+ All Categories
Home > Documents > ThinkingMachines-1986

ThinkingMachines-1986

Date post: 29-May-2018
Category:
Upload: zack-dagoba
View: 218 times
Download: 0 times
Share this document with a friend

of 19

Transcript
  • 8/9/2019 ThinkingMachines-1986

    1/19

    by looking at the whole problem at once. 1

  • 8/9/2019 ThinkingMachines-1986

    2/19

    I Imagine a 1000Mips computer thatE is simpletoprogram. That's right.1000million instructions per second.And simple toprogram.

    A computer that solves your problemsby working on all of the dataat once,instead of one word at a time.

    A computer that handleswordsandpicturesaseasilyasnumbers.

    A computer that isopeningnewfrontiersinappWions.

  • 8/9/2019 ThinkingMachines-1986

    3/19

    The Connection Machinea com- systems. Uniting immense pro-puter system is here. Reversing cessingpowerwith simplicitya twenty-year trendtoward of programming. Establishinggreaterandgreater complex- anew standardof large scaleity in large scale computing computingcapability.

  • 8/9/2019 ThinkingMachines-1986

    4/19

  • 8/9/2019 ThinkingMachines-1986

    5/19

    The Connection Machine system iseasy to use because it has enoughcomputing elements t o assign oneto eac h piece of data. It s 64,000processors scan whole data bases,process whole pictorial images,and simulate whole VLSI circuitssimultaneously. All proce ssors arelinked together in a dynamic net-work that reconfigures to matchthe way the problem works. T heresult is a simpler to programcomputer, one that delivers i tsenormo us computing powermo re naturally and directly.What do we mean by "simple toprogram"? We mean th at many ofyour mo st complicated and time -consuming subroutines can b ereplaced by single operations thatare executed directly by themachine. Not just regu lar opera-tions like vector arithmetic, but

    also operations that require com-plex patterns of data transfer, likenetwork propagation and sorting.We also mean that i t is easy touse . You inte rac t with the familiarprogramming environm ent of afront end syste m, running theedito rs, utilities, and operatingsystem to which you are alreadyaccustom ed. You wri te yourprogram in a sim ple extensio n of ahigh level programming language,like C or Lisp. The com piler andthe Connection Machine hardwareautomatically take care of how thedata is allocated to th e processors,how operations ar e synchronized,and what data is communicatedfrom one proces sor to another. It'sgood at those things, so you canconc entrate on the application, notthe computer.

  • 8/9/2019 ThinkingMachines-1986

    6/19

  • 8/9/2019 ThinkingMachines-1986

    7/19

    Looking at the w hole problem atonce means com puting on everyelem ent of a data structure a t thesame time. T he ConnectionMachine system's tens of thou-sands of process ors allow it to dothis in a straightforward way: byattaching a separate p rocessor toeach element of a data structure.The type of data structu redepends on th e application. Inmany language processing applica-tions, data level parallelism m eansa processor for every word orevery meaning. For data base appli-cations, it means a processor forevery document. In a numericsimulation, it may mean a process orfor every element of a matrix. Andin image processing, data level par-allelism means a processor forevery picture eleme nt, or pixel, inan image.The performance advantages ofdata level parallelism a re dram atic.The parallelism inhe rent in thedata structure s grows with the sizeof th e data. Proces sing all the pix-els in an image, for example, is asfast as processing one pixel,because they can all be computedat the same time. Awhole database canbe search ed in the time ittakes to search one document.When com puting the flow of a ir over

    an airplane wing, the C onnectionMachine system calculates the flowover all par ts of the wing simulta -neously, just like it works in reality.Speedups of 1000or more ar e com-mon with data level parallelism.If data level parallelism is so sim-ple and powerful, why don't all bigcomputers work this way? Becausein order to make these programswork, tens of thousands of proces-so rs have to work together.They have to communicate. TheConnection Machine system so lvesthe communications problem inhardware, so th e natural algorithmswork well.What makes the Connec tionMachine system work ar e the con-nections. Inside the machine ther eis a very flexible high-bandwidthcommunication network tha t movesdata between processors at billionsof bits pe r second. R outing circuitson every chip automatically ste erdata along the faste st paths, helpingto m ake programming simpler.Ther e is no need to adapt yourapplication to the stru ctur e of afixed arch itecture like a grid, ring,hypercube, or tree. Instead, theConnection Machine sy stem adaptsto your application, by dynamicallyforming the connections that a reneeded.

  • 8/9/2019 ThinkingMachines-1986

    8/19

    (A) In the simulation of a VLSIintegra ted circuit, the individualproblem elemen t is the transistor.Transi stors connect togethe r bytheir circuit wiring, which is uniqueto th e individual circuit. Sometransisto rs are wired to the transis-tor n ext door, som e are wired totransisto rs clear across the chip,and others a re wired to both. Forevery VLSI circuit it simulates,the Connection Machine syste mreconfigures its internal proces sorconnections t o match the wiring ofthe circuit.

    (B ) In picture processing, th e indi-vidual problem ele ment is thepixel. Pixels intera ct in a gridpattern, with th e st ate of each pixelinfluencing those adjacent t o it. Forimage processing, t he ConnectionMachine .system configures itselfas a grid, allowing each proces sorto pass information to the proces-sors above, below, and to the sides.

    (C ) In language processing, theindividual problem eleme nt is theword o r the meaning. Words andmeanings connect together inincreasingly complex structu rescalled semantic networks. T heConnection Machine system se tsup a connection between eachrelated word and meaning.

    (Dl Th e m ost demanding applica-tions have no fured topology. Theychange during the co urse of theproblem. So doe s the ConnectionMachine system . A completevision syste m may, for example,start its analysis by filtering theimag e in a grid topology. Globaloperations, such as computing theaverage intensity, are naturallytree-like in nature. Bringingtogethe r features into objectsrequires irregular patterns ofcommunication. Only theConnection Machine sys temadapts dynamically to thes echanging patterns.In short, the ConnectionMachine system's simplicity is alsoth e sour ce of its great power. Largenumbers of processors plus adynamically reconfigurable inte r-communications network combineto allow processing to go ondirectly at the data level, wherethousand-fold parallelism is natural.

  • 8/9/2019 ThinkingMachines-1986

    9/19

  • 8/9/2019 ThinkingMachines-1986

    10/19

    Th e ConnectionMachine syste mretrieves informa-tion by wholedocument com-parison instead ofby individual keyword.

    -&*,.

  • 8/9/2019 ThinkingMachines-1986

    11/19

    Th e ConnectionMachine architec-ture is ideal forth e whole rangeof vision algorithmsand applications.

  • 8/9/2019 ThinkingMachines-1986

    12/19

    VLSI simulationis ust one of thetwo-dimensionaland three-dimensionalnumeric problemsthe ConnectionMachine systemis solving.

  • 8/9/2019 ThinkingMachines-1986

    13/19

    Fluid dynamics isan outstandingexamp le of howth e power of theConnectionMachine systemallows innovativeusers to deal withphysical behaviordirectly.

  • 8/9/2019 ThinkingMachines-1986

    14/19

    Example Application:Numbers.The Connection Machine syst embrings a new level of power andnaturalne ss to numeric computingapplications. The simulation of aVLSI circuit shows how generalcommunication and massive com-puting power combine to providean easy-to-program solution. Th eproblem i s to feed a wave-form intoa circuit, then compu te the wave-form that t he circuit will produce atany other selected point. For acircuit with ten s of thousands oftransistors , all switching on andoff in parallel, the required num berof compu tations is immense.The C o~ ec ti on achine systemconfigures its communicationspattern t o match ea ch individual cir-cuit exactly. A com plete processoris assigned to each transistor andnode. T he linkages between proc-essors are configured to match thewires between transistors. T hesimulation match es the reality: allprocessors (transistors) compute(change state) in parallel.Processors evaluate mathemat-ical equations that describe th ebehavior of t he dev ice they ar esimulating. Then they cornmuni-cate their voltage, conductance,current, and charge to other de-vices. Finally, the syste m iterate sto reach co nsistent values for alldevices for that time s tep of thesimulation.VLSI simulat ion is just one ofthe two-dimensional and three-dimensional numeric problems thatthe Connection Machine syste m issolving. Other exam ples includematrix algebra and linearprogramming.

    Example Application: AggregateBehavior. Powerful though theyare, words, numbers, and picturesare simply abstractions of the worldaround us. They simplify behaviorthat is too complex to deal withdirectly.Fluid dynamics is an outstandingexam ple of how the power of theConnection Machine system allowsinnovative users to cu t throughthese abstractions and reach thephysical behavior directly. Byanalyzing the way fluids flow, scien-tis ts design mo re efficient aircraftand make more accurate weatherforecasts . Traditional methodshave modeled fluids by partial dif-ferential equations. The Connec-tion Machine system provides newinsights by more directly modelingthe behavior of ac tual molecules.Huge numbers of packets a reintroduced into the syste m, eachlike a tiny bundle of molecules.They move and jostle according tovery simple logical rules. No arith-metic of any kind is involved, sothe programming is simpler thantraditional methods.As th ese individual localinteractio ns evolve in huge num-bers (the Connection Machinesystem can update a billion of th eseindividual state s per second),the behav ior of fluids em erge s.Injected from an e dge , a fluid flowsin a regular way until an obstruc-tion (such a s an airfoil) is encoun-tered. Swirls and eddies appear.Over time they lose their energyand trail off.The simplicity of th e modelmakes simulation an interac tiveprocess for the first time. A changein the geom etry of th e objectrequires no lengthy reinitialization.Results for a new shape areavailable in a minute or two.

  • 8/9/2019 ThinkingMachines-1986

    15/19

    Programming and Using theConnection Machine System.We no longer have to acce pt com-plexity as the price of increasedperformance in high e nd comput-ers. The Connection Machine sys-tem achieves performance withouttricky com pilers, ultra-fast compo-nents , or ex otic packaging.Simple Construction.The Con-nection Machine system u ses a fewsimple parts over and aver again.Aspecial processor chip implem entssixteen processor s. 4096 of the sechips are used in the system as awhole. D esign rules throughout thesystem are conservative. The reare no e xotic technologies to causereliability problems. T he system isentire ly air cooled. No unusual sitepreparation is requ ired.If a failure do es occur, fault diag-nosis is rapid. E ach processorneed s only to check itself. 65,536processors do this testing in paral-lel. Mo st failures are isolated tothe chip level within five minutes.Hardware is repaired by replacingmodules. The re are only sevenkinds of modules in the system , soa complete se t of spa res is easilymaintained at ev ery site.

    FamiliarOperatingEnviron-ment.The use r interacts with theConnection Machine syste mthrough a conventional front endsystem, such as a VAF or a Symbol-ics 3600.2 Th e front end supportsthe operating environment. Instal-lation of a Connection Machinesystem does not mean introducinga new ope rating environment intoyour data center.The system i s programmed viathe fron t end, using familiar editorsand utilities. File structu res andnetwork protocols are supportedthere as well, a s are th e full rangeof sta nda rd VAX and Symbo licsperipherals.A CloserLook:Programmingatthe Data Level. Computers oper-ate by having a program (th e con-trol sequ ence) operating on a se t ofdata elements to solve a particularproblem. S ystem s that use "controllevel parallelism" apply their pro-ces sor s to individual sections of th eprogram. Th e Connection Machinesystem applies its parallelismdirectly to the data. T he amount ofparallelism that can be ex ploitedgrows with the size of the data.There is no need to segment theprogram.If the data struc tures containfewer than 65,536 elemen ts apiece,the Connection Machine systemmakes the a ssignmen t directly.

    If there a re more, the systemoperates in virtual processor m ode,simulating a large r number of pro-cessors in a way that is tran sparentto the user's program. The syste mcan easily sup port up to 1,000,000processors in the virtual processormode.Variable word length program-ming and dynamic reconfigurationcomplete th e task of matching theprogram's data exactly. Data in aConnection Machine sys tem maybe a s small as one bit or a s large asthousands of bits. For picture pro-cessing, one- to eight-bit values arecommon. For numeric processing,16- to 64-bit words appear mostfrequently. Language proces singvalues, such as words and sen-tences, can vary from a few bits tothousands. The ConnectionMachine system handles them allwith equal efficiency.Dynamic reconfigurationmatches the data's connectivity aswell as the da ta itself. When thedata structures are s et up, so arethe linkages. Any proce ssor cansend data to any other processor.Within th e Connection Machinelanguage environmen t, data con-nections are carried out automati-cally. At th e hardw are level, thesystem supports interconnectionwith a comm unications systemcalled the router. All 65,536processors can exchange datasimultaneously.

    The C* Programming Language.C* is a dire ct implem entation ofth e data level computing philosophyfor general purpo se computing.The language is so similar to C thatit requir es no se parate compiler.C* is implemented simply as a pre-processor to the standard C com-piler. C onsistent with the ten etsof data level parallelism, the re a reno change s to the basic commandstructure. Th e control sequenceof a C program d oes not chan gewithin C*. Enhanc ements are at thedata level, allowing data stru ctu resto be connected to individualprocessors for rapid execution.The *Lisp Programming Lan-gufLge.*Lisp is a direct implemen-h h o n of the data level philosophyfor artificial intelligence ap plica-tions. *Lisp adds data structure s toallow direct control of memo ryallocation and assig nme nt of d atavalues to processors. I t gives theprogrammer direct access to, andcontrol over, the Conn ectionMachine hardware, but does sofrom within th e L isp programminglanguage. Thus the *Lisp userretains all the productivity benefitsof the Lisp machine environment.

    The Conneetion Machine systemuses a few simplepartsoverandoveragain.A speeidpmmmrchipimplementssixteenproms-mrs.~ofkchiprrareusedin th e system as awhole.

  • 8/9/2019 ThinkingMachines-1986

    16/19

  • 8/9/2019 ThinkingMachines-1986

    17/19

    The Future Belongs toCom-putersThatLookAt the WholeProblem At Once.The Connection Machine systemoperates on a problem's en tiredata se t at once. It is a strikinglysimpler approach that opens upwhole new possibilities for problem-solving. For pictu re and imageprocessing, it provides performancelevels thousands of time s grea terthan conventional machines. Forlanguage processing, it providesaccess to information in the waypeople can most easily use it. Forscientific processing, it providesaccess to physical phenomenanever before captured in computersimulation.Yet the impact of data level com-puting is only just beginning. Asinnovative scientis ts ar e becomingconversant with the ConnectionMachine system's potential, theyare inventing new algorithms andapproaches to problem-solving.They ar e merging the image,language, and numeric process ingcapabilities of th e s ystem intohigher level a lgorithms. ThinkingMachines Corpora tion itself is atthe forefront of th is new wave ofcomputing. Byusing knowledge inall its forms, we see the potentialfor computing syst em s farmoreuseful and farmore powerful thananything available to us today.

    @CONNECTIONMACHINE is a re gis ter edtradem ark of ThinkingMachinesCorporation"The CUBE OF LIGHTS sym bol is a t rad e-mark of Thinking Machines Corporation"VAX is a trade mark of D igital Equipmentcorp."SYhdEOLICS 3600 is a trade mark ofSymbohcs, Inc.

  • 8/9/2019 ThinkingMachines-1986

    18/19

    It will be atrulyintelligent machine.

    Onethatcansee andhear and speak.

  • 8/9/2019 ThinkingMachines-1986

    19/19

    ThinkingMachinesCoqxmtion245FirstStreetCambridge,Massachusetts 02142(617)876-XU


Recommended