匠報告上搜索。

1、Session 12 Overview:Innovations from Outside the(ISSCCs)Box TECHNOLOGY DI RECTI ONS SUBCOMMI TTEEThere have been numerous new and i mportant developments outsi de the ci rcui t communi ty that can have profound i mpact on the soli d-state ci rcui ts soci ety,ei ther through emergence of new appli ca

2、ti ons,or,vi a offeri ng of new platforms for processi ng,communi cati ons,and sensi ng.Thi s i nvi ted sessi on seeks to provi de exposure to a few of such developments for the I SSCC audi ence who may not have day-to-day i nteracti ons wi th such subjects i n the hope of i nspi ri ng new ways of t

3、hi nki ng i n ci rcui t desi gn and collaborati on opportuni ti es.Session Chair:Kaushi k Sengupta Pri nceton Uni versi ty,Pri nceton,NJ Session Co-Chair:Fi rooz Aflatouni Uni versi ty of Pennsylvani a,Phi ladelphi a,PA 224 2025 I EEE I nternati onal Soli d-State Ci rcui ts ConferenceI SSCC 2025/SES

4、SI ON 12/I NNOVATI ONS FROM OUTSI DE THE(I SSCCS)BOX/OVERVI EW979-8-3315-4101-9/25/$31.00 2025 I EEE8:25 AM 12.2 p-Circuits:Neither Digital nor Analog Supri yo Datta,Purdue Uni versi ty,West Lafayette,I N I n Paper 12.2,Purdue presents a new paradi gm of computati on,whi ch i s nei ther analog nor d

5、i gi tal,ti tled the p-ci rcui t.The p-ci rcui t i s a stochasti c computi ng element,and not a determi ni sti c Boolean functi on of the i nputs.The output of the devi ce i s a random bi nary vari able whose probabi li ty of bei ng 1 i s gi ven by an analog functi on of the i nputs.Uti li zi ng suc

6、h devi ces,new computi ng archi tectures can emerge that can enable orders of magni tude reducti on i n energy for computi ng,opti mi zati on and learni ng problems.8:50 AM 12.3 Reversing Scattering to Perform Deep-Tissue Optical I maging and the Current Need for a Suitable Optoelectronic Solution C

7、hanghuei Yang,Cali forni a I nsti tute of Technology,Pasadena,CA I n paper 12.3,Caltech presents methodologi es for opti cal i magi ng through deep ti ssues.The hi gh opti cal turbi di ty of bi ologi cal ti ssues prevents sci enti sts and cli ni ci ans from performi ng deeply penetrati ng hi gh reso

8、luti on opti cal i magi ng through humans and ani mal models ali ke.The paper presents si gni ficant advances i n wavefront shapi ng that have allowed ti me-reversi ng scatteri ng and enabli ng focusi ng of li ght deep through bi ologi cal ti ssue for i magi ng and other appli cati ons.The paper als

9、o presents opportuni ti es for i ntegrated optoelectroni c soluti ons for transformati ve i mpact.9:15 AM 12.4 Skin-I nspired Electronics:An Emerging Sensing and Computing Platform Margheri ta Ronchi ni,Stanford Uni versi ty,Stanford,CA Aarhus Uni versi ty,Aarhus,Denmark I n Paper 12.4,Stanford pres

10、ents advances i n e-ski n technologi es for developi ng electroni c materi als i nspi red by ski n s properti es,such as stretchabi li ty,self-heali ng abi li ty and bi odegradabi li ty.Thi s class of new acti ve materi als can enable ski n-li ke i ntegrated ci rcui ts for neuromorphi c si gnal proc

11、essi ng to generate spi ke-trai n si gnals.The paper presents broad opportuni ti es for such ski n-li ke sensors and i ntegrated ci rcui ts for appli cati ons i n medi cal devi ces,roboti cs and wearable electroni cs.I SSCC 2025/February 18,2025/8:00 AM225 DI GEST OF TECHNI CAL PAPERS 8:00 AM 12.1 C

12、ircuits that Solve Optimization Problems by Exploiting Physics I nequalities Eli Yablonovi tch,Uni versi ty of Cali forni a,Berkeley,CA I n Paper 12.1,Berkeley presents an overvi ew of pri nci ple of computati on for arti fici al I ntelli gence and opti mi zati on based on Onsager s Pri nci ple of m

13、i ni mum entropy generati on,referred to as Onsager Computi ng as opposed to conventi onal Von Neumann Computi ng.Electri cal Onsager computers i s an emergi ng computi ng paradi gm wi th potenti al for drasti c reducti on of ti me and energy to soluti on compared to conventi onal machi nes.The pape

14、r presents several i nequali ti es i n physi cs as opportuni ti es to be uti li zed for energy-effici ent computi ng.12226 2025 I EEE I nternati onal Soli d-State Ci rcui ts ConferenceI SSCC 2025/SESSI ON 12/I NNOVATI ONS FROM OUTSI DE THE(I SSCCS)BOX/12.1979-8-3315-4101-9/25/$31.00 2025 I EEE12.1 C

15、ircuits that Solve Optimization Problems by Exploiting Physics I nequalities Eli Yablonovi tch,Qi xi n Feng,Sri Vadlamani,Patri ck Xi ao Uni versi ty of Cali forni a,Berkeley,CA Opti mi zati on i s vi tal to Engi neeri ng,Arti fici al I ntelli gence,and to many areas of Sci ence.Mathemati cally,we u

16、sually employ steepest-descent,or other di gi tal algori thms.For example,Deep Learni ng i s an opti mi zati on problem,Fi g.12.1.1.Everyday appli cati ons of opti mi zati on range from aerodynami c desi gn of vehi cles by flui d mechani cs and physi cal stress opti mi zati on of bri dges i n ci vi

17、l engi neeri ng;to scheduli ng of ai rli ne crews and routi ng of deli very trucks i n operati ons research.Furthermore,opti mi zati on i s also i ndi spensable i n machi ne learni ng,rei nforcement learni ng,computer vi si on,and speech processi ng.Gi ven the preponderance of massi ve datasets and

18、computati ons today,there has been a surge of acti vi ty i n the desi gn of hardware accelerators for neural network trai ni ng and i nference.But,every i nequali ty i n Physi cs,performs opti mi zati on i n the normal course of dynami cal evoluti onfor free.Nature provi des us wi th the followi ng

19、opti mi zati on pri nci ples:1.The Pri nci ple of Least Acti on;2.The Vari ati onal Pri nci ple of Quantum Mechani cs,see Fi g.12.1.2(top-left);3.The Pri nci ple of Mi ni mum Entropy Generati on,see Fi g.12.1.2(ri ght);4.The Fi rst Mode to Threshold method,see Fi g.12.1.2(bottom-left);5.The Pri nci

20、ple of Least Ti me,see Fi g.12.1.3;6.The Adi abati c Evoluti on method,see Fi g.12.1.4;7.Quantum Anneali ng.I n effect,Physi cs can provi de machi nes whi ch solve di gi tal opti mi zati on problems much faster than any conventi onal computer.The pri nci ple of Least Acti on i s the most fundamental

21、 pri nci ple i n physi cs.Newton s Laws of Mechani cs,Maxwell s Equati ons of Electromagneti sm,Schrdi nger s equati on i n Quantum Mechani cs,and Quantum Fi eld Theory can all be i nterpreted as mi ni mi zi ng a quanti ty called Acti on.For the speci al case of li ght propagati on,thi s reduces to

22、the pri nci ple of Least Ti me,as shown i n Fi g.12.1.3.A conservati ve system wi thout fri cti on or losses evolves accordi ng to the pri nci ple of Least Acti on.The fundamental equati ons of physi cs are reversi ble.A consequence of thi s reversi bi li ty i s the Li ouvi lle Theorem whi ch states

23、 that volumes i n phase space are left unchanged as the system evolves.Contrary-wi se,i n both a computer,and an opti mi zati on solver,the goal i s to have a speci fic soluti on wi th less uncertai nty or a smaller zone i n phase space than the i ni ti al state,an entropy cost first speci fied by L

24、andauer&Bennett.Thus,some degree of i rreversi bi li ty,or energy cost,i s needed,speci fied by the number of di gi ts i n the answer i n the Landauer/Bennett analysi s.An algori thm has to be desi gned and programmed i nto the reversi ble system(ei ther vi a software or vi a hardcodi ng of hardware

25、)to effect the reducti on i n entropy needed to solve the opti mi zati on problem.Comi ng up wi th a fast algori thm for NP-hard problems i s sti ll an open problem i n the field of reversi ble computi ng,whi ch i ncludes quantum computi ng.Thi s would requi re an energy cost,but not necessari ly a

26、requi rement for conti nuous power di ssi pati on.We look forward to computer sci ence breakthroughs that would allow the pri nci ple of Least Acti on to address unsolved problems.An alternati ve approach to computi ng would i nvolve physi cal systems that conti nuously di ssi pate power,ai di ng i

27、n the contracti on of phase space toward a final soluti on.Whi le exceedi ng the Landauer li mi t,such systems mi ght have the advantage of speed and si mpli ci ty.Thi s bri ngs us to the pri nci ple of Least Power di ssi pati on.“Mi ni mum Heat Generati on”i n the form of bi stable electri cal or o

28、pti cal ci rcui ts i s parti cularly adaptable 1 toward offeri ng di gi tal Opti mi zati on.For example,we provi de the electri cal ci rcui t whi ch can address the challengi ng I si ng problem,bi nary magnet energy mi ni mi zati on(Fi gs.12.1.5 and 12.1.6).Si nce Onsager,2(1931,Nobel Pri ze 1968)i

29、ntroduced the Pri nci ple of Mi ni mum Entropy Generati on we call thi s Onsager Computi ng,as opposed to conventi onal Von Neumann Computi ng.I n 1,we showed that many of the schemes for Physi cs-Based-Opti mi zati on,reduce to the Onsager Pri nci ple.Electri cal Onsager Computers run 10000 ti mes

30、faster have 10000 ti mes less energy-to-soluti on,than conventi onal machi nes.Furthermore,opti cal Onsager machi nes provi de an addi ti onal 1000 ti mes i ncrease i n speed.The appli cati on of Onsager s Pri nci ple i n ordi nary ci rcui ts i s parti cularly favorable toward quadrati c or bi li ne

31、ar Fi gures-of-Meri t,si nce heat generati on i n ci rcui ts i s I2R=(I1+I2+Ii+)2R,where the current Ii i s summed over every branchi i n the ci rcui t.The I si ng problem of magneti c energy mi ni mi zati on i n a group of magnets M,i s i somorphi c to the Onsager Pri nci ple,si nce the magneti c c

32、oupli ng energi es are of the bi li near form,MiMj,the same as the heat generati on terms 2IiIjR.I n due course,we may learn how to use each of these seven Physi cs pri nci ples to perform opti mi zati on.Let us consi der the pri nci ple of mi ni mum entropy generati on i n di ssi pati ve physi cal

33、systems,such as resi sti ve electri cal ci rcui ts.I t was shown by Onsager 2,that the equati ons of li near systems,li ke resi stor networks,can be re-expressed as the mi ni mi zati on pri nci ple of a functi on f(i1,i2,in)for currents i n vari ous branches of the resi stor network.For a resi stor

34、network,the functi on f contai ns the power di ssi pati on,or entropy generati on.By re-expressi ng a meri t functi on i n terms of power di ssi pati on,the ci rcui t i tself wi ll find the mi ni mum of the meri t functi on,or mi ni mum power di ssi pati on.Opti mi zati on i s generally accompani ed

35、 by constrai nts.For example,perhaps the constrai nt i s that the final answers must be restri cted to be 1.Such a di gi tally constrai ned opti mi zati on produces answers compati ble wi th any di gi tal computer.There has been a seri es of machi nes created i n the physi cs and engi neeri ng commu

36、ni ty to devi se physi cs-based engi nes for the I si ng problem.The I si ng challenge i s to find the mi ni mum energy configurati on of a large set of magnets.I t s a very hard problem even when the magnets are restri cted to only two ori entati ons,North pole up or down.Our mai n i nsi ght i n th

37、i s paper i s that most of these I si ng solvers use hardware based on the Pri nci ple of Mi ni mum Entropy generati on.The natural evoluti on of these machi nes i s toward a good,low-power-di ssi pati on final state.Further,almost all of them i mplement the well-known Lagrange Multi pli ers method

38、for constrai ned opti mi zati on.We survey the mi ni mi zati on pri nci ples of physi cs and the i mportant opti mi zati on algori thms deri ved from them.These physi cal opti mi zati on machi nes are i ntended to converge to opti ma that are agnosti c to i ni ti al condi ti ons.By agnosti c to i ni

39、 ti al condi ti ons,we mean systems that converge to the global opti mum,or a good local opti mum,i rrespecti ve of the i ni ti al poi nt for the search.Much early work was by Yamamoto.These entropy generati ng machi nes range from coupled Opti cal Parametri c Osci llators,to RLC electri cal ci rcui

40、 ts,to coupled exci ton-polari tons,and si li con photoni c coupled arrays.These types of machi nes have the advantage that they solve di gi tal problems i n the analog domai n,whi ch can be orders-of-magni tude faster and more energy-effici ent than conventi onal di gi tal chi ps that are li mi ted

41、 by latency and the energy cost.But then the analog ci rcui t provi des a di gi tal answer,li ke a fli p-flop.Wi thi n the framework of these di ssi pati ve machi nes,constrai nts can be readi ly i ncluded.I n effect,these machi nes perform constrai ned opti mi zati on equi valent to the techni que

42、of Lagrange multi pli ers.We i llustrate thi s connecti on by surveyi ng 7 physi cally di sti nct machi nes and show that each mi ni mi zes entropy generati on i n i ts own way,subject to constrai nts;correspondi ng to Lagrange multi pli er opti mi zati on.I n effect,Onsager machi nes perform local

43、steepest descent i n the entropy generati on rate.They can become stuck i n local opti ma.At the very least,they perform a rapi d search for local opti ma,thus reduci ng the search space for the global opti mum.These machi nes are also adaptable toward searchi ng i n an expanded phase space,and othe

44、r techni ques for approachi ng a global opti mum.Karp 3 showed that many famous Computer Sci ence problems have quadrati c or bi li near Fi gures-of-Meri t,and that i f the magneti c I si ng problem could be solved,then many of the other NP-hard problems i n Computer Sci ence could also be solved.No

45、netheless,there are many computati onal challenges that do not reduce to quadrati c Fi gures-of-Meri t,for whi ch ci rcui t desi gn i s easy,and for whi ch there are many examples.The challenge i s to desi gn the non-obvi ous ci rcui ts.One such non-obvi ous ci rcui t i s to i nvert large matri ces.

46、Mi ller 4 showed an opto-electroni c ci rcui t whi ch i mplements Gaussi an eli mi nati on usi ng tunable 22 couplers.Mi ller s approach i s to mi ni mi ze the power i n an unused output port,that would have been wasted.But mi ni mi zi ng the wasted power i s exactly what Onsager calls for.Of course

47、,every electroni c ci rcui t,automati cally fulfills Onsager s Pri nci ple,so the desi gn challenge remai ns.Figure 12.1.1:Deep learning is an optimization problem.Figure 12.1.2:The variational principle of Quantum Mechanics adjusts the wave function until it has the minimum.energy.Circuit showing O

48、nsager principle of least power dissipation/least entropy production.First-to-gain-threshold mechanism in a multi-mode laser finds the least lossy mode.Figure 12.1.3:The lifeguards path refracts as it enters the lower speed medium,the same as Snells Law.Figure 12.1.4:I n the Adiabatic Method the gro

49、und state remains the ground state even as the Hamiltonian goes from simple to hard.I n Quantum Mechanics the corresponding mechanism is called“Quantum Annealing”.Figure 12.1.5:An I sing machine based on bistable LC parametric oscillators connected together with ferromagnetic OR anti-ferromagnetic c

50、oupling.Figure 12.1.6:The colored circle represents the wiring connections between parametric oscillators;Red representing Ferromagnetic wiring,and Blue representing Anti-Ferromagnetic wiring.The answer emerges after about 0.4 microseconds,which is about 40 clock periods.I SSCC 2025/February 18,2025

51、/8:00 AM227 DI GEST OF TECHNI CAL PAPERS 12 2025 I EEE I nternati onal Soli d-State Ci rcui ts ConferenceI SSCC 2025 PAPER CONTI NUATI ONS AND REFERENCES979-8-3315-4101-9/25/$31.00 2025 I EEERe fe re nc e s:1 Sri K.Vadlamani,T.Patri ck Xi ao,and Eli Yablonovi tch,“Physi cs successfully i mplements l

52、agrange multi pli er opti mi zati on,”Proceedi ngs of Nati onal Academy of Sci ences USA,vol.117,no.43,pp.26639-26650(2020).2 L.Onsager,“Reci procal relati ons i n i rreversi ble processes.I I,”Physi cal Revi ew,vol.38,pp.2265-2279,(1931).3 Ri chard M.Karp,“Reduci bi li ty among combi natori al prob

53、lems,”Plenum Press,pp.85-103(1972).4 D.A.B.Mi ller,“Self-configuri ng uni versal li near opti cal component,”Photoni cs Research,vol.1,pp.1-15(2013).228 2025 I EEE I nternati onal Soli d-State Ci rcui ts ConferenceI SSCC 2025/SESSI ON 12/I NNOVATI ONS FROM OUTSI DE THE(I SSCCS)BOX/12.2979-8-3315-410

54、1-9/25/$31.00 2025 I EEE12.2 p-Circuits:Neither Digital nor Analog Mi ng-Che Li1,Archi sman Ghosh1,Ri si Jai swal1,Lakshmi Ani rudh Ghantasala1,2,Behtash Behi n-Aei n2,Shreyas Sen1,Supri yo Datta1 1Purdue Uni versi ty,West Lafayette,I N 2Ludwi g Computi ng,Mi ll Valley,CA Custom i ntegrated ci rcui

55、ts ai m to solve i mportant problems wi th ultra-hi gh effici ency,maki ng use of analog and di gi tal ci rcui ts,wi th well-known trade-offs.Thi s work i s about a new paradi gm whi ch i s nei ther analog nor di gi tal,we call i t a p-ci rcui t See for example,1-12.I n terms of i nputs and outputs,

56、our p-ci rcui ts look li ke di gi tal ci rcui ts(Fi g.12.2.1),thus requi ri ng no ADC s or DAC s.However,the output i s not a Boolean functi on of the i nputs.I t i s a random bi nary vari able whose probabi li ty P(B=1)of bei ng 1 i s gi ven by an analog functi on A of the i nputs whi ch takes on v

57、alues conti nuously between 0 and 1.Building p-circuits:Fi gure 12.2.2 shows how a p-ci rcui t can be bui lt usi ng bui ldi ng blocks each of whi ch takes four i nputs,combi nes them to compute an analog quanti ty A(b1,b2,b3,b4)and generates a bi nary output B wi th P(B=1)=A as shown i n Fi g.12.2.1

58、.Thi s si ngle output i s repli cated four-ways as shown to faci li tate the creati on of a two-di mensi onal array through ti li ng.The ci rcui t operates by sequenti ally updati ng the output of each of the N bui ldi ng blocks(or p-gates)based on i ts current i nputs.Each of the N p-bi ts i s upda

59、ted it(denoti ng i terati ons)ti mes.What problems can we solve wi th thi s two-di mensi onal array?For starters,we can solve Quadrati c Unconstrai ned Bi nary Opti mi zati on(QUBO)problems descri bed by cost functi ons E of the form E=-i bi hi 0.5 i,j(bi Wi j bj)where the i ndi ces(i,j)run over the

60、 si tes of the two-di mensi onal latti ce,and the wei ght matri x Wij i s non-zero only i f i and j are nearest nei ghbors.Opti mi zati on requi res us to find configurati ons b that mi ni mi ze the cost functi on E.A stati sti cal approach to thi s problem i s to generate samples from a probabi li

61、ty di stri buti on functi on(PDF)P e xp(-E(b)so that low E configurati ons appear wi th hi gh probabi li ty.Thi s can be done i f we generate new samples from the exi sti ng sample by modi fyi ng a si ngle p-bi t bk out of the collecti on b such that P(bk=1)=1+exp(-k)(-1)where k=E(bk=0)E(bk=1)=hk+j(

62、Wkj bj).The algori thm i s i mplemented by repeatedly performi ng a core operation,consi sti ng of:(1)looki ng at n bi nary i nputs(where n depends on the number of non-zero elements of Wkj for a gi ven k),and(2)generati ng a random bi nary output followi ng the probabi li ty gi ven above.Several AS

63、I C i mplementati ons 13-25 use si mi lar concepts i n commerci al process.We wi ll now descri be an ASI C i mplementati on usi ng a commerci al 65nm process,whi ch solves a class of QUBO problems 1.ASI C design:Fi gure 12.2.3 shows the system archi tecture:The ASI C,compri si ng a 1,440 p-bi t comp

64、uter,employs a 72 p-gate array as i ts pri mary computi ng uni ts for stochasti c computi ng.Addi ti onally,i t i ncorporates two types of memory:wei ght memory and p-bi t memory.These memori es serve the purpose of provi di ng wei ght values Wkj from the cost functi on E that define the speci fic Q

65、UBO problem.The i terati ve process i nvolves 1,440 p-bi ts,whi ch are updated over 20 cycles.Duri ng each cycle,72 p-bi ts are updated through 72 p-gates.Fi gure 12.2.3(bottom-left)shows the i mplementati on of a p-gate,whi ch takes up to 7 p-bi ts as i nputs along wi th thei r correspondi ng wei g

66、hts.Followi ng the equati on k=hk+j(Wkj bj),the probabi li ty i s calculated usi ng an exponenti al LUT and compared wi th a random number generated by an Xoshi ro128+PRNG,as depi cted i n Fi g.12.2.3(bottom-ri ght).After each cycle,an updated p-bi t i s wri tten back i nto the p-bi t memory.Each i

67、terati on updates all p-bi ts to generate a new sample from the PDF,P e xp(-E(b).ASI C measurement results:Fi gure 12.2.4 shows measured results from our I C.Fi gure 12.2.4(top)shows the soluti on to the problem descri bed i n 1.The convergence i ndex reaches a threshold value after some i terati on

68、s.Fi gure 12.2.4(bottom-left)shows the measured power.The p-bi t computer consumes 328uW acti ve power at 10MHz at 0.5V core voltage,wi th leakage power contri buti ng to an addi ti onal 57.42uW.Notably,approxi mately 70%of thi s power i s consumed by the p-gate array.Energy per operation:Fi gure 12

69、.2.5(left)compares the energy cost of the core operati on di scussed earli er,esti mated for each of four opti ons,namely,(1)CPU,(2)125MHz FPGA,(3)10MHz ASI C and(4)clockless ci rcui t wi th s-MTJ s.We evaluated the energy based on solvi ng a QUBO problem li ke the one descri bed earli er except tha

70、t the latti ce i s 3D and non-rectangular 1 for whi ch each p-bi t looks at 5 i nputs rather than 4.The number of p-bi ts N=1,440,whi le the problem requi red it=25,000 i terati ons.We have also used the same archi tecture to solve other QUBO problems featuri ng di fferent values of(N,it)requi ri ng

71、 di fferent amounts of tota l e ne rgy,but the e ne rgy pe r ope ra tion i s characteri sti c of the hardware used to i mplement i t.Fi gure 12.2.5(ri ght)shows the energy and ti mes for several other i mplementati ons reported i n the li terature,whi ch we di scuss below i n the secti on“How we com

72、pare.”The costli est i mplementati on(Fi g.12.2.5 left)i s on a CPU for whi ch we esti mate uJ per operati on,whi le our ASI C i mplementati on requi res pJ per operati on,whi ch i s si x orders of magni tude smaller.Note that these energy esti mates should be appli cable to any algori thm that can

73、be i mplemented by repeatedly performi ng the same core operati on requi ri ng a p-gate wi th 5 bi nary i nputs,allowi ng us to evaluate the analog functi on si mply usi ng an LUT wi th 25=32 entri es.But i f each p-gate were to look at many more i nputs,the core operati on may consume more energy.B

74、ut how versati le i s our core operati on and the p-gate i mplementi ng i t?We beli eve i t can be used way beyond the QUBO problem descri bed above as i llustrated by the followi ng example from a common generati ve model.Future directions:Fi gure 12.2.6 shows a seri es of transformati ons each of

75、whi ch has a form si mi lar to what we di scussed,namely ck=F(hk+j(Wkj bj)turni ng a set b i nto a set c .Gi ven a speci fic non-li near functi on F,trai ni ng algori thms have been developed that can find appropri ate W,h for each transformati on such that a random i nput i mage i s transformed i n

76、to a recogni zable i mage.However,we cannot i mplement the standard algori thms wi th the p-gates we di scussed si nce,(1)the non-li near functi on F i s usually determi ni sti c whi le our p-gates are probabi li sti c,and(2)F operates on a large number of conti nuous vari ables whi le our p-gate op

77、erates on a small number of bi nary vari ables.We need to change these trai ni ng condi ti ons so that the resulti ng W,h can be i mplemented usi ng p-gates for whi ch Fi g.12.2.6 provi des a proof-of-concept,hopefully a steppi ng stone to more complex generati ve models li ke di ffusi on models or

78、even large language models.Compared to the analog(or multi-bi t di gi tal)gates commonly used i n i mplementi ng deep neural networks(DNN s),the advantage of p-gates i s that they work wi th bi nary i nputs.But these bi nary quanti ti es do not just approximate the analog information,they embed it s

79、tatistically.One mi ght thi nk that we would have to average many samples to get acceptable results,but the results i n Fi g.12.2.6 were obtai ned with just one sample.We beli eve that the true power of p-ci rcui ts li es i n provi di ng a natural platform for such probabi li sti c appli cati ons an

80、d algori thms.How do we compare?I si ng computi ng i s of course not new to thi s communi ty 13-28,and the pri or results i n Fi g.12.2.5(ri ght)show energy costs rangi ng from pJ to J across vari ous desi gns.Note,however,that ti me-to-soluti on(TTS)and energy-to-soluti on depend on the desi gn cho

81、i ces,anneali ng method,and the target I si ng problem.We also note that much previ ous li terature 17,19,22-24 has appli ed i n-memory computi ng techni ques,si gni ficantly reduci ng the energy cost associ ated wi th repeatedly loadi ng varyi ng wei ghts compared to classi c di gi tal desi gns,li

82、ke our ASI C.We propose that the energy cost per operation per spin for a gi ven number of i nputs would be a fai r metri c for compari ng di fferent I si ng solvers.Secondly,i f all p-bi ts along wi th the wei ghts can fit on a chi p then the enti re ci rcui t can operate autonomously wi thout cloc

83、ki ng and thi s can reduce the energy cost si gni ficantly.I ndeed our SPI CE si mulati ons of clockless ci rcui ts usi ng experi mentally benchmarked models for stochasti c magneti c tunnel juncti on s(s-MTJ s)suggest 20W x 50ps fJ per operati on 1,9.Such s-MTJ s have only been demonstrated i n lab

84、oratori es but si mi lar numbers should be achi evable wi th other standard devi ces that can be taped out.The di fficulty,however,i s to scale up thi s clockless operati on to address large problems.Fi nally,we note that thi s paper i s not about a parti cularly energy-effici ent i mplementati on o

85、f I si ng computi ng,of whi ch there are many already.I ndeed we have not yet i mplemented several standard techni ques li ke compute-i n-memory or the ones noted i n Fi g.12.2.5(ri ght).I t i s well-known that speci ali zati on can reduce energy and delay,the challenge i s to make i t broadly appli

86、 cable.Our pri mary message here i s that a very broad vari ety of problems can be addressed through repeated appli cati on of the core bui ldi ng block or p-gate shown i n Fi g.12.2.1 and hence the need to benchmark i ts energy and delay per elementary operation for di fferent i mplementati ons.As

87、we have di scussed,p-gates ari se naturally i n I si ng computi ng but are relati vely unknown i n the context of feed-forward DNN s whi ch are the staple of modern arti fici al i ntelli gence(AI)29,30.We hope this pa pe r will e nc oura ge the use of p-ga te s be yond the na rrow c onfine s of I si

88、ng c omputing.We end by noti ng an i nteresti ng si mi lari ty between p-ci rcui ts and quantum ci rcui ts that i nvolve qubi ts coherently coupled to other qubi ts.A system of N qubi ts has an associ ated wavefuncti on wi th 2N components whose squared magni tude gi ves the PDF.Quantum ci rcui ts m

89、ulti ply the wavefuncti on by a uni tary matri x U,whi le p-ci rcui ts multi ply the PDF by a stochasti c matri x W.Both U a nd W are huge matri ces of si ze 2N x 2N,too large for di rect computati on.But p-ci rcui ts can be used to generate samples effici ently that follow the correct PDF generated

90、 by W.However,they are not very effecti ve i f the problems i nvolve a complex uni tary matri x U,li ke the quantum Fouri er transform(QFT)i n Shor s algori thm,whi ch can be sampled effici ently only wi th quantum computers.Ac knowle dge me nt:We thank Drs.Shuvro Chowdhury and Kerem Camsari for sha

91、ri ng detai ls of thei r worki n 1 on a quantum Monte Carlo problem whose ASI C i mplementati on i s descri bed i nFi g.12.2.3 of thi s paper.Thi s work was parti ally supported by ONR-MURI Grant No.N000142312708,OptNet:Opti mi zati on wi th p-Bi t Networks.Compe ting inte re sts:SD has a financi al

92、 i nterest i n Ludwi g Computi ng.Figure 12.2.1:Building blocks for analog(top left),digital(top right),and p-circuits(bottom).P(B=1)denotes probability of B being 1 rather than 0.Figure 12.2.2:Building a circuit by tiling together 4-input p-gates.I f each p-gate has more inputs then a more elaborat

93、e layout is required.Figure 12.2.3:System level architecture of the prototype p-bit ASI C.An example p-circuits architecture(bottom left).Xoshiro128+PRNG(bottom right).Figure 12.2.4:Measurement results of an example p-bit ASI C fabricated with commercial process.Possibility of improvements.Figure 12

94、.2.5:Energy per operation per spin of the core operation,estimated for each of four options as discussed in text(left).Energy to solution and TTS plot for recent state-of-the-art I sing solvers(right).Figure 12.2.6:A sequence of p-bit layers trained to turn a random initial image into a recognizable

95、 image.Random imageRandom image evolves into a 3x128x128 imageIntermediate layers use p-bitsNote:Image is blurry because only 1 sample was used difor d)nCPUEstimate dhChplatgnignt refeffifdesiper operatinergy?e 2023t lformstspin per on?i y.y.lydirectld darempccobebetet nnon nnocacadmd anbs,aemprobla

96、ingmIsrrent rentdiffediffetst entargegnsigddeant Differesitcu-Maxdnentio emenotnemproblt tcucu-Maxtcub-anMaxt txcucu-Maxtcua-tMaxTFIMtcutx-Maxtcux-Maxtgcu-TS.Tvslution Soto gy rgEners:rssolverIsingor forested llecCo-MaxorkMs wisThia17x18fMax13 Max24 19n23-2226 6CIM DigitalyllFuOSCalog:RAnbased-tchal

97、og:LaAn28tdher x6 aa1Mtarge26152415ults f alescg(loJ)(er spinionoperatper M125Energy AFPGCPU eury owdhCh?MHzGA,?2023et al.?This work?HzMHASIC,s ClocklesCTION:10 MOJE?peEPROJE I SSCC 2025/February 18,2025/8:25 AM229 DI GEST OF TECHNI CAL PAPERS 12 2025 I EEE I nternati onal Soli d-State Ci rcui ts Co

98、nferenceI SSCC 2025 PAPER CONTI NUATI ONS AND REFERENCES979-8-3315-4101-9/25/$31.00 2025 I EEEFigure 12.2.7:I C Micrograph and specification.Re fe re nc e s:1 S.Chowdhury et al.,“Accelerated quantum Monte Carlo wi th probabi li sti c computers.”Communic a tions Physic s 6.1(2023):85.2 Bunai yan,Sale

99、h,Supri yo Datta,and Kerem Y.Camsari.“Hei senberg machi nes wi th programmable spi n ci rcui ts.”Physic a l Re v ie w Applie d 22.1(2024):014014.3 S.Ni azi et al.,“Trai ni ng deep Boltzmann networks wi th sparse I si ng machi nes.”Na ture Ele c tronic s(2024):1-10.4 N.S.Si ngh et al.,“CMOS plus stoc

100、hasti c nanomagnets enabli ng heterogeneous computers for probabi li sti c i nference and learni ng.”Na ture Communic a tions 15.1(2024):2685.5 W.Whi tehead et al.,“CMOS-compati ble I si ng and Potts anneali ng usi ng si ngle-photon avalanche di odes.”Na ture Ele c tronic s 6.12(2023):1009-1019.6 J.

101、Kai ser et al.,“Hardware-aware i n si tu learni ng based on stochasti c magneti c tunnel juncti ons.”Physic a l Re v ie w Applie d 17.1(2022):014016.7 Hassan,Orchi,Supri yo Datta,and Kerem Y.Camsari.“Quanti tati ve evaluati on of hardware bi nary stochasti c neurons.”Physic a l Re v ie w Applie d 15

102、.6(2021):064046.8 R.Fari a et al.,“Hardware desi gn for autonomous bayesi an networks.”F rontie rs in c omputa tiona l ne urosc ie nc e 15(2021):584797.9 R.F.Sutton et al.,“Autonomous Probabi li sti c Coprocessi ng Wi th Petafli ps per Second,”I EEE Ac c e ss,vol.8,pp.157238-157252,2020.10 W.A.Borde

103、rs et al.,“I nteger factori zati on usi ng stochasti c magneti c tunnel juncti ons.”Na ture 573.7774(2019):390-393.11 K.Y.Camsari et al.,“Stochasti c p-bi ts for i nverti ble logi c.”Physic a l Re v ie w X 7.3(2017):031014.12 Behi n-Aei n,Behtash,Vi nh Di ep,and Supri yo Datta.“A bui ldi ng block fo

104、r hardware beli ef networks.”Sc ie ntific re ports 6.1(2016):29893.13 M.Yamaoka et al.,“20k-spi n I si ng chi p for combi nati onal opti mi zati on problem wi th CMOS anneali ng,”2015 I EEE I nt.Solid-Sta te Circ uits Conf.(I SSCC)Dig.Te c h.Pa pe rs,pages 1-3,2015.14 T.Takemoto et al.,“A 2 30k-spi

105、n multi chi p scalable anneali ng processor based on a processi ng-i n-memory approach for solvi ng large-scale combi natori al opti mi zati on problems,”2019 I EEE I nt.Solid-Sta te Circ uits Conf.(I SSCC)Dig.Te c h.Pa pe rs,pages 52-54,2019.15 K.Yamamoto et al.,“Stati ca:A 512-spi n 0.25m-wei ght

106、full-di gi tal anneali ng processor wi th a near-memory all-spi n-updates-at-once archi tecture for combi natori al opti mi zati on wi th complete spi n-spi n i nteracti ons,”2020 I EEE I nt.Solid-Sta te Circ uits Conf.(I SSCC)Dig.Te c h.Pa pe rs,pages 138-140,2020.16 T.Takemoto et al.,“A 144kb anne

107、ali ng system composed of 916kb anneali ng processor chi ps wi th scalable chi p-to-chi p connecti ons for large-scale combi natori al opti mi zati on problems,”2021 I EEE I nt.Solid-Sta te Circ uits Conf.(I SSCC)Dig.Te c h.Pa pe rs,volume 64,pages 64-66,2021.17 Y.Su,H.Ki m and B.Ki m,“CI M-Spi n:A

108、0.5-to-1.2V Scalable Anneali ng Processor Usi ng Di gi tal Compute-I n-Memory Spi n Operators and Regi ster-Based Spi ns for Combi natori al Opti mi zati on Problems,”2020 I EEE I nte rna tiona l Solid-Sta te Circ uits Confe re nc e -(I SSCC),San Franci sco,CA,USA,2020,pp.480-482.18 Y.Su,T.T.-H.Ki m

109、 and B.Ki m,“FlexSpi n:A Scalable CMOS I si ng Machi ne wi th 256 Flexi ble Spi n Processi ng Elements for Solvi ng Complex Combi natori al Opti mi zati on Problems,”2022 I EEE I nte rna tiona l Solid-Sta te Circ uits Confe re nc e (I SSCC),San Franci sco,CA,USA,2022,pp.1-3.19 J.Bae,W.Oh,J.Koo and B

110、.Ki m,“CTLE-I si ng:A 1440-Spi n Conti nuous-Ti me Latch-Based i sli ng Machi ne wi th One-Shot Fully-Parallel Spi n Updates Featuri ng Equali zati on of Spi n States,”2023 I EEE I nte rna tiona l Solid-Sta te Circ uits Confe re nc e (I SSCC),San Franci sco,CA,USA,2023,pp.142-144.20 S.Xi e et al.,“S

111、nap-SAT:A One-Shot Energy-Performance-Aware All-Di gi tal Compute-i n-Memory Solver for Large-Scale Hard Boolean Sati sfiabi li ty Problems,”2023 I EEE I nte rna tiona l Solid-Sta te Circ uits Confe re nc e (I SSCC),San Franci sco,CA,USA,2023,pp.420-422.21 K.Kawamura et al.,“Amorphi ca:4-Repli ca 51

112、2 Fully Connected Spi n 336MHz Metamorphi c Annealer wi th Programmable Opti mi zati on Strategy and Compressed-Spi n-Transfer Multi-Chi p Extensi on,”2023 I EEE I nte rna tiona l Solid-Sta te Circ uits Confe re nc e (I SSCC),San Franci sco,CA,USA,2023,pp.42-44.22 J.Bae,J.Koo,C.Shi m and B.Ki m,“LI

113、SA:A 5764 All-i n-One Repli ca-Spi ns Conti nuous-Ti me Latch-Based I si ng Computer Usi ng Massi vely-Parallel Random-Number Generati ons and Repli ca Equali zati ons,”2024 I EEE I nte rna tiona l Solid-Sta te Circ uits Confe re nc e (I SSCC),San Franci sco,CA,USA,2024,pp.284-286.23 J.Bae,C.Shi m a

114、nd B.Ki m,“e-Chi mera:A Scalable SRAM-Based I si ng Macro wi th Enhanced-Chi mera Topology for Solvi ng Combi natori al Opti mi zati on Problems Wi thi n Memory,”2024 I EEE I nte rna tiona l Solid-Sta te Circ uits Confe re nc e (I SSCC),San Franci sco,CA,USA,2024,pp.286-288.24 J.Song et al.,“A Vari

115、ati on-Tolerant I n-eDRAM Conti nuous-Ti me I si ng Machi ne Featuri ng 15-Level Coeffici ents and Leaked Negati ve-Feedback Anneali ng,”2024 I EEE I nte rna tiona l Solid-Sta te Circ uits Confe re nc e (I SSCC),San Franci sco,CA,USA,2024,pp.490-492.25 Y.-C.Chu et al.,“A Fully I ntegrated Anneali ng

116、 Processor for Large-Scale Autonomous Navi gati on Opti mi zati on,”2024 I EEE I nte rna tiona l Solid-Sta te Circ uits Confe re nc e (I SSCC),San Franci sco,CA,USA,2024,pp.488-490.26 I.Ahmed,P.-W.Chi u and C.H.Ki m,“A Probabi li sti c Self-Anneali ng Compute Fabri c Based on 560 Hexagonally Coupled

117、 Ri ng Osci llators for Solvi ng Combi natori al Opti mi zati on Problems,”2020 I EEE Symposium on VLSI Circ uits,Honolulu,HI,USA,2020,pp.1-2.27 S.Xi e et al.,“I si ng-CI M:A Reconfigurable and Scalable Compute Wi thi n Memory Analog I si ng Accelerator for Solvi ng Combi natori al Opti mi zati on P

118、roblems,”i n I EEE J ourna l of Solid-Sta te Circ uits,vol.57,no.11,pp.3453-3465,Nov.2022.28 Y.Zhou et al.,“A Compute-i n-Memory Anneali ng Processor wi th I nteracti on Coeffici ent Reuse and Sparse Energy Computati on for Solvi ng Combi natori al Opti mi zati on Problems,”I EEE J ourna l of Solid-

119、Sta te Circ uits,vol.59,no.9,pp.3094-3105,Sept.2024.29 Shekhovtsov,Alexander,and Vi ktor Yanush.“Rei ntroduci ng strai ght-through esti mators as pri nci pled methods for stochasti c bi nary networks.”DAGM German Conference on Pattern Recogni ti on.Cham:Spri nger I nternati onal Publi shi ng,2021.30

120、 Li,Yang,et al.“Bi nary-Stochasti ci ty-Enabled Hi ghly Effici ent Neuromorphi c Deep Learni ng Achi eves Better-than-Software Accuracy.”Advanced I ntelli gent Systems 6.1(2024):2300399.230 2025 I EEE I nternati onal Soli d-State Ci rcui ts ConferenceI SSCC 2025/SESSI ON 12/I NNOVATI ONS FROM OUTSI

121、DE THE(I SSCCS)BOX/12.3979-8-3315-4101-9/25/$31.00 2025 I EEE12.3 Reversing Scattering to Perform Deep-Tissue Optical I maging and the Current Need for a Suitable Optoelectronic Solution Changhuei Yang Cali forni a I nsti tute of Technology,Pasadena,CA Bi ologi cal ti ssues are hi ghly turbi d i n t

122、he opti cal regi me the mean opti cal scatteri ng length i s on the order of 100 mi crons.Thi s extreme turbi di ty prevents sci enti sts and cli ni ci ans from performi ng deeply penetrati ng hi gh resoluti on opti cal i magi ng through humans and ani mal models ali ke.The challenge associ ated wi

123、th deep-ti ssue opti cal i magi ng i s aki n to the challenge of focusi ng li ght through a dense fog scatteri ng prevents meani ngful soluti ons to thi s problem by tradi ti onal means.Thi s i s why hi gh-resoluti on opti cal i magi ng and opti cal exci tati on methods cannot reach more than 1 mm i

124、 nto bi ologi cal ti ssues.Deep penetrati ng i magi ng modali ti es,such as x-ray i magi ng,MRI,and ultrasound i magi ng,are excellent at reveali ng structures wi thi n bi ologi cal enti ti es,but provi de no bi ochemi cal detecti on capabi li ty.On the other hand,there i s a wi de vari ety of opti

125、cal domai n methods capable of reveali ng bi ochemi cal content,i ncludi ng fluorescence,harmoni c generati on,Raman spectroscopy and absorpti on spectral si gnatures.I f we can overcome opti cal ti ssue turbi di ty,opti cal i magi ng can dramati cally change the way we do bi osci ence research and

126、practi ce medi ci ne.Over the past 2 decades,advances i n opti cal wavefront shapi ng have made si gni ficant stri des i n tackli ng opti cal turbi di ty and enabli ng focusi ng of li ght through bi ologi cal ti ssues 1.The promi si ng research di recti on i s at a cri ti cal juncture,because i t re

127、qui res an i mportant optoelectroni c soluti on i n order to move forward meani ngfully.Much li ke MRI requi red the engi neeri ng and creati on of a hi gh tesla magnet to truly reali ze i ts potenti al,deep ti ssue opti cal i magi ng requi res a si gni ficant but very achi evable scale-up of a spec

128、i ali zed optoelectroni c system to become practi cally useful.Wavefront shapi ng and opti cal phase conjugati on.Scatteri ng i s a determi ni sti c and reversi ble process.A poi nt source of li ght transmi tted through a fog would present a hi ghly di sordered opti cal wavefront on the other si de,

129、where the phase and ampli tude of the di sorder wavefront can randomly vary at a granulari ty of a speckle grai n(lambda/2).But i f we are able to measure part of the wavefront and generate a counter-propagati ng wavefront(phase conjugate wavefront).Thi s phase conjugate wavefront would be to retrac

130、e the pathways back through the fog and reconstruct the ori gi nal li ght poi nt.Thi s process i s someti mes referred to as ti me-reversed opti cal playback,because the li ght field acts as i f i ts ti me-di recti on has been reversed.We demonstrated that thi s concept can work for ti ssue secti on

131、s of up to thi ckness 7 mm 2,where the average number of scatteri ng events for each photon i s upwards of hundreds.The more of the i ni ti al wavefront we measure,the stronger the reconstructi on peak.I n pri nci ple,there i s no theoreti cal li mi t for depth penetrati on,as scatteri ng i s nulli

132、fied through thi s method.I n practi ce,the orders of magni tude weaker opti cal absorpti on i n ti ssues would eventually exti ngui sh the vast majori ty of li ght for large ti ssue depths cm s to meters.Thi s approach for focusi ng i s di fferent from the conventi onal lens based focusi ng i n a s

133、i gni ficant way-the focus resoluti on i s not li mi ted by the numeri cal aperture of the i magi ng system.The scatteri ng nature of the ti ssue actually create a si tuati on where the scattered li ght can arri ve at the focal spot from all di recti ons,leadi ng to the possi bi li ty of getti ng di

134、 ffracti on li mi ted focusi ng i n the ri ght ci rcumstances.Di gi tal opti cal phase conjugati on.Whi le purely optomateri al means for performi ng opti cal phase conjugati on exi st,they are associ ated wi th low playback returns 3.By carefully ali gni ng a camera and a spati al li ght modulator

135、wi th pi xel-to-pi xel preci si on,i t i s possi ble to create a di gi tal opti cal phase conjugator that can provi de vastly superi or playback gai ns i n opti cal power 4.Thi s i s useful because the i ni ti al li ght source may be very weak,and by usi ng thi s approach a much stronger li ght fiel

136、d can be generated and refocused back to the i ni ti al li ght source.The general scheme of a di gi tal phase conjugati on system i s shown i n Fi g.12.3.2.The di sordered wavefront emergi ng from the scatteri ng medi um i s i nterferometri cally i nterfered wi th a reference field,so that the phase

137、 i nformati on i s encoded i nto ampli tude vari ati ons(di gi tal holography).The combi ned li ght field i s then detected by a camera,and the i nformati on i s then processed to reveal the phase vari ati ons.Thi s phase wavefront i s then di gi tally phase conjugated(reverse the phase si gn)and th

138、e result i s sent to the spati al li ght modulator.A planar li ght field reflected from the spati al li ght modulator would then i mpri nt the i nformati on and form the opti cal phase conjugate li ght field that can then travel back through the scatteri ng medi um to regenerate the ori gi nal li gh

139、t source.The performance of such a system i s characteri zed by the so-called peak-to-background(PBR)i ntensi ty rati o.As a general rule,the PBR scales as the number of pi xels used i n the di gi tal phase conjugati on process,and i t scales i nversely as the number of speckle grai ns i n the focus

140、ed spot.A di gi tal opti cal phase conjugati on system can generally manage to keep 100,000 effecti ve pi xels i n good enough ali gnment.Ulti mately,thi s li mi t i s due to the di fficulty of keepi ng the spati al li ght modulator and camera well-i maged onto each other.I t i s a li mi t that can

141、be easi ly overcome i f we can bui ld the two system on the same substrate.Deep ti ssue opti cal focusi ng and i ts appli cati ons.Thi s ti me-reversal scatteri ng nulli ng approach can i n turn be combi ned wi th di fferent gui destar approaches to generate controllable and scannable opti cal foci

142、deep i n ti ssue 1.One promi nent strategy i s to use ultrasound modulati on to tag an i ni ti al li ght field deep wi thi n ti ssues to form a vi rtual gui destar 5.Through thi s means,we have demonstrated that i t i s possi ble to perform hi gh-resoluti on fluorescence i magi ng i n ti ssues at un

143、precedented depth of 5mm 6.Thi s represents the first demonstrati on of bi ochemi cally sensi ti ve hi gh-resoluti on deep ti ssue opti cal i magi ng.Thi s controlled focus deep wi thi n bi ologi cal ti ssues can,i n pri nci ple,be used wi th other opto-chemi cal i nteracti ons to perform other form

144、s of bi ochemi cal sensi ng and i magi ng deep i n ti ssues.I nteresti ng,thi s opti cal focus can also be used for purposes other than i magi ng.For example,we have demonstrated that i t i s possi ble to use the opti cal focus to selecti vely sti mulate optogeneti cally tagged neurons i n li vi ng

145、brai n secti ons 7.The next step:I ntegrated di gi tal opti cal phase conjugati on.Thi s technology requi res hi gh reacti on ti me.Mi nute movements of the scatterers wi thi n ti ssues can throw off the phase conjugate wavefront soluti ons i f the playback has a si gni ficant lag.Ti ssue opti cal d

146、ecorrelati on occurs at a rate of sub-mi lli seconds for mi lli meters depth i n li ve ani mal models.Current opti cal phase conjugati on approach has a lag ti me on the order of tens of mi lli seconds.Thi s i s due to the i nformati on pi peli ne bottlenecks associ ated wi th i mage data flowi ng o

147、ut of the camera to the computer and then eventually to the spati al li ght modulator.An i ntegrated camera and spati al li ght modulator(pi xel-to-pi xel matched)bui lt on the same substrate where the i nformati on only flows between the camera pi xel and i ts matchi ng spati al li ght modulator pi

148、 xel would dramati cally cut the lag and enable thi s exci ti ng research di recti on to progress to novel practi cal appli cati ons.An i ntegrated system wi th a reacti on speed of 100 mi croseconds and a pi xel count of 1 megapi xels would be an appropri ate starti ng poi nt.I t i s also worth not

149、i ng that bi nary phase and ampli tude modulati on would be adequate,and hi gher bi t depth does not as si gni ficantly add to the capabi li ty i n contrast to hi gher pi xel counts.Such a technology i s also broadly useful i n other appli cati ons.For example,i n defense appli cati ons,thi s approa

150、ch to perform hi gh speed focusi ng can potenti ally be used to mai ntai n laser focus on an i ncomi ng mi ssi le duri ng a cloudy day to destroy the mi ssi le.I n autonomous navi gati on appli cati ons,such methods keep LI DAR operati onal through foggy condi ti ons.Figure 12.3.1:Demonstration of p

151、hase conjugation as a means for reversing tissue turbidity.Figure 12.3.2:General scheme of a digital optical phase conjugation system.Figure 12.3.3:The use of ultrasound tagging to serve as a virtual“guidestar”in combination with digital optical conjugation can enable optical focusing of light deep

152、in biological tissue.Figure 12.3.4:The combination of ultrasound“guidestar”and digital optical phase conjugation enables deep tissue(5 mm)high-resolution fluorescence imaging through the rendering of a scannable tight time-reversal optical focus.Figure 12.3.5:The technique can be used to tightly gen

153、erate a localized light spot for activating optogenetically tagged neurons deep in brain tissue.Figure 12.3.6:The missing platform technology for the next phase of this research direction:integrated digital phase conjugation mirror.I SSCC 2025/February 18,2025/8:50 AM231 DI GEST OF TECHNI CAL PAPERS

154、 12 2025 I EEE I nternati onal Soli d-State Ci rcui ts ConferenceI SSCC 2025 PAPER CONTI NUATI ONS AND REFERENCES979-8-3315-4101-9/25/$31.00 2025 I EEERe fe re nc e s:1 R.Horstmeyer,H.Ruan,and C.Yang,“Gui destar-assi sted wavefront-shapi ng methods for focusi ng li ght i nto bi ologi cal ti ssue,”Na

155、 ture Photon,vol.9,no.9,pp.563571,Sep.2015,doi:10.1038/nphoton.2015.140.2 E.J.McDowell,M.Cui,I.M.Vellekoop,V.Senekeri myan,Z.Yaqoob,and C.Yang,“Turbi di ty suppressi on from the balli sti c to the di ffusi ve regi me i n bi ologi cal ti ssues usi ng opti cal phase conjugati on,”J.Biome d.Opt.,vol.15

156、,no.2,p.025004,2010,doi:10.1117/1.3381188.3 Z.Yaqoob,D.Psalti s,M.S.Feld,and C.Yang,“Opti cal phase conjugati on for turbi di ty suppressi on i n bi ologi cal samples,”Na ture Photon,vol.2,no.2,pp.110115,Feb.2008,doi:10.1038/nphoton.2007.297.4 M.Cui and C.Yang,“I mplementati on of a di gi tal opti c

157、al phase conjugati on system and i ts appli cati on to study the robustness of turbi di ty suppressi on by phase conjugati on,”Opt.Expre ss,vol.18,no.4,p.3444,Feb.2010,doi:10.1364/OE.18.003444.5 X.Xu,H.Li u,and L.V.Wang,“Ti me-reversed ultrasoni cally encoded opti cal focusi ng i nto scatteri ng med

158、i a,”Nature Photon,vol.5,no.3,pp.154157,Mar.2011,doi:10.1038/nphoton.2010.306.6 Y.M.Wang,B.Judkewi tz,C.A.Di Marzi o,and C.Yang,“Deep-ti ssue focal fluorescence i magi ng wi th di gi tally ti me-reversed ultrasound-encoded li ght,”Na t Commun,vol.3,no.1,p.928,Jun.2012,doi:10.1038/ncomms1925.7 H.Ruan

159、 et al.,“Deep ti ssue opti cal focusi ng and optogeneti c modulati on wi th ti me-reversed ultrasoni cally encoded li ght,”Sc i.Adv.,vol.3,no.12,p.eaao5520,Dec.2017,doi:10.1126/sci adv.aao5520.232 2025 I EEE I nternati onal Soli d-State Ci rcui ts ConferenceI SSCC 2025/SESSI ON 12/I NNOVATI ONS FROM

160、 OUTSI DE THE(I SSCCS)BOX/12.4979-8-3315-4101-9/25/$31.00 2025 I EEE12.4 Skin-I nspired Electronics:An Emerging Sensing and Computing Platform Margheri ta Ronchi ni1,2,Wei chen Wang1,Yuya Ni shi o1,Yati ng Yao1,Zhenan Bao1 1Stanford Uni versi ty,Stanford,CA 2Aarhus Uni versi ty,Aarhus,Denmark Ski n

161、i s the i nterface between our body and the envi ronment.I ts uni que properti es and sensi ng capabi li ti es allow us to percei ve the world around us.I nformati on about shape,texture,sti ffness,temperature,and pai n conveyed by touch faci li tates our abi li ty to perform tasks and mani pulate o

162、bjects dexterously 1.Addi ng sensory feedback i n prostheti c li mbs wi ll enable users to i nteract wi th the envi ronment naturally and effecti vely,i nstead of relyi ng solely on vi sual cues.Tacti le sensi ng technologi es,i n the form of hi ghly dense sensor arrays and neuromorphi c encodi ng c

163、i rcui ts coupled to a neuromodulator,are poi sed to revoluti oni ze prostheti c devi ces.On the other hand,such a sensi ng and si gnal computi ng system archi tecture can provi de a general platform for seamless and mi ni mally i nvasi ve wearable and i mplantable devi ces(Fi g.12.4.1).To reali ze

164、ski n-li ke functi ons,ski n-i nspi red electroni cs(i.e.e-ski n)should emulate several cri ti cal features of natural ski n such as stretchabi li ty,durabi li ty,self-heali ng abi li ty and,above all,mechani cal compli ance 2.The development of stretchable electroni cs represents a transformati ve

165、advancement i n the field of e-ski n.Conventi onal si li con electroni cs,bei ng ri gi d and bri ttle,requi res speci al mechani cal desi gn to be i ncorporated i nto stretchable electroni cs,e.g.buckled serpenti ne structure 3.I ntri nsi cally stretchable electroni cs addresses these li mi tati ons

166、 by i ncorporati ng polymer-based electroni c materi als that possess hi gh electri cal performance under si gni ficant strai n,allowi ng for the creati on of conformable devi ces that can adapt to the full range of dynami c movements of the body.So far,developi ng hi gh-performance i ntri nsi cally

167、 stretchable electroni c materi als and i ntegrati ng them i nto scalable hi gh-resoluti on fabri cati on methods have posed a major challenge to the adopti on of thi s technology.Here,we provi de an overvi ew of the evoluti on of e-ski n systems and soft ci rcui ts over ti me,emphasi zi ng i nnovat

168、i ons i n molecular desi gn,fabri cati on protocol developments,devi ce engi neeri ng,and ci rcui t desi gn.The poor scalabi li ty of stretchable electroni cs pri mari ly stems from two factors:1)the scarci ty of hi gh-performance strai n-tolerant electroni c materi als,as crack formati on leads to

169、substanti al reducti on i n the contacts/i nterconnects conducti vi ty,degraded semi conductor charge carri er mobi li ty,and di electri c leakage;2)the lack of hi gh-resoluti on patterni ng techni ques to accurately and reli ably pattern these materi als,e.g.photoresi sts commonly used i n tradi ti

170、 onal photoli thography are not chemi cally orthogonal to acti ve polymeri c electroni c materi als and may degrade thei r performance 4.To overcome the first li mi tati on,the CONPHI NE(conjugated-polymer/elastomer phase separati oni nduced elasti ci ty)methodology was di scovered 5 as a general st

171、rategy to i ncrease polymer semi conductor stretchabi li ty,wi th enhanced charge carri er mobi li ty gi ven the appropri ate matri x selecti on and processi ng condi ti ons 6.A thi n film of hi gh-mobi li ty semi conducti ng conjugated polymer was blended wi th a soft thermoplasti c elastomer of si

172、 mi lar surface energy to form a robust stretchable polymer semi conductor.The nanoconfinement of the polymer semi conductor i nduced by the i ncorporati on of the thermoplasti c elastomer rubber-li ke matri x resulted i n:1)reduced conformati onal di sorder and i mproved charge transport;2)reduced

173、crystalli ni ty and a lower glass transi ti on temperature,whi ch gave a more ducti le semi conductor network.The nanoconfined semi conductor nanofibers deri ved from the phase-separati on faci li tated charge transport and allowed mobi li ty to preserve i ts value even when the film was stretched u

174、p to 100%5.Si mi lar materi al desi gn and processi ng concepts enabled a collecti on of hi gh-performance electroni c materi als,such as self-healable and stretchable semi conductors 7,stretchable and bi odegradable polymer semi conductors 8,and hi ghly conducti ve transparent stretchable conductor

175、s 9,10.Next,we revi ew the progress wi th stretchable ci rcui t fabri cati on.Fi gure 12.4.2 shows the development of such ci rcui ts wi th i ntri nsi cally stretchable materi als.Wi th the hi gh-performance CONPHI NE semi conductors,the first generati on of stretchable transi stor arrays was develo

176、ped,wi thout usi ng any unreli able transfer processes 11.CONPHI NE semi conductor was patterned by O2 plasma etchi ng usi ng a copper mask and a fluori nated sacri fici al layer,whi le the di electri c layer was photo-patterned by UV-i ni ti ated azi de-crossli nki ng reacti on.The drai n and sourc

177、e electrodes were patterned by shadow maski ng,whi ch li mi ted the densi ty of the transi stor arrays to about 350 transi stors/cm2.Nevertheless,thi s was the first scalable process for fabri cati on of i ntri nsi cally stretchable transi stor arrays wi th mobi li ty on par wi th that of amorphous-

178、Si.Basi c ci rcui t elements,such as pseudo-CMOS i nverter,NAND gate,and an ampli fier i nterfaced to a strai n sensor showed the potenti al of thi s technology for si gnal-processi ng at the edge 11.A later work bui lt on the all-elastomer fabri cati on led to strai n-i nsensi ti ve hi gh-gai n amp

179、li fiers by engi neeri ng sti ffness of elastomer substrate 12.Both analog and di gi tal ci rcui ts benefited from the strai n-i nsensi ti ve mechani cal structure.I n parti cular,a NOR gate,a ri ng osci llator and the first stretchable two-stage pseudo-D ampli fier wi th a gai n of over 120 were de

180、monstrated.To further i ncrease the patterni ng resoluti on,the PhotoAssi st method was i ntroduced 4 that uti li zed di rect photo-patterni ng of all the i nvolved polymeri c electroni c materi als i n a transi stor.Thi s requi red development of sui table photo-patterni ng chemi stry that does not

181、 degrade electroni c performance and si multaneously generates clean patterns.As a result,drai n and source electrode channel length of 2m could be reali zed wi th an overall transi stor densi ty of 42,000 transi stors/cm2.Recent advancements led to the creati on of transi stor arrays wi th hi gh dr

182、i vi ng current(2.0 A m1 under 5V)and channel lengths as small as 0.9m 13.These results were achi eved thanks to the development of a smooth patterned gate 10,a hi gh-mobi li ty hi gh puri ty semi conducti ng carbon nanotube(S-CNT)network film,low-contact-resi stance source/drai n contacts wi th met

183、alli c carbon nanotubes(M-CNTs),hi gh permi tti vi ty di electri c(ni tri le-butadi ene rubber,NBR),and low-sheet-resi stance hi gh-resoluti on patterned i nterconnects(eutecti c galli umi ndi um alloy,EGaI n).The fabri cati on process gave transi stor arrays an average field-effect mobi li ty of mo

184、re than 20 cm2 V1 s1 under 100%strai n wi th unprecedented devi ce densi ty of 100,000 transi stors/cm2 and a yi eld up to 99.3%13.At the ci rcui t level,a 527-stage ri ng osci llator wi th 1MHz swi tchi ng speed(Fi g.12.4.3),and a 2,500 uni ts cm2 dense acti ve-matri x tacti le sensor for brai lle

185、recogni ti on were demonstrated 13.More recently,we have further i mproved thi s fabri cati on process to i ncorporate an encapsulati on and a dual-gate structure.Combi ned wi th rati onal ci rcui t desi gn,we reali zed analog-to-di gi tal converters(ADCs)wi th hi gh gai n and low strai n dependence

186、(Fi g.12.4.4).Thi s represents a major step forward for hi ghly i ntegrated,large-scale i ntri nsi cally stretchable ci rcui ts.A summary of charge carri er mobi li ty and correspondi ng transi stor densi ty i s provi ded i n Fi g.12.4.5.E-ski n prostheti c system represents a general sensi ng and c

187、omputi ng system enabled by sensors development and i ntri nsi cally stretchable ci rcui ts development.Past e-ski n desi gns i ncluded an organi c mechanoreceptor 17,an arti fici al afferent nerve 18,and a muscle spi ndle-based propri ocepti ve feedback 19.The organi c mechanoreceptor transduced th

188、e pressure appli ed to a mi crostructured resi sti ve tacti le sensor i nto frequency-modulated di gi tal pulses i n the range 0-to-200Hz,emulati ng the physi ologi cal response of slow-adapti ng mechanoreceptors.The generated firi ng patterns were then used for the optogeneti c sti mulati on of a m

189、ouse geneti cally-modi fied somatosensory neurons 17.The ri gi d electroni c components used for si gnal encodi ng were recently replaced by the i ntri nsi cally stretchable ci rcui t shown i n Fi g.12.4.6,made by a seven-stage ri ng osci llator and an edge detector,together compri si ng a total of

190、54 transi stors 20.The arti fici al afferent nerve showed the power of combi ni ng spi ke-trai n si gnal outputs from multi ple pressure sensors and organi c osci llators as gate i nputs to an i on-gated synapti c transi stor to produce patterns of postsynapti c current 18.The most advanced e-ski n

191、i s a monoli thi c soft e-ski n closed-loop system for multi modal sensati on,neuromorphi c encodi ng,and motor actuati on 20.The development of a tri layer di electri c enabled ci rcui ts operati on wi th a 3V supply that di ssi pated 100 ti mes less than previ ous generati ons of stretchable trans

192、i stors.The above sensi ng and neuromorphi c encodi ng desi gn can be appli ed to a vari ety of wearable and i mplantable sensor appli cati ons.Wi th further scali ng down of e-ski n power consumpti on and footpri nt,we envi si on that next generati ons of i ntri nsi cally stretchable complementary

193、logi c(see Fi g.12.4.7),i mproved devi ce characteri sti cs and clever ci rcui t desi gn for strai n-i nsensi ti ve operati on wi ll greatly broaden the scope of potenti al appli cati ons of ski n-li ke stretchable electroni c sensors and ci rcui ts.Figure 12.4.1:Examples of applications of skin-ins

194、pired electronics.I llustration contains BioRender ContentFigure 12.4.2:Timeline of the circuits implemented in intrinsically stretchable electronics.Number of transistors is indicated in round brackets.References are in square brackets.All circuits were PMOS designs except the seven-stage CMOS ring

195、 oscillator.Figure 12.4.3:Schematic circuit diagram(a)and optical microscopic image(b)of a stretchable 527-stage ring oscillator(RO)with 1056 transistors.I nset:Size comparison between the RO,an active-matrix stretchable transistor array and a US quarter.Scale bar,5mm.(c)Frequency spectrum of the ou

196、tput signal from the three-stage RO.Ref 13.Figure 12.4.4:Comparison of recently developed intrinsically stretchable analog-to-digital converter with reported flexible versions.FOM:Walden figure-of-merit;ENOB:effective number of bits;fs:Nyquist sampling rate;SNR:Signal-to-noise ratio.Figure 12.4.5:De

197、velopment of soft intrinsically stretchable transistor array fabrication processes.Gen 1:Ref 11;Gen 2:Ref 4;Gen 3:Ref 13.Figure 12.4.6:Low-voltage-driven soft stretchable circuit electronic skin system for generating biomimetic pulse trains.(a)I mage of a stretchable seven-stage ring oscillator(RO)-

198、edge detector(ED)circuit.(b)Schematic of the biomimetic sensor-circuit system for sensory information encoding.(c)Output frequency of the sensor-RO-ED system under different pressures.Ref 20.I SSCC 2025/February 18,2025/9:15 AM233 DI GEST OF TECHNI CAL PAPERS 12 2025 I EEE I nternati onal Soli d-Sta

199、te Ci rcui ts ConferenceI SSCC 2025 PAPER CONTI NUATI ONS AND REFERENCES979-8-3315-4101-9/25/$31.00 2025 I EEEFigure 12.4.7:Status of current intrinsically stretchable CMOS circuit fabrication.Box:Challenges faced in skin-inspired electronics.Re fe re nc e s:1 A.Chortos,J.Li u and Z.Bao,“Pursui ng p

200、rostheti c electroni c ski n,”Na t.Ma te r.,vol.15,pp.937-950,Jul.2016.2 C.Zhao et al.,“Ski n-i nspi red soft bi oelectroni c materi als,devi ces and systems,”Na t.Re v.Bioe ng.,vol.2,pp.671-690,Jun.2024.3 J.A.Rogers,T.Someya and Y.Huang,“Materi als and mechani cs for stretchable electroni cs,”Sc ie

201、 nc e,vol.327,no.5973,pp.1603-1607,Mar.2010.4 Y.-Q.Zheng et al.,“Monoli thi c opti cal mi croli thography of hi gh-densi ty elasti c ci rcui ts,”Sc ie nc e,vol.373,no.6550,pp.88-94,Jul.2021.5 J.Xu et al.,“Hi ghly stretchable polymer semi conductor films through the nanoconfinement effect,”Sc ie nc e

202、,vol.355,no.6320,pp.59-64,Jan.2017.6 J.Xu et al.,“Multi-scale orderi ng i n hi ghly stretchable polymer semi conducti ng films,”Na t.Ma te r.,vol.18,pp.594-601,Apr.2019.7 J.Y.Oh et al.,“I ntri nsi cally stretchable and healable semi conducti ng polymer for organi c transi stors,”Na ture,vol.539,pp.4

203、11-415,Nov.2016.8 H.Tran et al.,“Stretchable and fully degradable semi conductors for transi ent electroni cs,”ACS Ce nt.Sc i.,vol.5.,no.11,pp.1884-1891,Nov.2019.9 Y.Wang et al.,“A hi ghly stretchable,transparent,and conducti ve polymer,”Sc i.Adv.,vol.3,no.3,p.e1602076,Mar.2017.10 Y.Ji ang et al.,“T

204、opologi cal supramolecular network enabled hi gh-conducti vi ty,stretchable organi c bi oelectroni cs,”Sc ie nc e,vol.375,no.6587,pp.1411-1417,Mar.2022.11 S.Wang et al.,“Ski n electroni cs from scalable fabri cati on of an i ntri nsi cally stretchable transi stor array,”Na ture,vol.555,pp.83-88,Feb.

205、2018.12 W.Wang et al.,“Strai n-i nsensi ti ve i ntri nsi cally stretchable transi stors and ci rcui ts,”Na t.Ele c tron.,vol.4,pp.143-150,Jan.2021.13 D.Zhong et al.,“Hi gh-speed and large-scale i ntri nsi cally stretchable i ntegrated ci rcui ts,”Na ture,vol.627,pp.313-320,Mar.2024.14 N.Papadopoulos

206、 et al.,“Flexi ble selfbi ased 66.7nJ/c.s.6bi t 26S/s successi ve-approxi mati on C-2C ADC wi th offset cancellati on usi ng uni polar metal-oxi de TFTs”,Proc.I EEE Custom I ntegr.Ci rcui ts Conf.(CI CC),Apr.2017,pp.1-4.15 S.Abdi ni a et al.,“A 4b ADC manufactured i n a fully-pri nted organi c compl

207、ementary technology i ncludi ng resi stors,”2013 I EEE I nternati onal Soli d-State Ci rcui ts Conference Di gest of Techni cal Papers,Feb.2013,pp.106-107.16 H.Mari en,et al.,“An Analog Organi c Fi rst Order CT ADC on a Flexi ble Plasti c Substrate wi th 26.5 dB Precessi on,”I SSCC Di gest of Techni

208、 cal Papers,Feb.2010,pp.136-137.17 B.C.-K.Tee et al.,“A ski n-i nspi red organi c di gi tal mechanoreceptor,”Sc ie nc e,vol.350,no.6258,pp.313-316,Oct.2015.18 Y.Ki m et al.,“A bi oi nspi red flexi ble organi c arti fici al afferent nerve,”Sc ie nc e,vol.360,no.6392,pp.998-1003,Jun.2018.19 Y.Lee et al.,“A low-power stretchable neuromorphi c nerve wi th propri ocepti ve feedback,”Na t.Biome d.Eng.,vol.7,pp.511-519,Aug.2022.20 W.Wang et al.,“Neuromorphi c sensori motor loop embodi ed by monoli thi cally i ntegrated,low-voltage,soft e-ski n,”Sc ie nc e,vol.380,no.6646,pp.735-742,May 2023.