01010011010101101001011010011001101001101 01010011001100101010011001100101010101010 10101010101011110100101001010011100100101 01001101010101011001100110011001010011001 01001010010100101010101001010101010101100 10100110011001100101001100101001100101001 10011001100111111011011110110111011101101 11011011010010010010011101010100101010101 0101010101010101010100101010111011011011 0001001101110110110110101001110011000101 0101010101110100100101010100111001101101 10110101011011011101110111011101110110101 01011011010001110010101011001100110010100 11001010011001010011001010010100101110011 00110010101011100010001001001000100000001 01111110010101010110101010101010101010010 10101111010010100101001110010010101001101 01010101100110011001100101001100101001010 01010010101010100101010101010110010100110 01100110010100110010100110010100110011001 10011111101101111011011101110110111011011 0100100100100111010101001010101010101010 10101010101010010101011101100101011110100 1010010100111001001010100110101010101100 1100110011001010011001010010111000000000 0100101001010101010010101010100010110010 1001100110011001010011001010011001010011 Neuromorphic Computing: From Materials to Systems Architecture Report of a Roundtable Convened to Consider Neuromorphic Computing Basic Research Needs October 29-30, 2015 Gaithersburg, MD Organizing Committee Ivan K. Schuller (Chair), University of California, San Diego Rick Stevens (Chair), Argonne National Laboratory and University of Chicago Neuromorphic  Computing:  From   Materials  to  Systems  Architecture     Report  of  a  Roundtable  Convened  to   Consider  Neuromorphic  Computing   Basic  Research  Needs     October  29-­‐‑30,  2015   Gaithersburg,  MD         Organizing  Committee     Ivan  K.  Schuller  (Chair),  University  of  California,  San  Diego     Rick  Stevens  (Chair),  Argonne  National  Laboratory  and     University  of  Chicago           DOE  Contacts     Robinson  Pino,  Advanced  Scientific  Computing  Research     Michael  Pechan,  Basic  Energy  Sciences       Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     2   Contents   EXECUTIVE  SUMMARY  ........................................................................................................................  3   WHY  NEUROMORPHIC  COMPUTING?  .............................................................................................  5   The  Need  for  Enhanced  Computing  ..........................................................................................................  6   VON  NEUMANN  vs.  NEUROMORPHIC  ..............................................................................................  7   System  Level  .....................................................................................................................................................  7   Device  Level  ......................................................................................................................................................  8   Performance  .....................................................................................................................................................  9   EXISTING  NEUROMORPHIC  SYSTEMS  .........................................................................................  10   Architectures  ..................................................................................................................................................  10   Demonstrations  .............................................................................................................................................  12   IMPLEMENTATION  NEEDS  ..............................................................................................................  14   Neuromorphic  Concepts  .............................................................................................................................  14   Building  Blocks  ..............................................................................................................................................  15   PROPOSED  IMPLEMENTATION  .....................................................................................................  16   Architecture  ....................................................................................................................................................  16   Properties  ........................................................................................................................................................  17   Devices  ..............................................................................................................................................................  18   Materials  ..........................................................................................................................................................  19   OPEN  ISSUES  ........................................................................................................................................  21   Materials/Devices  .........................................................................................................................................  21   System  ...............................................................................................................................................................  22   INTERMEDIATE  STEPS  .....................................................................................................................  24   CONCLUSIONS  .....................................................................................................................................  25   Acknowledgements  ......................................................................................................................................  26   APPENDICES  ........................................................................................................................................  27   Acronyms  .........................................................................................................................................................  27   Glossary  ............................................................................................................................................................  28   Tables  ................................................................................................................................................................  29   WORKSHOP  LOGISTICS  ....................................................................................................................  34   Roundtable  Participants  ............................................................................................................................  34   Roundtable  Summary  ..................................................................................................................................  35   Roundtable  Agenda  ......................................................................................................................................  36   Disclaimer  .......................................................................................................................................................  37         Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     3   EXECUTIVE  SUMMARY     Computation   in   its   many   forms   is   the   engine   that   fuels   our   modern   civilization.   Modern   computation—based   on   the   von   Neumann   architecture—has   allowed,   until   now,   the   development   of   continuous   improvements,   as   predicted   by   Moore’s   law.   However,   computation  using  current  architectures  and  materials  will  inevitably—within  the  next  10   years—reach  a  limit  because  of  fundamental  scientific  reasons.             DOE   convened   a   roundtable   of   experts   in   neuromorphic   computing   systems,   materials   science,   and   computer   science   in   Washington   on   October   29-­‐‑30,   2015   to   address   the   following  basic  questions:     Can  brain-­‐‑like  (“neuromorphic”)  computing  devices  based  on  new  material  concepts   and   systems   be   developed   to   dramatically   outperform   conventional   CMOS   based   technology?   If   so,   what   are   the   basic   research   challenges   for   materials   sicence   and   computing?   The  overarching  answer  that  emerged  was:   The  development  of  novel  functional  materials  and  devices  incorporated  into   unique  architectures  will  allow  a  revolutionary  technological  leap  toward  the   implementation  of  a  fully  “neuromorphic”  computer.   To  address  this  challenge,  the  following  issues  were  considered:   The   main   differences   between   neuromorphic   and   conventional   computing   as   related   to:   signaling   models,   timing/clock,   non-­‐‑volatile   memory,   architecture,   fault   tolerance,   integrated  memory  and  compute,  noise  tolerance,  analog  vs.  digital,  and  in  situ  learning   New  neuromorphic  architectures  needed  to:  produce  lower  energy  consumption,  potential   novel  nanostructured  materials,  and  enhanced  computation   Device  and  materials  properties  needed  to  implement  functions  such  as:  hysteresis,  stability,   and  fault  tolerance   Comparisons   of   different   implementations:   spin   torque,   memristors,   resistive   switching,   phase  change,  and  optical  schemes  for  enhanced  breakthroughs  in  performance,  cost,  fault   tolerance,  and/or  manufacturability   The  conclusions  of  the  roundtable,  highlighting  the  basic  research  challenges  for  materials   science  and  computing,  are:   1.   Creating   the   architectural   design   for   neuromorphic   computing   requires   an   integrative,   interdisciplinary   approach   between   computer   scientists,   engineers,   physicists,  and  materials  scientists   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     4   2.   Creating   a   new   computational   system   will   require   developing   new   system   architectures  to  accommodate  all  needed  functionalities   3.   One   or   more   reference   architectures   should   be   used   to   enable   comparisons   of   alternative  devices  and  materials   4.   The  basis  for  the  devices  to  be  used  in  these  new  computational  systems  require  the   development   of   novel   nano   and   meso   structured   materials;   this   will   be   accomplished   by   unlocking   the   properties   of   quantum   materials   based   on   new   materials  physics       5.   The   most   promising   materials   require   fundamental   understanding   of   strongly   correlated   materials,   understanding   formation   and   migration   of   ions,   defects   and   clusters,   developing   novel   spin   based   devices,   and/or   discovering   new   quantum   functional  materials   6.   To  fully  realize  open  opportunities  requires  designing  systems  and  materials  that   exhibit   self-­‐‑   and   external-­‐‑healing,   three-­‐‑dimensional   reconstruction,   distributed   power  delivery,  fault  tolerance,  co-­‐‑location  of  memory  and  processors,  multistate—   i.e.,  systems  in  which  the  present  response  depends  on  past  history  and  multiple   interacting  state  variables  that  define  the  present  state   7.   The   development   of   a   new   brain-­‐‑like   computational   system   will   not   evolve   in   a   single  step;  it  is  important  to  implement  well-­‐‑defined  intermediate  steps  that  give   useful  scientific  and  technological  information     Successfully  addressing  these  challenges  will  lead  to  a  new  class  of  computers  and  systems   architectures.  These  new  systems  will  exploit  massive,  fine-­‐‑grain  computation;  enable  the   near   real-­‐‑time   analysis   of   large-­‐‑scale   data;   learn   from   examples;   and   compute   with   the   power   efficiency   approaching   that   of   the   human   brain.   Future   computing   systems   with   these  capabilities  will  offer  considerable  scientific,  economic,  and  social  benefits.     This  DOE  activity  aligns  with  the  recent  White  House  “A  Nanotechnology-­‐‑Inspired  Grand   Challenge  for  Future  Computing”  issued  on  October  20th,  2015  with  the  goal  to  “Create  a   new  type  of  computer  that  can  proactively  interpret  and  learn  from  data,  solve  unfamiliar   problems  using  what  it  has  learned,  and  operate  with  the  energy  efficiency  of  the  human   brain”.   This   grand   challenge   addresses   three   Administration   priorities:   the   National   Nanotechnology   Initiative   (NNI),   the   National   Strategic   Computing   Initiative   (NSCI),   and   the  BRAIN  initiative.       Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     5   WHY  NEUROMORPHIC  COMPUTING?       Computers   have   become   essential   to   all   aspects   of   modern   life—from   process   controls,   engineering,  and  science  to  entertainment  and  communications—and  are  omnipresent  all   over  the  globe.  Currently,  about  5–15%  of  the  world’s  energy  is  spent  in  some  form  of  data   manipulation,  transmission,  or  processing.       In  the  early  1990s,  researchers  began  to  investigate  the  idea  of  “neuromorphic”  computing.   Nervous  system-­‐‑inspired  analog  computing  devices  were  envisioned  to  be  a  million  times   more   power   efficient   than   devices   being   developed   at   that   time.   While   conventional   computational   devices   had   achieved   notable   feats,   they   failed   in   some   of   the   most   basic   tasks  that  biological  systems  have  mastered,  such  as  speech  and  image  recognition.  Hence   the   idea   that   taking   cues   from   biology   might   lead   to   fundamental   improvements   in   computational  capabilities.     Since  that  time,  we  have  witnessed  unprecedented  progress  in  CMOS  technology  that  has   resulted  in  systems  that  are  significantly  more  power  efficient  than  imagined.  Systems  have   been   mass-­‐‑produced   with   over   5   billion   transistors   per   die,   and   feature   sizes   are   now   approaching   10   nm.   These   advances   made   possible   a   revolution   in   parallel   computing.   Today,  parallel  computing  is  commonplace  with  hundreds  of  millions  of  cell  phones  and   personal   computers   containing   multiple   processors,   and   the   largest   supercomputers   having  CPU  counts  in  the  millions.       “Machine  learning”  software  is  used  to  tackle  problems  with  complex  and  noisy  datasets   that  cannot  be  solved  with  conventional  “non-­‐‑learning”  algorithms.  Considerable  progress   has  been  made  recently  in  this  area  using  parallel  processors.  These  methods  are  proving   so  effective  that  all  major  Internet  and  computing  companies  now  have  “deep  learning”—   the   branch   of   machine   learning   that   builds   tools   based   on   deep   (multilayer)   neural   networks—research   groups.   Moreover,   most   major   research   universities   have   machine   learning  groups  in  computer  science,  mathematics,  or  statistics.  Machine  learning  is  such  a   rapidly  growing  field  that  it  was  recently  called  the  “infrastructure  for  everything.”     Over   the   years,   a   number   of   groups   have   been   working   on   direct   hardware   implementations   of   deep   neural   networks.   These   designs   vary   from   specialized   but   conventional  processors  optimized  for  machine  learning  “kernels”  to  systems  that  attempt   to   directly   simulate   an   ensemble   of   “silicon”   neurons,   better   known   as   neuromorphic   computing.   While   the   former   approaches   can   achieve   dramatic   results,   e.g.,   120   times   lower   power   compared   with   that   of   general-­‐‑purpose   processors,   they   are   not   fundamentally  different  from  existing  CPUs.  The  latter  neuromorphic  systems  are  more  in   line  with  what  researchers  began  working  on  in  the  1980s  with  the  development  of  analog   CMOS-­‐‑based  devices  with  an  architecture  that  is  modeled  after  biological  neurons.  One  of   the   more   recent   accomplishments   in   neuromorphic   computing   has   come   from   IBM   research,  namely,  a  biologically  inspired  chip  (“TrueNorth”)  that  implements  one  million   spiking   neurons   and   256   million   synapses   on   a   chip   with   5.5   billion   transistors   with   a   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     6   typical  power  draw  of  70  milliwatts.  As  impressive  as  this  system  is,  if  scaled  up  to  the  size   of  the  human  brain,  it  is  still  about  10,000  times  too  power  intensive.     Clearly,  progress  on  improvements  in  CMOS  and  in  computer  hardware  more  generally  will   not   be   self-­‐‑sustaining   forever.   Well-­‐‑supported   predictions,   based   on   solid   scientific   and   engineering  data,  indicate  that  conventional  approaches  to  computation  will  hit  a  wall  in   the  next  10  years.  Principally,  this  situation  is  due  to  three  major  factors:  (1)  fundamental   (atomic)   limits   exist   beyond   which   devices   cannot   be   miniaturized,   (2)   the   local   energy   dissipation  limits  the  device  packing  density,  and  (3)  the  increase  and  lack  of  foreseeable   limit  in  overall  energy  consumption  are  becoming  prohibitive.  Novel  approaches  and  new   concepts   are   needed   in   order   to   achieve   the   goals   of   developing   increasingly   capable   computers  that  consume  decreasing  amounts  of  power.     The  Need  for  Enhanced  Computing     The  DOE  has  charted  a  path  to  Exascale  computing  by  early  in  the  next  decade.  Exascale   machines  will  be  orders  of  magnitude  faster  than  the  most  powerful  machines  today.  Even   though  they  will  be  incredibly  powerful,  these  machines  will  consume  between  20  and  30   megawatts  of  power  and  will  not  have  intrinsic  capabilities  to  learn  or  deal  with  complex   and   unstructured   data.   It   has   become   clear   that   the   mission   areas   of   DOE   in   national   security,   energy   sciences,   and   fundamental   science   will   need   even   more   computing   capabilities  than  what  can  be  delivered  by  Exascale  class  systems.  Some  of  these  needed   capabilities  will  require  revolutionary  approaches  for  data  analysis  and  data  understanding.         Neuromorphic   computing   systems   are   aimed   at   addressing   these   needs.   They   will   have   much   lower   power   consumption   than   conventional   processors   and   they   are   explicitly   designed   to   support   dynamic   learning   in   the   context   of   complex   and   unstructured   data.   Early  signs  of  this  need  show  up  in  the  Office  of  Science  portfolio  with  the  emergence  of   machine   learning   based   methods   applied   to   problems   where   traditional   approaches   are   inadequate.   These   methods   have   been   used   to   analyze   the   data   produced   from   climate   models,   in   search   of   complex   patterns   not   obvious   to   humans.   They   have   been   used   to   recognize  features  in  large-­‐‑scale  cosmology  data,  where  the  data  volumes  are  too  large  for   human   inspection.   They   have   been   used   to   predict   maintenance   needs   for   accelerator   magnets—so   they   can   be   replaced   before   they   fail—to   search   for   rare   events   in   high-­‐‑ energy  physics  experiments  and  to  predict  plasma  instabilities  that  might  develop  in  fusion   reactors.  These  novel  approaches  are  also  being  used  in  biological  research  from  searching   for   novel   features   in   genomes   to   predicting   which   microbes   are   likely   to   be   in   a   given   environment   at   a   given   time.   Machine   learning   methods   are   also   gaining   traction   in   designing  materials  and  predicting  faults  in  computer  systems,  especially  in  the  so-­‐‑called   “materials  genome”  initiative.  Nearly  every  major  research  area  in  the  DOE  mission  was   affected  by  machine  learning  in  the  last  decade.  Today  these  applications  run  on  existing   parallel  computers;  however,  as  problems  scale  and  dataset  sizes  increase,  there  will  be   huge  opportunities  for  deep  learning  on  neuromorphic  hardware  to  make  a  serious  impact   in  science  and  technology.     Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     7   Neuromorphic  computing  may  even  play  a  role  in  replacing  existing  numerical  methods   where   lower   power   functional   approximations   are   used   and   could   directly   augment   planned   Exascale   architectures.   Important   questions   for   the   future   are   which   areas   of   science   are   most   likely   to   be   impacted   by   neuromorphic   computing   and   what   are   the   requirements  for  those  deep  neural  networks.  Although  this  roundtable  did  not  focus  on  an   application  driven  agenda,  it  is  increasingly  important  to  identify  these  areas  and  to  further   understand  how  neuromorphic  hardware  might  address  them.     VON  NEUMANN  vs.  NEUROMORPHIC     System  Level     Traditional  computational  architectures  and  their  parallel  derivatives  are  based  on  a  core   concept  known  as  the  von  Neumann  architecture  (see  Figure  1).  The  system  is  divided  into   several  major,  physically  separated,  rigid  functional  units  such  as  memory  (MU),  control   processing   (CPU),   arithmetic/logic   (ALU),   and   data   paths.   This   separation   produces   a   temporal   and   energetic   bottleneck   because   information   has   to   be   shuttled   repeatedly   between  the  different  parts  of  the  system.  This  “von  Neumann”  bottleneck  limits  the  future   development   of   revolutionary   computational   systems.   Traditional   parallel   computers   introduce   thousands   or   millions   of   conventional   processors   each   connected   to   others.   Aggregate   computing   performance   is   increased,   but   the   basic   computing   element   is   fundamentally   the   same   as   that   in   a   serial   computer   and   is   similarly   limited   by   this   bottleneck.     In  contrast,  the  brain  is  a  working  system  that  has  major  advantages  in  these  aspects.  The   energy   efficiency   is   markedly—many   orders   of   magnitude—superior.   In   addition,   the   memory   and   processors   in   the   brain   are   collocated   because   the   constituents   can   have   different  roles  depending  on  a  learning  process.  Moreover,  the  brain  is  a  flexible  system   able   to   adapt   to   complex   environments,   self-­‐‑programming,   and   capable   of   complex   processing.  While  the  design,  development,  and  implementation  of  a  computational  system   similar  to  the  brain  is  beyond  the  scope  of  today’s  science  and  engineering,  some  important   steps  in  this  direction  can  be  taken  by  imitating  nature.       Clearly   a   new   disruptive   technology   is   needed   which   must   be   based   on   revolutionary   scientific   developments.   In   this   “neuromorphic”   architecture   (see   Figure   1),   the   various   computational   elements   are   mixed   together   and   the   system   is   dynamic,   based   on   a   “learning”   process   by   which   the   various   elements   of   the   system   change   and   readjust   depending  on  the  type  of  stimuli  they  receive.                           Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     8    von  Neumann  Architecture                        Neuromorphic  Architecture                               Figure  1.  Comparison  of  high-­‐‑level  conventional  and  neuromorphic  computer  architectures.  The  so-­‐‑ called  “von  Neumann  bottleneck”  is  the  data  path  between  the  CPU  and  the  memory  unit.  In  contrast,  a  neural   network  based  architecture  combines  synapses  and  neurons  into  a  fine  grain  distributed  structure  that  scales   both  memory  (synapse)  and  compute  (soma)  elements  as  the  systems  increase  in  scale  and  capability,  thus   avoiding  the  bottleneck  between  computing  and  memory.     Device  Level     A  major  difference  is  also  present  at  the  device  level  (see  Figure  2).  Classical  von  Neumann   computing   is   based   on   transistors,   resistors,   capacitors,   inductors   and   communication   connections   as   the   basic   devices.   While   these   conventional   devices   have   some   unique   characteristics  (e.g.,  speed,  size,  operation  range),  they  are  limited  in  other  crucial  aspects   (e.g.,   energy   consumption,   rigid   design   and   functionality,   inability   to   tolerate   faults,   and   limited  connectivity).  In  contrast,  the  brain  is  based  on  large  collections  of  neurons,  each  of   which   has   a   body   (soma),   synapses,   axon,   and   dendrites   that   are   adaptable   and   fault   tolerant.  Also,  the  connectivity  between  the  various  elements  in  the  brain  is  much  more   complex  than  in  a  conventional  computational  circuit  (see  Figure  2).           a)             b)                                 Figure  2.  Interconnectivity  in  a)  conventional  and  b)  neuronal  circuits.     Output     Device   Central  Processing  Unit   (CPU)   Control  Unit   Memory  Unit   Arithmetic  /   Logic  Unit   Input     Device   Dendrites Axon Synapses Soma Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     9   Performance     A   comparison   of   biological   with   technological   systems   is   revealing   in   almost   all   aspects.   Although  the  individual  constituents  of  silicon  systems  naively  seem  to  exhibit  an  enhanced   performance   in   many   respects,   the   system   as   a   whole   exhibits   a   much-­‐‑reduced   functionality.  Even  under  reasonable  extrapolations  of  the  various  parameters,  it  appears   that  the  improvement  in  computational  capacity  of  conventional  silicon  based  systems  will   never   be   able   to   approach   the   density   and   power   efficiency   of   a   human   brain.   New   conceptual  development  is  needed.       Inspection   of   the   delay   time   versus   power   dissipation   for   devices   in   many   competing   technologies  is  quite  revealing  (see  Figure  3).  The  neurons  and  synapses  exist  in  a  region  in   this   phase   diagram   (upper   left   corner)   where   no   other   technologies   are   available.   The   energy  dissipation  in  a  synapse  is  orders  of  magnitude  smaller  although  the  speed  is  much   slower.         Figure  3.  Delay  time  per  transistor  versus  the  power  dissipation.  The  operating  regime  for  neuromorphic   devices  is  in  the  upper  left  corner  indicating  the  extreme  low  power  dissipation  of  biological  synapses  and  the   corresponding  delay  time.  Systems  built  in  this  region  would  be  more  “brain-­‐‑like”  in  their  power  and  cycle   times  (after  userweb.eng.gla.ac.uk).     Table  1  compares  the  performance  of  biological  neurons  and  neuron  equivalents  built  from   typical  CMOS  transistors  currently  used  in  computers.  While  the  values  listed  in  this  table   may  not  be  exact  and  are  debatable,  the  differences  are  so  large  that  it  is  clear  that  new   scientific  concepts  are  needed,  including  possibly  to  the  system  architecture  itself.  While   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     10   silicon   devices   may   exhibit   certain   advantages,   the   overall   system   computational   capabilities—its   fault   tolerance,   energy   consumption,   and   ability   to   deal   with   large   data   sets—is  considerably  superior  in  biological  brains.  Moreover,  there  are  whole  classes  of   problems   that   conventional   computational   systems   will   never   be   able   to   address   even   under  the  most  optimistic  scenarios.         Biology                   Silicon     Advantage   Speed   1  msec   1  nsec   1,000,000x   Size   1µm  -­‐  10µm   10nm  -­‐  100nm   1,000x   Voltage   ~  0.1V   Vdd  ~1.0V   10x   Neuron  Density   100K/mm2   5k/mm2   20x   Reliability   80%   <  99.9999%   1,000,000x   Synaptic  Error  Rate   75%   ~  0%   >109   Fan-­‐out  (-­‐in)   103 -­‐104   3-­‐4   10,000x   Dimensions   Pseudo  3D   Pseudo  3D   Similar   Synaptic  Op  Energy   ~  2  fJ   ~10pJ   5000x   Total  Energy   10  Watt   >>103  Watt   100,000x   Temperature   36C  -­‐  38C   5C  -­‐  60C   Wider  Op  Range   Noise  effect   Stochastic  Resonance   Bad     Criticality   Edge   Far       Table  1.  Comparison  of  biological  and  silicon  based  systems.  This  table  shows  a  comparison  of  neurons   built  with  biology  to  equivalent  structures  built  with  silicon.  Red  is  where  biology  is  ahead;  black  is  where   silicon  is  ahead.  The  opportunity  lies  in  combining  the  best  of  biology  and  silicon.     Technology  that  has  the  advantages  of  both  biological  and  engineered  materials,  with  the   downsides   of   neither,   is   needed.   Thus,   major   changes   are   required   in   nanoscale   device   designs,   new   functional   materials,   and   novel   software   implementations.   The   general   philosophy  of  building  a  conventional  computational  system  relies  on  the  ability  to  produce   billions  of  highly  controlled  devices  that  respond  the  same  way  to  a  well-­‐‑defined  stimulus   or   signal.   In   contrast,   neuromorphic   circuits   elements   (especially   synapses)   intrinsically   are   expected   to   respond   differently   depending   on   their   past   history.   Consequently,   the   design  as  well  as  the  implementation  of  new  architectures  must  be  dramatically  modified.     EXISTING  NEUROMORPHIC  SYSTEMS     Architectures       A   range   of   computing   architectures   can   support   some   form   of   neuromorphic   computing   (see  Table  3  in  the  appendix  for  a  partial  list  of  historical  efforts  to  build  chips  that  directly   implement   neural   network   abstractions).   At   one   end   of   the   spectrum   are   variations   of   general-­‐‑purpose   CPU   architectures   with   data   paths   that   are   optimized   for   execution   of   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     11   mathematical   approximations   of   neural   networks.   These   artificial   neural   network   accelerators   support   the   fast   matrix   oriented   operations   that   are   the   heart   of   neural   network   methods.   This   architectural   approach   can   beat   general-­‐‑purpose   CPUs   in   power   efficiency   by   a   large   margin   by   discarding   those   features   that   are   not   needed   by   neural   network  algorithms  and  may  be  well  suited  for  integration  into  existing  systems  as  neural   network  accelerators  (see  Figure  4).     At  the  other  end  of  the  spectrum  are  direct  digital  or  analog  implementations  of  networks   of  relatively  simple  neurons,  typically  based  on  an  abstract  version  of  a  leaky  integrate-­‐‑ and-­‐‑fire  (LIF)  neuron.  These  have  discrete  implementations  of  synapses  (simple  storage),   soma,  and  axons  (integrating  and  transmitting  signals  to  other  neurons).  Implementations   can  be  in  analog  circuits  or  digital  logic.  Analog  implementations  of  spike  timing  dependent   plastic  (STDP)  synapses—a  form  of  learning  synapse—have  been  demonstrated  with  a  few   transistors,  and  an  analog  leaky  integrate-­‐‑and-­‐‑fire  model  of  the  soma  required  around  20   transistors.  Axons  in  this  case  are  implemented  as  conventional  wires  for  local  connections   within  a  chip.     Figure  4.  Variation  of  CPU  data  path  optimized  for  execution  of  a  mathematical  approximation  of  a   neural  network.  This  variation  shows  dramatic  power  reduction  compared  to  general  purpose  CPU  when   executing  key  machine  learning  kernels  including  neural  network  execution.  It  uses  a  relatively  simple  data   path,  and  many  fixed-­‐‑width  multiple/add  units  enable  this  architecture  to  perform  well  on  the  dense  floating-­‐‑ point  workload  characteristic  of  neural  network  training.  This  type  of  architecture  is  aimed  at  a  low-­‐‑power   accelerator  add-­‐‑on  to  existing  CPUs.  Figure:  Chen,  Tianshi,  et  al.  IEEE  Micro  (2015):  24-­‐‑32.   For  long-­‐‑distance  (off-­‐‑chip)  signaling,  an  (electronic  or  optical)  analog  to  digital  to  analog   conversion  can  be  used.  The  digital  equivalent  implementation  can  take  considerably  more   transistors   depending   on   the   degree   of   programmability.   The   TrueNorth   system,   for   example,   is   estimated   to   require   about   10   times   as   many   transistors   as   does   an   analog   equivalent;  many  of  these  transistors  are  used  to  provide  a  highly  programmable  soma,  but   do  not  currently  support  on-­‐‑chip  learning.  Power  differences  between  analog  and  digital   implementations  are  also  significant,  with  analog  being  several  orders  of  magnitude  more   efficient.  Even  the  best  current  analog  chips,  however,  are  four  to  five  orders  of  magnitude   less  power  efficient  than  biological  neurons.       In  this  report,  we  focus  on  the  abstract  hardware  implementation  of  an  abstraction  of  a   neural  type  network.  Clearly  novel  devices  based  on  new  functional  materials  will  have  a   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     12   major   impact   for   such   an   implementation.   Therefore,   collaborative   research   in   the   synthesis  of  new  nanostructured  materials,  design  and  engineering  of  nanoscaled  devices   and   implementation   of   creative   architectures   is   needed.   This   is   precisely   the   approach   necessary   to   dramatically   improve   power   and   density   needed   to   reach   “brain-­‐‑like”   performance  and  “brain-­‐‑like”  power  levels.   Demonstrations     Recent  neuromorphic  processor  test  chips  have  typically  implemented  256  neurons  and   65K–256K   synapses,   whereas   IBM’s   TrueNorth   chip,   announced   in   2015,   contained   one   million  neurons  and  256  million  synapses.  Some  groups  are  experimenting  with  test  chips   including   novel   synapse   devices   based   on   memristors.   Full-­‐‑system-­‐‑scale   neuromorphic   prototypes   under   development   include   the   one   billion-­‐‑neuron   SpiNNaker   hybrid   CPU/Neuron   project   at   the   University   of   Manchester   and   the   wafer-­‐‑scale   neuromorphic   hardware  system  “FACETS”  being  developed  at  the  University  of  Heidelberg  that  contains   180K  neurons  and  4  x  107  synapses  per  wafer.  The  completed  system  should  contain  many   such  wafers.  The  FACETS  system  is  unique  in  that  it  supports  more  than  10,000  synapses   per   neuron,   making   it   potentially   able   to   simulate   systems   that   are   more   biologically   plausible  and  perhaps  more  powerful.   Project  Name   Programmable   Structure   Component   Complexity   (Neuron/Synapse)   On-­‐Chip   Learning   Materials/Devices   Desired   Neurons  and   synapses   <  5  /  <  5   Yes   Novel  Materials?   Darwin6   Neurons  and  synapses   >  5  /  >  5   Yes   Fabbed  with  existing   CMOS  processes   DANNA1   Neurons  and  synapses   2  /  2   Yes   FPGA,  ASIC   TrueNorth2   Fixed  (Synapses   on/off)   10  /  3   No   Fabbed  with  existing   CMOS  processes   Neurogrid3   Fixed  (Synapses   on/off)   79  /  8   No   Fabbed  with  existing   CMOS  processes   BrainScaleS4   Neurons  and  synapses   Variable   Yes   Wafer-­‐Scale   ASIC   SpiNNaker5   Neurons  and  synapses   Variable   Yes   ARM  Boards,  Custom   Interconnection     Table  2.  Comparison  of  some  recent  neuromorphic  device  implementations.   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     13   A  common  architectural  pattern  in  neuromorphic  hardware  is  to  arrange  the  synapses  in  a   dense   crossbar   configuration   with   separate   blocks   of   logic   to   implement   learning   in   the   synapse  and  integrate-­‐‑and-­‐‑fire  for  the  soma  (see  Figure  5).  A  challenge  for  such  systems  is   the  fan-­‐‑in  fan-­‐‑out  ratios  for  the  crossbars.  Real  biological  neurons  can  have  up  to  20,000   synapse  connections  per  neuron.  Existing  neuromorphic  chips  tend  to  limit  the  number  of   synapses  to  256  per  neuron.       We  currently  do  not  understand  precisely  when  or  if  artificial  neural  networks  will  need  to   have  the  number  of  connections  that  approach  those  in  biological  systems.  However,  some   deep   learning   networks   in   production   have   layers   with   all-­‐‑to-­‐‑all   connections   between   many   thousand   neurons,   but   these   approaches   also   use   methods   to   “regularize”   the   networks  by  dropping  some  connections.  It  might  be  possible  to  get  by  with  limitations  in   the  number  of  synapses  per  neuron  in  the  thousands.  It  has  also  recently  been  determined   that   dendrites   are   not   just   passive   channels   of   communication   from   the   synapse   to   the   soma,  but  that  they  might  also  play  a  role  in  pattern  recognition  by  filtering  or  recognizing   patterns   of   synaptic   inputs   and   transmitting   potentials   only   in   some   cases.   This   might   require  that  we  revise  the  ideas  of  simple  synapses  connected  by  wires.     Figure  5.  Example  of  a  simple  neuromorphic  architecture.  This  diagram  illustrates  the  dominance  of  the   synapse  crossbar  in  most  neuromorphic  architectures  implementations  and  explains  in  part  why  most  groups   are  focused  on  implementation  of  synapses  as  the  key  scalable  component.  Since  synapse  area  will  dominate   most  designs,  it  is  imperative  that  designs  minimize  synapse  area.   For  historical  comparison,  a  list  of  early  neuromorphic  chips  and  their  scale  is  available  in   Table  3  in  the  Appendix.  While  many  of  these  implementations  have  produced  considerable   advances,  none  are  based  on  developing  completely  novel  approaches  nor  based  on  new   neuromorphic  materials/devices  nor  approach  the  performance  expected  from  a  brain-­‐‑like   device.         Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     14   IMPLEMENTATION  NEEDS     In  this  section,  we  will  outline  the  important  concepts  and  building  blocks  that  will  most   likely  be  used  to  implement  a  neuromorphic  computer.   Neuromorphic  Concepts     The  following  concepts  play  an  important  role  in  the  operation  of  a  system,  which  imitates   the  brain.  It  should  be  mentioned  that  sometimes  the  definitions  listed  below  are  used  in   slightly  different  ways  by  different  investigators.     Spiking.   Signals   are   communicated   between   neurons   through   voltage   or   current   spikes.   This   communication   is   different   from   that   used   in   current   digital   systems,   in   which   the   signals   are   binary,   or   an   analogue   implementation,   which   relies   on   the   manipulation   of   continuous  signals.  Spiking  signaling  systems  are  time  encoded  and  transmitted  via  “action   potentials”.     Plasticity.  A  conventional  device  has  a  unique  response  to  a  particular  stimulus  or  input.  In   contrast,   the   typical   neuromorphic   architecture   relies   on   changing   the   properties   of   an   element  or  device  depending  on  the  past  history.  Plasticity  is  a  key  property  that  allows  the   complex  neuromorphic  circuits  to  be  modified  (“learn”)  as  they  are  exposed  to  different   signals.     Fan-­‐‑in/fan-­‐‑out.   In   conventional   computational   circuits,   the   different   elements   generally   are   interconnected   by   a   few   connections   between   the   individual   devices.   In   the   brain,   however,   the   number   of   dendrites   is   several   orders   of   magnitude   larger   (e.g.,   10,000).   Further   research   is   needed   to   determine   how   essential   this   is   to   the   fundamental   computing  model  of  neuromorphic  systems.     Hebbian   learning/dynamical   resistance   change.   Long-­‐‑term   changes   in   the   synapse   resistance   after   repeated   spiking   by   the   presynaptic   neuron.   This   is   also   sometimes   referred  to  as  spike  time-­‐‑dependent  plasticity  (STDP).  An  alternative  characterization  in   Hebbian  learning  is  “devices  that  fire  together,  wire  together”.     Adaptability.   Biological   brains   generally   start   with   multiple   connections   out   of   which,   through   a   selection   or   learning   process,   some   are   chosen   and   others   abandoned.   This   process  may  be  important  for  improving  the  fault  tolerance  of  individual  devices  as  well  as   for  selecting  the  most  efficient  computational  path.  In  contrast,  in  conventional  computing   the  system  architecture  is  rigid  and  fixed  from  the  beginning.       Criticality.  The  brain  typically  must  operate  close  to  a  critical  point  at  which  the  system  is   plastic  enough  that  it  can  be  switched  from  one  state  to  another,  neither  extremely  stable   nor  very  volatile.  At  the  same  time,  it  may  be  important  for  the  system  to  be  able  to  explore   many  closely  lying  states.  In  terms  of  materials  science,  for  example,  the  system  may  be   close  to  some  critical  state  such  as  a  phase  transition.       Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     15     Accelerators.   The   ultimate   construction   of   a   neuromorphic–based   thinking   machine   requires   intermediate   steps,   working   toward   small-­‐‑scale   applications   based   on   neuromorphic  ideas.  Some  of  these  types  of  applications  require  combining  sensors  with   some  limited  computation.         Building  Blocks       In  functional  terms,  the  simplest,  most  naïve  properties  of  the  various  devices  and  their   function  in  the  brain  areas  include  the  following.   1.   Somata  (also  known  as  neuron  bodies),  which  function  as  integrators  and   threshold  spiking  devices   2.   Synapses,  which  provide  dynamical  interconnections  between  neurons   3.   Axons,  which  provide  long-­‐‑distance  output  connection  between  a  presynaptic  to  a   postsynaptic  neuron   4.   Dendrites,  which  provide  multiple,  distributed  inputs  into  the  neurons     To  implement  a  neuromorphic  system  that  mimics  the  functioning  of  the  brain  requires   collaboration   of   materials   scientists,   condensed   matter   scientists,   physicists,   systems   architects,   and   device   designers   in   order   to   advance   the   science   and   engineering   of   the   various  steps  in  such  a  system.  As  a  first  step,  individual  components  must  be  engineered   to  resemble  the  properties  of  the  individual  components  in  the  brain.       Synapse/Memristor.   The   synapses   are   the   most   advanced   elements   that   have   thus   far   been   simulated   and   constructed.   These   have   two   important   properties:   switching   and   plasticity.  The  implementation  of  a  synapse  is  frequently  accomplished  in  a  two-­‐‑terminal   device  such  as  a  memristor.  This  type  of  devices  exhibits  a  pinched  (at  V=0),  hysteretic  I-­‐‑V   characteristic.         Soma/Neuristor.  These  types  of  devices  provide  two  important  functions:  integration  and   threshold   spiking.   Unlike   synapses,   they   have   not   been   investigated   much.   A   possible   implementation  of  such  a  device  consists  of  a  capacitor  in  parallel  with  a  memristor.  The   capacitance   (“integration”)   and   spiking   function   can   be   engineered   into   a   single   two-­‐‑ terminal  memristor.     Axon/Long  wire.   The   role   of   the   axon   has   commonly   (perhaps   wrongly)   been   assumed   simply  to  provide  a  circuit  connection  and  a  time  delay  line.  Consequently,  little  research   has  been  done  on  this  element  despite  the  fact  that  much  of  the  dissipation  may  occur  in   the  transmission  of  information.  Recent  research  indicates  that  the  axon  has  an  additional   signal-­‐‑conditioning  role.  Therefore,  much  more  research  is  needed  to  understand  its  role   and  how  to  construct  a  device  that  resembles  its  function.     Dendrite/Short  wire.  The  role  of  dendrites  is  commonly  believed  simply  to  provide  signals   from  multiple  neurons  into  a  single  neuron.  This  in  fact  emphasizes  the  three-­‐‑dimensional   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     16   nature  of  the  connectivity  of  the  brain.  While  pseudo-­‐‑3D  systems  have  been  implemented   in   multilayer   (~8)   CMOS-­‐‑based   architecture,   a   truly   3D   implementation   needs   further   research  and  development.  In  addition,  recent  work  in  neuroscience  has  determined  that   dendrites   can   also   play   a   role   in   pattern   detection   and   subthreshold   filtering.   Some   dendrites  have  been  shown  to  detect  over  100  patterns.     Fan-­‐‑in/Fan-­‐‑out.   Some   neurons   have   connections   to   many   thousands   of   other   neurons.     One  axon  may  perhaps  connect  to  ten  thousand  or  more  dendrites.  Current  electronics  is   limited   to   fan-­‐‑in/fan-­‐‑out   of   a   few   tens   of   terminals.   New   approaches   to   high-­‐‑radix   connections  may  be  needed;  currently,  crossbars  are  used  in  most  neuromorphic  systems   but  they  have  scaling  limits.     Many   of   the   needed   functions   can   be   (and   have   been)   implemented   in   complex   CMOS   circuits.  However,  these  not  only  occupy  much  real  estate,  but  also  are  energy  inefficient.   The  latter  perhaps  is  a  crucial  fundamental  limitation  as  discussed  above.  Thus,  for  the  next   step  in  the  evolution  of  brain-­‐‑like  computation,  it  is  crucial  to  build  these  types  of  devices  from   a  single  material  that  is  sufficiently  flexible  to  be  integrated  at  large-­‐‑scale  and  have  minimal   energy  consumption.       PROPOSED  IMPLEMENTATION     Architecture     Ultimately,   an   architecture   that   can   scale   neuromorphic   systems   to   “brain   scale”   and   beyond  is  needed.  A  brain  scale  system  integrates  approximately  1011  neurons  and  1015   synapses   into   a   single   system.   The   high-­‐‑level   neuromorphic   architecture   illustrated   in   Figure  1  consists  of  several  large-­‐‑scale  synapse  arrays  connected  to  soma  arrays  such  that   flexible   layering   of   neurons   (including   recurrent   networks)   is   possible   and   that   off-­‐‑chip   communication   uses   the   address   event   representation   (AER)   approach   to   enable   digital   communication   to   link   spiking   analog   circuits.   Currently,   most   neuromorphic   designs   implement  synapses  and  somata  as  discrete  sub-­‐‑circuits  connected  via  wires  implementing   dendrites  and  axons.  In  the  future,  new  materials  and  new  devices  are  expected  to  enable   integrated  constructs  as  the  basis  for  neuronal  connections  in  large-­‐‑scale  systems.  For  this,   progress   is   needed   in   each   of   the   discrete   components   with   the   primary   focus   on   identification   of   materials   and   devices   that   would   dramatically   improve   the   implementation  of  synapses  and  somata.     One  might  imagine  a  generic  architectural  framework  that  separates  the  implementation  of   the   synapses   from   the   soma   in   order   to   enable   alternative   materials   and   devices   for   synapses  to  be  tested  with  common  learning/spiking  circuits  (see  Figure  6).  A  reasonable   progression   for   novel   materials   test   devices   would   be   the   following:   (1)   single   synapse-­‐‑ dendrite-­‐‑axon-­‐‑soma  feasibility  test  devices,  (2)  chips  with  dozens  of  neurons  and  hundreds   of  synapses,  followed  by  (3)  demonstration  chips  with  hundreds  of  neurons  and  tens  of   thousands  of  synapses.   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     17   Once  hundreds  of  neurons  and  tens  of  thousands  of  synapses  have  been  demonstrated  in  a   novel   system,   it   may   be   straightforward   to   scale   these   building   blocks   to   the   scale   of   systems  competitive  with  the  largest  CMOS  implementations.     State-­‐‑of-­‐‑the-­‐‑art  neural  networks  that  support  object  and  speech  recognition  can  have  tens   of  millions  of  synapses  and  networks  with  thousands  of  inputs  and  thousands  of  outputs.   Simple   street-­‐‑scene   recognition   needed   for   autonomous   vehicles   require   hundreds   of   thousands  of  synapses  and  tens  of  thousands  of  neurons.  The  largest  networks  that  have   been  published—using  over  a  billion  synapses  and  a  million  neurons—have  been  used  for   face  detection  and  object  recognition  in  large  video  databases.     Figure  6.  Block  diagram  of  a  hybrid  neuromorphic  processor  for  synapse  materials  testing.  The  idea  is   that  novel  materials  could  be  tested  in  a  “harness”  that  uses  existing  CMOS  implementations  of  learning  and   soma.  A  framework  such  as  this  could  be  used  to  accelerate  testing  of  materials  at  some  modest  scale.   Properties   Development   of   neuromorphic   computers,   materials   and/or   devices   are   needed   that   exhibit  some  (or  many)  of  the  following  properties:     1.   Multistate  behavior,  in  which  a  physical  property  may  have  different  values  for  the   same  control  parameters,  depending  on  past  history.   2.   Sensitivity  to  external  stimuli  such  as  current,  voltage,  light,  H  field,  temperature  or   pressure  to  provide  desirable  functionalities.   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     18   3.   Threshold  behavior,  in  which  the  material  may  drastically  change  its  properties   after  repetitive  application  of  the  same  stimulus.   4.   Fault  tolerance,  so  that  the  responses  are  repeatable  in  a  statistical  sense,  but  not   necessarily  with  extremely  high  reliability.   5.   Nonvolatility,  such  that  the  properties  are  maintained  for  a  long  time  without  the   need  for  refreshing  processes  or  energy  dissipation  to  hold  state.   6.   Temperature  window,  in  which  the  properties  can  be  controlled  and  maintained.   7.   Insensitivity  to  noise,  through  phenomena  such  as  stochastic  resonances  caused  by   inherent  nonlinearities  in  the  material.  Counter  intuitively;  the  addition  of  noise  to  a   periodic  signal  may  enhance  its  intensity.   8.   Low  energy,  in  which  switching  from  one  state  to  another  is  obtained  and   maintained  with  low  dissipation  without  the  need  for  energy-­‐‑costly  refreshing.   9.   Compatibility  with  other  materials  already  in  use  in  these  types  of  systems.   In  order  to  make  this  program  a  reality,  several  material-­‐‑specific  needs  must  be  met.  In   general,  the  types  of  material  systems  that  have  been  investigated  in  this  context  include   strongly   correlated   oxides,   phase   change   materials,   metal   inclusions   in   insulators,   spin   torque   devices,   ionic   liquid-­‐‑solid   interfaces,   and   magnetic   nanostructures.   In   addition,   many   of   the   complex   materials   are   close   to   some   type   of   electronic   and/or   structural   instability.   The   use   of   these   materials   for   neuromorphic   applications   requires   extensive   knowledge   of   their   behavior   under   highly   nonlinear,   nonequilibrium   conditions   in   heterogeneous   structures   at   the   appropriate   nanometer   scale,   presenting   an   ambitious   materials/condensed   matter   challenge.   Because   of   the   broad   range   of   materials   and   systems,  this  cannot  be  cast  as  a  single,  universal  aim.       Specifically,   researchers   must   quantitatively   address   issues   related   to   synthesis,   characterization  (e.g.,  static,  dynamic,  and  in  operando),  measurements  at  short  time  scales,   interactions   with   different   types   of   electromagnetic   radiation,   and   nanoscale   inhomogeneities  and  defects.  The  problem  is  complex  enough  that  it  requires  the  attention   of  multiple  investigators  with  different  expertise  and  techniques.  Much  of  this  work  can  be   done   in   small-­‐‑scale   laboratories   such   as   at   universities;   however,   certain   resources   are   available  and  accessible  only  at  major  facilities  such  as  national  laboratories.   Devices     The  main  basis  for  the  current  digital  computers  is  a  three-­‐‑terminal  device  that  has  gain:   the  transistor.  In  this  device,  the  drain  current  is  controlled  by  the  application  of  a  gate   voltage,  as  shown  in  Figure  7.  For  a  fixed  gate  voltage,  the  I-­‐‑V  characteristic  is  reversible.   The  control  is  provided  by  the  changes  in  the  output  current  as  a  function  of  gate  voltage.   In  this  case,  for  fixed  parameters  the  output  is  always  the  same.  Typically,  these  devices  are   built   from   ultra-­‐‑pure   semiconductors   (e.g.,   Si,   GaAs)   where   extreme   control   has   to   be   exercised  to  minimize  the  defect  density.       Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     19     One   possible   implementation   of   a   neuromorphic   synapse   is   a   hysteretic,   two-­‐‑terminal   device:  the  memristor  (see  Figure  8).  Typically,  these  types  of  devices  exhibit  a  pinched  (at   zero   voltage)   I-­‐‑V   characteristic   that   is   hysteretic.   The   type   of   hysteresis   depends   on   the   particular   device   implementation   and   material.   An   experimentally,   easily   accessible   memristor   material   can   be   built   from   a   strongly   correlated   oxide   such   as   TiO2.   The   behavior  of  a  memristor  has  been  well  established,  although  the  detailed  mechanism  and   how  to  modify  it  are  still  being  investigated.  This  type  of  I-­‐‑V  characteristics  lends  itself  to   plasticity,  namely,  changing  properties  depending  on  the  past  history.     It   is   expected   that   data   transfer   will   be   considerably   reduced   in   the   type   of   new   neuromorphic  architectures,  with  the  consequent  energetic  savings.  However,  this  cannot   be   completely   eliminated   as   exemplified   by   the   preponderance   of   connections   through   dendrites  and  axons  in  the  brain.  As  a  consequence,  the  interesting  issue  arises  whether  it   would  be  possible  to  replace  some  of  the  electrically  based  data  transfer  to  more  energy   efficient  optical  communications  using  reconfigurable  optical  elements  and  antennas.  This   may  even  be  the  basis  of  the  development  of  a  largely  optically  based  data  processor.   Materials   The   development   of   neuromorphic   devices   also   requires   the   use   of   strongly   correlated,   possibly  complex,  heterogeneous  materials  and  heterostructures  that  are  locally  active  so   that  they  can  be  used  to  create  an  action  potential.  In  many  systems  that  are  used  as  a  basis   for  unconventional  neuromorphic  computing,  the  ultimate  aim  is  to  control  the  dynamical   resistance  (i.e.,  I-­‐‑V  characteristic)  of  a  material  or  device.  The  control  parameters  can  be   categorized   generally   into   two   major   classes:   electronically   driven   and   defect   driven.   Figure 8. I-V characteristic of a memristor, the basis of a possible neuromorphic architecture. This I-V is controlled by the history of the device. Figure 7. I-V characteristic of a transistor, the basis of von Neumann architecture. This I-V is controlled by the voltage applied to the gate. Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     20   Within  each  class,  many  materials  systems  and  devices  have  been  investigated  in  different   contexts   and   depth.   A   (probably)   non-­‐‑exhaustive   list   is   given   in   Tables   5   and   6   in   the   Appendix.     Electronically  driven  materials  rely  on  changes  of  a  physical  property  due  to  variation  in   some   control   parameter   such   as   temperature,   current,   voltage,   magnetic   field,   or   electromagnetic   irradiation.   Generally,   they   become   useful   if   they   exhibit   negative   differential   resistance   or   capacitance   and/or   “threshold   switching.”   While   for   some   systems  the  basic  physics  is  well  understood,  for  others,  even  the  underlying  phenomena   are   controversial.   The   types   of   materials   that   are   actively   being   investigated   and   the   fundamental  issues  that  are  being  addressed  are  related  to  strong  electronic  correlations,   phase   transitions,   charge,   and   spin   propagation.   In   addition,   efforts   are   underway   to   identify  applications  in  the  areas  of  spintronics,  ferroelectric  storage,  and  sensors.   Electronically   driven   transitions   include   spin   torque,   ferroelectric,   phase   change,   and   metal-­‐‑insulator   transition   based   devices.   The   detailed   physics   of   these   systems   differs   significantly  depending  on  the  phenomena  and  the  material  system.  In  general,  however,  a   scientific  bottleneck  exists  because  the  detailed  atomic-­‐‑scale  dynamics  of  these  systems  is   still  largely  unknown.     Defect  driven  materials  take  advantage  of  some  of  the  fundamental  properties,  which  are   strongly   affected   by   the   presence   of   inhomogeneities.   These   include   the   formation   of   metallic  filaments,  inhomogeneous  stress,  and  uniform  and  nonuniform  oxygen  diffusion.   Strong   debates   arise   concerning   how   well   these   phenomena   can   be   controlled.   Nevertheless,   a   number   of   experiments   have   proven   that   the   inhomogeneities   can   be   controlled  by  control  parameters  such  as  voltage,  temperature,  or  electric  fields.  Typical   systems  of  this  type  are  oxygen  ionic  diffusion  in  oxides,  formation  of  metallic  filaments,   and  ionic  front  motion  across  metal-­‐‑insulator-­‐‑metal  trilayers.  Moreover,  the  endurance  of   these  types  of  devices  has  been  shown  to  exceed  one  trillion  cycles—not  as  much  as  DRAM   or  SRAM—although  research  is  expected  to  significantly  increase  this.     Optically  controlled  materials  and  devices  had  a  much  more  limited  use  in  this  context.   This   probably   is   because   the   eventual   system   that   will   emerge   will   likely   be   mostly   electronic.   On   the   other   hand,   the   development   of   tunable   optical   elements   may   add   an   additional  functionality  to  materials,  which  may  be  useful  partially  in  this  context.   Fault  tolerance  is  one  of  the  principal  requirements  of  the  materials  to  be  used.  Defects   and   their   effect   play   a   crucial   role   in   the   behavior   of   these   materials   especially   when   incorporated  into  devices.  This  has  not  been  the  case  with  the  materials  that  have  been   used  until  now  in  silicon-­‐‑based  technology,  which  strives  for  ever  higher  perfection  Thus,   the  new,  fault-­‐‑tolerant  materials  may  actually  improve  device  performance.     Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     21   OPEN  ISSUES       The  most  important  general  issue  that  needs  extensive  research  and  is  not  clearly  defined   is   how   to   integrate   individual   devices   into   a   working   (although   limited)   system   (“accelerator”)   that   will   serve   as   a   proof   of   concept.   Moreover,   this   system   should   be   potentially  scalable,  although  the  exact  way  to  do  this  may  not  be  known  at  present  time.   Below  we  list  some  of  the  open  issues  that  arise  when  considering  materials/devices  and   systems  almost  independently.           Materials/Devices     Many  open  issues  and  questions  remain  regarding  properties  of  materials  and  devices  used   and  proposed  for  neuromorphic  implementations.  Because  it  is  practically  impossible  to   produce  a  comprehensive  literature  review  in  this  brief  report,  we  list  here  a  few  striking   examples.   Resistive  switching.  Some  memristor  devices  use  as  a  basis  the  metal-­‐‑insulator  transition   of  simple  transition-­‐‑metal  oxides.  The  switching  mechanism  in  these  is  based  on  a  first-­‐‑ order   electronic   transition   that   is   generally   coincident   with   a   symmetry-­‐‑changing   structural   phase   transition.   Resistive   devices   based   on   these   materials,   so-­‐‑called   locally   active   memristors,   exhibit   unusual   hysteretic   I-­‐‑V   characteristics   of   different   types.   They   can  show  oscillations  in  the  presence  of  a  DC  bias,  can  inject  previously  stored  energy  into  a   circuit   to   provide   power   amplification   or   voltage   amplification   for   electrical   pulses   or   spikes,  and  can  exhibit  chaos  under  controlled  conditions.  The  physics  of  these  materials  is   still  being  investigated,  although  the  properties  can  be  controlled  well.  For  these  types  of   strongly  correlated  materials,  researchers  continue  to  debate  intensely  regarding  the  role   that  Mott,  Peierls,  Hubbard,  or  Slater  mechanisms  play  in  the  transition,  the  relevant  time   scales,  the  correlation  of  structural  and  resistive  transitions,  the  effect  of  proximity,  and  the   way  these  effects  may  be  controlled  (e.g.,  by  epitaxial  clamping).  These  issues  are  related  to   the   specific   properties   of   the   materials   in   questions   and   are   being   investigated   in   many   other   contexts.   Nevertheless,   their   properties   can   be   controlled   sufficiently   well   that   a   number  of  neuromorphic  devices  have  been  implemented.     Filament  formation.  Many  of  the  neuromorphic  devices  utilized  as  synaptic  memory  rely   on  the  formation  of  filaments  or  conductive  channels  in  the  material  between  two  metallic   electrodes.   These   may   occur   because   of   an   intrinsic   phase   transition   present,   such   as   amorphous-­‐‑crystalline   or   metal-­‐‑insulator   transitions,   or   because   of   redistribution   of   defects/ions  that  modulates  local  electrical  properties.  Understanding  the  mechanisms  in   the  formation  and  destruction  of  filaments  and  the  effect  of  preexisting  defects  is  crucial  for   understanding   the   reproducibility   of   these   devices.   Particularly   important   is   the   role   of   pre-­‐‑existing  defects  and  the  way  these  modify  the  formation  of  filaments.   Spin  torque  switching.  In  spin  torque  devices,  the  resistive  transition  is  controlled  by  the   magnitude   of   a   current   through   (in   principle)   a   four-­‐‑layer   device.   Spin   torque   devices   already   have   been   implemented   for   unrelated   spintronics   applications   and   are   being   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     22   incorporated   into   commercial   nonvolatile   magnetic   random   access   memory.   The   magnitude  of  the  spin  torque  effect  and  the  role  the  Oersted  field  are  subjects  of  important   research,   especially   when   the   size   of   the   devices   is   reduced   to   the   nanoscale   or   when   devices  are  closely  packed,  as  needed  here.  Understanding  defects  and  their  effect  on  the   stability   of   ferromagnetic   materials   is   important   in   order   to   improve   the   endurance   of   materials.     Ionic  diffusion.  Defect-­‐‑induced  devices  and  materials  rely  on  the  controlled  formation  of   filaments   or   the   diffusion   of   ionic   fronts   between   two   metallic   electrodes   to   control   the   conductance  and  thus  provide  large  resistance  switching.  Examples  are  the  formation  of   metallic   shorts   across   metal-­‐‑insulator-­‐‑metal   trilayers,   the   change   in   the   resistance   of   an   insulator  due  to  an  ionic  front  diffusion  between  two  metallic  electrodes,  and  the  diffusion   of  oxygen  induced  by  a  metallic  tip  from  an  ionic  liquid  into  an  oxide.  For  these  types  of   defect-­‐‑based  devices  and  phenomena,  basic  research  is  needed  on  several  major  issues:  the   importance  of  preexisting  defects  on  the  diffusion  of  light  elements,  thermophoresis,  the   formation  of  filaments,  the  reproducibility  from  cycle  to  cycle,  and  electromigration  and  the   consequent  effects  on  endurance.   Nano  and  mesoscale.   A   number   of   open   issues   straddle   both   the   materials   and   devices   contemplated  here.  In  particular,  incorporating  these  materials  into  computational  systems   will   require   reducing   them   to   nanometer   scale   in   functional   structures.   Therefore,   understanding   the   materials   and   especially   their   interface   behavior   at   the   nano   and   mesoscale  is  crucial.  While  at  large-­‐‑scale  these  phenomena  may  be  well  understood,  when   reduced  to  the  nanoscale,  additional  effects  become  important.  For  instance,  Oersted  fields   and   dipolar   interactions   may   become   more   important   than   at   the   micron   scale   in   spin   torque  devices;  filament  formation  may  become  highly  inhomogeneous  and  uncontrolled  in   defect-­‐‑based  devices;  and  the  importance  and  magnitude  of  enhanced  fields  in  connection   with   roughness   become   especially   important   in   nanoscale   devices   based   on   ionic   conduction.  A  key  issue  in  scaling  down  to  the  nanoscale  is  the  fact  that  while  the  device   heat   capacity   decreases,   its   thermal   resistance   increases,   which   can   lead   to   huge   Joule   heating-­‐‑induced   temperature   increases   and   enormous   temperature   gradients   when   electrically  biased.  This  issue  requires  detailed  studies  of  materials  properties,  ideally  in   operando.   In   many   cases,   complex   interactions   will   appear   at   different   nano   and   meso   scales  which  can  only  be  solved  by  the  experimental  capabilities  available  at  DOE  facilities.   This   will   also   necessarily   include   some   capabilities   for   fabricating   test   structures   using   what  is  traditionally  called  back-­‐‑end-­‐‑of-­‐‑line  (BEOL)  processing  techniques  (i.e.,  not  using   full   CMOS   capabilities).   The   size   scales   and   the   reproducibility   required   for   meaningful   analysis  will  mean  that  some  lithographic  capabilities  at  the  50  nm  and  smaller  scales  will   be   required,   which   can   be   obtained   by   electron   beam   lithography   or   refurbished   and   therefore  relatively  inexpensive  UV  steppers.   System     As  we  consider  the  building  of  large-­‐‑scale  systems  from  neuron  like  building  blocks,  there   are   a   large   number   of   challenges   that   must   be   overcome.   One   challenge   arises   from   the   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     23   need  for  dense  packaging  of  neurons  in  order  to  achieve  comparable  volumes  to  brains.     This  implies  dense  3D  packing  with  a  range  of  problems  associated  with  assembly,  power   delivery,   heat   removal   and   topology   control.   Another   set   of   challenges   arises   from   the   abstract  nature  of  neuro-­‐‑inspired  computation  itself.  How  close  to  nature  must  we  build  to   gain  the  benefits  that  evolution  has  devised?  Can  we  develop  computational  abstractions   that  have  many  of  the  advantages  of  biology  but  are  easier  to  construct  with  non-­‐‑biological   materials   and   non-­‐‑biological   assembly   processes?   How   will   such   systems   be   designed?   How   will   they   be   programmed   and   how   will   they   interact   with   the   vast   computational   infrastructure  that  is  based  on  conventional  technologies?       A  number  of  critical  issues  remain  as  we  consider  the  artificial  physical  implementation  of   a  system  that  partially  resembles  a  brain-­‐‑like  architecture:   1.   What  are  the  minimal  physical  elements  needed  for  a  working  artificial  structure:   dendrite,  soma,  axon,  and  synapse?   2.   What  are  the  minimal  characteristics  of  each  one  of  these  elements  needed  in  order   to  have  a  first  proven  system?   3.   What  are  the  essential  conceptual  ideas  needed  to  implement  a  minimal  system:   spike-­‐‑dependent  plasticity,  learning,  reconfigurability,  criticality,  short-­‐‑  and  long-­‐‑ term  memory,  fault  tolerance,  co-­‐‑location  of  memory  and  processing,  distributed   processing,  large  fan-­‐‑in/fan-­‐‑out,  dimensionality?  Can  we  organize  these  in  order  of   importance?   4.   What  are  the  advantages  and  disadvantages  of  a  chemical  vs.  a  solid-­‐‑state   implementation?   5.   What  features  must  neuromorphic  architecture  have  to  support  critical  testing  of   new  materials  and  building  block  implementations?   6.   What  intermediate  applications  would  best  be  used  to  prove  the  concept?     These   and   certainly   additional   questions   should   be   part   of   a   coherent   approach   to   investigating  the  development  of  neuromorphic  computing  systems.  The  field  could  also   use  a  comprehensive  review  of  what  has  been  achieved  already  in  the  exploration  of  novel   materials,  as  there  are  a  number  of  excellent  groups  that  are  pursuing  new  materials  and   new  device  architectures.  Many  of  these  activities  could  benefit  from  a  framework  that  can   be  evaluated  on  simple  applications.       At  the  same  time,  there  is  a  considerable  gap  in  our  understanding  of  what  it  will  take  to   implement   state-­‐‑of-­‐‑the-­‐‑art   applications   on   neuromorphic   hardware   in   general.   To   date,   most   hardware   implementations   have   been   rather   specialized   to   specific   problems   and   current   practice   largely   uses   conventional   hardware   for   the   execution   of   deep   learning   applications   and   large-­‐‑scale   parallel   clusters   with   accelerators   for   the   development   and   training   of   deep   neural   networks.   Moving   neuromorphic   hardware   out   of   the   research   phase  into  applications  and  end  use  would  be  helpful.  This  would  require  advances  which   support  training  of  the  device  itself  and  to  show  performance  above  that  of  artificial  neural   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     24   networks   already   implemented   in   conventional   hardware.   These   improvements   are   necessary  both  regarding  power  efficiency  and  ultimate  performance.       INTERMEDIATE  STEPS     This   section   identifies   the   major   milestones   needed   toward   the   development   of   a   neuromorphic   computer.   We   should   highlight   that   every   step   must   be   based   on   earlier   steps  and  connected  to  eventual  implementation  of  next  steps.  This  can  be  considerably   advanced   through   the   construction   of   appropriate   compact   theoretical   models   and   numerical  simulations  that  are  calibrated  through  experimentation.  It  is  also  important  to   point   out   that   this   field   is   in   its   earlier   stages   of   development   and   therefore   sufficient   flexibility  should  be  maintained  at  every  stage.  This  should  not  be  viewed  as  a  well-­‐‑defined   development  task  but  as  a  research  project.  Therefore,  it  is  important  that  at  every  stage   several  competing  projects  are  implemented  to  allow  for  the  best  solution  to  emerge.    The   key  ingredients  in  these  intermediate  steps  could  be:   General  Aim.  As  a  general  goal,  it  would  be  desirable  to  develop  well-­‐‑defined  intermediate   application  such  as  needed  in  the  fields  of  vision,  speech,  and  object  recognition  to  prove   the  reality  of  a  program  as  described  here.    Simulations.   There   are   opportunities   to   leverage   large-­‐‑scale   computing   in   the   development  of  simulators  for  neuromorphic  designs  and  to  develop  a  deep  understanding   of  materials  and  device.  These  simulations  could  be  used  to  refine  architectural  concepts,   improve  performance  parameters  for  materials  and  devices,  and  to  generate  test  data  and   signals  to  help  support  accelerated  testing  as  new  materials,  devices  and  prototypes  are   developed.   Devices.   Development   and   engineering   of   novel   devices   perhaps   based   on   some   type   of   memristive  or  optically  bistable  property  is  needed.  This  should  include  incorporation  into   well-­‐‑defined  systems  and  be  based  on  well-­‐‑understood  materials  science.     Material   Science.   Synthesis,   characterization   and   study   of   new   functional,   tunable   materials   with   enhanced   properties   are   needed   to   integrate   into   novel   neuromorphic   devices.         We  envision  the  following  stages  in  the  development  of  such  a  project:                   1.   Identify  conceptual  design  of  neuromorphic  architectures   2.   Identify  devices  needed  to  implement  neuromorphic  computing   3.   Define  properties  needed  for  prototype  constituent  devices   4.   Define  materials  properties  needed     5.   Identify  major  materials  classes  that  satisfy  needed  properties   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     25   6.   Develop  a  deep  understanding  of  the  quantum  materials  used  in  these  applications     7.   Build  and  test  devices  (e.g.,  synapse,  soma,  axon,  dendrite)   8.   Define  and  implement  small  systems,  and  to  the  extent  possible,   integrate/demonstrate  with  appropriate  research  and  development  results  in   programming  languages,  development  and  programming  environments,  compilers,   libraries,  runtime  systems,  networking,  data  repositories,  von  Neumann-­‐‑ neuromorphic  computing  interfaces,  etc.   9.   Identify  possible  “accelerator”  needs  for  intermediate  steps  in  neuromorphic   computing  (e.g.,  vision,  sensing,  data  mining,  event  detection)   10.   Integrate  small  systems  for  intermediate  accelerators   11.   Integrate  promising  devices  into  end-­‐‑to-­‐‑end  system  experimental  chips  (order  10   neurons,  100  synapses)   12.   Scale  promising  end-­‐‑to-­‐‑end  experiments  to  demonstration  scale  chips  (order  100   neurons  and  10,000  synapses)   13.   Scale  successful  demonstration  chips  to  system  scale  implementations  (order   millions  of  neurons  and  billion  synapses)   14.   Scale  successful  demonstration  chips  to  system  scale  implementations  (order   millions  of  neurons  and  billion  synapses)     We  have  outlined  in  this  report  many  of  the  open  issues  and  opportunities  for  architecture   and  materials  science  research  needed  to  realize  the  vision  of  neuromorphic  computing.     The  key  idea  is  that  by  adopting  ideas  from  biology  and  by  leveraging  novel  materials,  we   can   build   systems   that   can   learn   from   examples,   process   large-­‐‑scale   data   adjust   their   behavior  to  new  inputs  and  do  all  of  these  with  the  power  efficiency  of  the  brain.  Taking   steps   in   this   direction   will   continue   the   development   of   data   processing   in   support   of   science  and  society.       CONCLUSIONS     The  main  conclusions  of  the  roundtable  were:   1.   Creating   the   architectural   design   for   neuromorphic   computing   requires   an   integrative,   interdisciplinary   approach   between   computer   scientists,   engineers,   physicists,  and  materials  scientists   2.   Creating   a   new   computational   system   will   require   developing   new   system   architectures  to  accommodate  all  needed  functionalities   3.   One   or   more   reference   architectures   should   be   used   to   enable   comparisons   of   alternative  devices  and  materials   4.   The  basis  for  the  devices  to  be  used  in  these  new  computational  systems  require  the   development   of   novel   nano   and   meso   structured   materials;   this   will   be   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     26   accomplished   by   unlocking   the   properties   of   quantum   materials   based   on   new   materials  physics       5.   The   most   promising   materials   require   fundamental   understanding   of   strongly   correlated   materials,   understanding   formation   and   migration   of   ions,   defects   and   clusters,   developing   novel   spin   based   devices,   and/or   discovering   new   quantum   functional  materials   6.   To  fully  realize  open  opportunities  requires  designing  systems  and  materials  that   exhibit   self-­‐‑   and   external-­‐‑healing,   three-­‐‑dimensional   reconstruction,   distributed   power  delivery,  fault  tolerance,  co-­‐‑location  of  memory  and  processors,  multistate—   i.e.,  systems  in  which  the  present  response  depends  on  past  history  and  multiple   interacting  state  variables  that  define  the  present  state   7.   The   development   of   a   new   brain-­‐‑like   computational   system   will   not   evolve   in   a   single  step;  it  is  important  to  implement  well-­‐‑defined  intermediate  steps  that  give   useful  scientific  and  technological  information       Acknowledgements   We  thank  all  participants  of  the  workshop  for  their  valuable  input  under  much  time   constraints.  We  thank  Juan  Trastoy  for  critical  reading  and  input  for  Table  5.  We  thank   Emily  Dietrich  for  editorial  support.               Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     27   APPENDICES   Acronyms     3D:  Three-­‐‑dimensional       AER:  Address  Event  Representation     ALU:  Arithmetic/Logic  Unit     BEOL:  Back-­‐‑end-­‐‑of-­‐‑line  (BEOL)       CMOS:  Complementary  Metal  Oxide  Semiconductor     CNN:  Convolutional  Neural  Network     CPU:  Central  Processing  Unit     ColdRAM:    Memory  organized  for  infrequent  access  patterns     DNN:  Deep  Neural  Network     DMA:  Logic  for  supporting  Direct  Memory  Access       HotRAM:  Memory  organized  for  frequent  access  patterns     InstRAM:  Instruction  Memory     LIF:  Leaky  Integrate  and  Fire  (neuron)     LTPS:  Long-­‐‑Term  Plasticity  Synapses     MLU:  Machine  Learning  Unit,  functional  units  optimized  for  ML  operations     MU:  Memory  Unit     OutputRAM:  Memory  for  holding  output  of  operations     STDP:  Spike  Time-­‐‑Dependent  Plasticity     STPS:  Short-­‐‑Term  Plasticity  Synapses     VLSI:  Very  Large-­‐‑Scale  Integration   Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     28   Glossary     Axon:  Transmitting  signals  to  other  neurons,  axons  provide  long-­‐‑distance  output   connection  between  a  presynaptic  to  a  postsynaptic  neuron     Charge  trapping:  Mobile  electrons  maybe  trapped  sometimes  in  defects  invariably  present   in  materials     Crystalline-­‐‑amorphous  transition:  Some  materials  change  their  physical  structure  from   an  ordered  crystalline  to  a  completely  disordered  (amorphous)  phase;  the  electronic   properties  of  these  two  phases  maybe  radically  different     Deep  learning:  Deep  learning  is  the  branch  of  machine  learning  that  builds  tools  based  on   deep  (multilayer)  neural  networks     Dendrite:  Providing  multiple,  distributed  inputs  into  the  neurons,  the  role  of  dendrites  is   commonly  believed  simply  to  provide  signals  from  multiple  neurons  into  a  single  neuron     Depress:  To  increase  resistance     Domain  wall  motion:  In  some  cases,  a  magnetic  material  reverts  its  magnetization  by  the   motion  of  a  separation  (“wall”)  in  between  two  well-­‐‑defined  magnetization  areas     Filament  formation:  The  resistance  of  two  electrodes  separated  by  an  insulator  may   change  drastically  if  a  conducting  filament  forms;  this  can  form  because  of  intrinsic  reasons   or  due  to  the  motion  of  atoms  in  the  insulator     Hebbian  learning:  Change  occurs  in  the  synapse  resistance  after  repeated  spiking  by  the   presynaptic  neuron  before  the  postsynaptic  neuron;  this  is  also  sometimes  referred  to  as   spike  time-­‐‑dependent  plasticity  (STDP)     Ionic  motion:  Certain  solids  and  liquids  ions  may  move  under  different  driving  forces  such   as  high  voltages,  currents  and  temperature     Learning:  Conditioning;  process  by  which  the  various  elements  of  the  system  change  and   readjust  depending  on  the  type  of  stimuli  they  receive     Machine  learning:  The  branch  of  computer  science  that  deals  with  building  systems  that   can  learn  from  and  make  predictions  on  data     Magnetic  tunnel  junction:  A  device  based  on  quantum  mechanical  tunneling  that  changes   it  resistance  depending  on  the  relative  magnetization  of  the  magnetic  electrodes;  the  basis   for  most  spintronics  applications     Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     29   Parallel  computing:  Differing  from  serial  von  Neumann,  parallel  computing  is   distinguished  by  the  kind  of  interconnection  between  processors  and  between  processors   and  memory;  most  parallel  computers  today  are  large-­‐‑ensembles  of  von  Neumann   processors  and  memory     Plasticity:  Synaptic  resistance;  key  property  that  allows  the  complex  neuromorphic   circuits  to  be  modified  (“learn”)  as  they  are  exposed  to  different  signals       Potentiation:  Increased  in  conductance     Soma:  Neuron  body,  where  primary  input  integration  and  firing  occurs     Synapse:  Space  between  axon  and  dendrites  that  allows  for  signals  to  be  transmitted  from   the  presynaptic  to  the  postsynaptic  neuron.     Vacancy  motion:  In  certain  solids  and  liquids,  the  absence  of  an  ion  (“vacancy”)  may  move   under  different  driving  forces  such  as  high  voltages,  currents  and  temperature     Tables     Table  2  References  (Table  2  above)   1.   Dean,  Mark  E.,  Catherine  D.  Schuman,  and  J.  Douglas  Birdwell.  "Dynamic  adaptive   neural  network  array."  Unconventional  Computation  and  Natural  Computation.   Springer  International  Publishing,  2014.  129-­‐‑141.     2.   Merolla,  Paul  A.,  et  al.  "A  million  spiking-­‐‑neuron  integrated  circuit  with  a  scalable   communication  network  and  interface."  Science  345.6197  (2014):  668-­‐‑673.   3.   Benjamin,  Ben  Varkey,  et  al.  "Neurogrid:  A  mixed-­‐‑analog-­‐‑digital  multichip  system   for  large-­‐‑scale  neural  simulations."  Proceedings  of  the  IEEE  102.5  (2014):  699-­‐‑716.   4.   Brüderle,  Daniel,  et  al.  "A  comprehensive  workflow  for  general-­‐‑purpose  neural   modeling  with  highly  configurable  neuromorphic  hardware  systems."  Biological   cybernetics  104.4-­‐‑5  (2011):  263-­‐‑296.   5.   Furber,  Steve  B.,  et  al.  "Overview  of  the  spinnaker  system  architecture."  Computers,   IEEE  Transactions  on  62.12  (2013):  2454-­‐‑2467.   6.   Shen  JunCheng,  et.  al.  “Darwin:  a  Neuromorphic  Hardware  Co-­‐‑Processor  based  on   Spiking  Neural  Networks”.  SCIENCE  CHINA  Information  Sciences,  DOI:   10.1360/112010-­‐‑977       Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     30     Neural  networks  digital  hardware  implementations   Name   Architecture   Learn   Precision   Neurons   Synapses   Speed   Slice  architectures         Micro  devices  MD-­‐1220 a   Feedforward,  ML   No   1  x  16  bits   8   8   1.9  MCPS   NeuraLogix  NLX-­‐420 a   Feedforward,  ML   No   1-­‐16  bits   16   Off-­‐chip   300  CPS   Philips  Lneuro-­‐1   Feedforward,  ML   No   1-­‐16  bits   16  PE   64   26  MCPS   Philips  Lneuros-­‐2.3   N.A.   No   16-­‐32  bits   12  PE   N.A.   720  MCPS             SIMD         Inova  N64000 a   GP,  SIMD,  Int   Program   1-­‐16  bits   64  PE   256k   870  MCPS         220  MCUPS   Hecht-­‐Nielson  HNC  100-­‐NAP b   GP,  SIMD,  FP   Program   32  bits   4  PE   512k   250  MCPS         Off-­‐chip   64  MCUPS   Hitachi  WSI   Wafer,  SIMD   Hopfield   9  x  8  bits   576   32k   138  MCPS   Hitachi  WSI   Wafer,  SIMD   BP   9  x  8  bits   144   N.A.   300  MCUPS   Neuricam  NC3001  TOTEM   Feedforward,  ML,  SIMD   No   32  bits   1-­‐32   32k   1  GCPS   Neuricam  NC3003  TOTEM   Feedforward,  ML,  SIMD   No     32  bits   1-­‐32   64k   750  MCPS   RC  Module  NM6403   Feedforward,  ML   Program   1-­‐64  x  1-­‐64  bits   1-­‐64   1-­‐64   1200  MCPS             Systolic  array         Siemans         MA-­‐16   Matrix  ops   No   16  bits   16  PE   16  x  16   400  MCPS             Radial  basis  functions         Nestors/Intel  NI1000 c   RBF   RCE,  PNN,  program   5  bits   1  PE   245  x  1024   40  kpat/s   IBM  ZISC036   RBF   ROI   8  bits   36   64  x  36   250  kpat/s   Silicon  recognition  ZISC78   RBF   KNN,  L1.  LSUP   N.A.   78   N.A.   1  Mpat/s             Other  chips         SAND/1   Feedforward,  ML,  RBF   No   40  bits   4  PE   Off-­‐chip   200  MCPS       Kohonen         MCE  MT  19003   Feedforward,  ML   No   13  bits   8   Off-­‐chip   32  MCPS     Table  3.  Historical  hardware  implementations  of  neural  networks.         Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     31     Feature   Description   Clock  Free   Fully  asynchronous   Scale  Free   Activity  can  vary  from  local  to  system  level  scales  depending  upon  context   Symbol  Free   No  single  neuron  or  synapse  represents  any  single  item/concept   Grid  Free   Small  world  network  geometry  allows  feature  integration  from   heterogeneous  and  non  local  brain  areas   Dendritic  Neuron   Nonlinear  signal  processing  via  dendrites  in  each  neuron   Synaptic  Plasticity   Most  synapses  exhibit  plasticity  at  various  time  scales  (secs  to  hrs)   Synaptic  Path  Length   Approx.  constant  number  of  hops  between  different  brain  areas   Dense  Connectivity   Each  neuron  connects  to  between  1000-­‐10000  other  neurons   Modular  Cortex   Six  layered  modular  architecture  that  repeats  across  architecture   Broadcasting   Brain  areas  that  broadcast  signals  (neuromodulatory)  to  all  other  parts       Table  4.  Neuromorphic  system  level  architecture  features.       Table  5.  List  of  materials  systems  for  neuromorphic  applications.  Characteristics  obtained  from  the   literature  of  the  different  material  systems  put  forward  for  neuromorphic  applications.  The  following   abbreviations  are  used:  Vom=Vacancy  motion;  Ff=Filament  formation;  Iom=ionic  motion;  DWm=domain  wall   motion;  Cat=crystalline  amorphous  transition;  Ct=Charge  trapping;  Ps=Polarization  switching;  RT=room   temperature;  SET=change  from  high  to  low  resistance  state,  and  MTJ=magnetic  tunnel  junction.           CATEGORY    SYSTEM   MECHANISM   WRITE   SPEED   ON  /   OFF   DATA   RETENTION  (@   1  Hz)   TEMP  (ºC)     POWER   (W)   ISSUES   REFERENCE   Oxides   HfOX/TiOX/HfOX/T iOX   Vom-­‐front   10s     ns   1000   >  15  min   20-­‐85     10 -­‐4   Abrupt  SET  process  (only  depression   is  possible)   http://onlinelibrary.wiley.com/doi/10.1 002/adma.201203680/abstract   VO2   Ff   1s   100       70   10 -­‐2   Temperature,  memory  duration,  50   V  pulses,  only  potentiation   http://scitation.aip.org/content/aip/jour nal/apl/95/4/10.1063/1.3187531   Nb2O5/Pt   Vom   100s  ns   10   >  500  years   RT   10 -­‐4   Simple  planar  micron  scale  structure   http://ieeexplore.ieee.org/xpl/articleDe tails.jsp?arnumber=1425686   WOX   Vom-­‐front   100s    µs   1.4   >  3  hours   RT   10 -­‐5   Low  ON/OFF  ratio   http://dx.doi.org/10.1109/TED.2014.23 19814   Nb-­‐doped-­‐a-­‐STO   Vom-­‐filament   10s  µs   1000   >  1  day   27-­‐125   10 -­‐4   Electroforming  needed   http://onlinelibrary.wiley.com/doi/10.1 002/adfm.201501019/abstract   Pt/TiO2/Pt   Vom   10  ns   10   >  35,000  years   RT   <10 -­‐4   30  nm  wire  width  crossbar  structure   http://onlinelibrary.wiley.com/doi/10.1 002/adfm.201202170/abstract   Phase   Change   GeSbTe   CAt   1  ns   100   >  3  hours   RT   10 -­‐3       https://www.sciencemag.org/content/3 36/6088/1566.full   Optical   C-­‐Nanotubes   Photo/Electrical   gating   10s         200   >  2  days   RT   10 -­‐6   Very  slow  and  difficult  to  implement   http://onlinelibrary.wiley.com/doi/10.1 002/adma.200902170/full   GaLaSO   Photodark   Ms   1.1       RT   10 -­‐1   Proof  of  concept   http://onlinelibrary.wiley.com/doi/10.1 002/adom.201400472/abstract   Metal   Inclusions   Ag  on  a-­‐Si   Ff   10s    ns   1000   11  days   RT   10 -­‐8   Electroforming,  short  endurance   http://pubs.acs.org/doi/abs/10.1021/nl 073225h   Cosputtered  a-­‐Si   and  Ag   Iom-­‐front   100s  µs   8   5  years   RT   10 -­‐6   Low  ON/OFF  ratio   http://pubs.acs.org/doi/abs/10.1021/nl 904092h   Organic   Au/Pentacene/Si NWs/Si   Ct     120       RT   10 -­‐5   Speed  is  not  clear   http://scitation.aip.org/content/aip/jour nal/apl/104/5/10.1063/1.4863830   Ferro   electric   BTO/LSMO   Tunneling   Ps   10s  ns   300   >  15  min   RT   10 -­‐6       http://www.nature.com/nmat/journal/v 11/n10/full/nmat3415.html   Magnetic   MgO-­‐based  MTJ   DWm       1.1       RT   10 -­‐4   Needs  external  magnetic  field,  Low   ON/OFF  ratio   http://www.nature.com/nphys/journal/ v7/n8/full/nphys1968.html   Liquid-­‐Solid     Ionic   liquid/SmNiO3   Iom   10s  ms   11     35-­‐160     Requires  gating  circuit,  slow   https://doi.org/10.1038/ncomms3676     Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     33   Summary  of  common  inorganic  storage  media  and  corresponding  switching  characteristics   Storage  medium   Switching  mode   ON/OFF  ration   Operation  speed   Endurance  (cycles)   Binary  oxides         MgOx   Unipolar,  bipolar   >10 5   -­‐   >4  x  10 2   AIOx   Unipolar,  bipolar   >10 6   <10  ns;  <10  ns   >10 4   SiOx   Unipolar,  bipolar   >10 7   <100  ps;  <100  ps   >10 8   TIOx   Unipolar,  bipolar   >10 5   <5  ns;  <5  ns   >2  x  10 6   CrOx   Bipolar   >10 2   <4  µs;  <5  µs   >6  x  10 4   MnOx   Unipolar,  bipolar   >10 4   <100  ns;  <200  ns   >10 5   FeOx   Bipolar   >10 2   <10  ns;  <10  ns   >6  x  10 4   CoOx   Unipolar,  bipolar   >5  x  10 3   <20  ns;  <20  ns   >10 3   NiOx   Unipolar,  bipolar   >10 6   <10  ns;  <20  ns   >10 6   CuOx   Unipolar,  bipolar   >10 5   <50  ns;  <50  ns   >1.2  x  10 4   ZnOx   Unipolar,  bipolar   >10 7   <5  ns:  <5  ns   >10 6   GaOx   Bipolar   >10 2   <400  ns;  <600  ns   >10 4   GeOx   Unipolar,  bipolar   >10 9   <20  ns;  <20  ns   >10 6   ZrOx   Unipolar,  bipolar   >10 6   <10  ns;  <10  ns   >10 4   NbOx   Unipolar,  bipolar   >10 8   <100  ns;  <100  ns   >10 7   MoOx   Unipolar,  bipolar   >10   <1  µs;  <1  µs   >10 6   HfOx   Unipolar,  bipolar   >10 5   <300  ps;  <300  ps   >10 10   TaOx   Unipolar,  bipolar   >10 9   <105  ps  ;  <120  ps   >10 12   WOx   Unipolar,  bipolar   >10 4   <300  ns;  <50  ns   >10 8   CeOx   Unipolar,  bipolar   >10 5   <1  µs  :  <200  ns   >10 4   GdOx   Unipolar,  bipolar   >5  x  10 5   <1  ns;  <1  ns   >10 7   YbOx   Unipolar,  bipolar   >10 5   -­‐   >10 5   LuOx   Unipolar,  bipolar   >10 4   <10  ns;  <30  ns   >8  x  10 2             Ternary  and  more  complex  oxides         LaAIO3   Bipolar   >10 4   -­‐   >10 2   SrTiO3   Bipolar   >10 5   <5  ns:  <5  ns   >10 6   BaTiO3   Unipolar,  bipolar   >10 4   <10  ns;  <70  ns   >10 5   LC(or  S)MO   Bipolar   >10 3   <25  ns;  <25  ns   >10 3   PCMO   Bipolar   >10 3   <8  ns;  <8  ns   >10 10   BiFeO3   Unipolar,  bipolar   >10 5   <50  ns;  <100  µs   >10 3             Chalcogenides         Cu2S   Bipolar   >10 6   <100  µs;  <100  µs   >10 5   GeSx   Bipolar   >10 5   <50  ns;  <50  ns   >7.5  x  10 6   Ag2S   Bipolar   >10 6   <200  ns;  <200  ns   -­‐   GexSey   Bipolar   >10 6   <100  ns;  <100  ns   >3.2  x  10 10             Nitrides         AIN   Unipolar,  bipolar   >10 3   <10  ns;  <10  ns   >10 8   SiN   Unipolar,  bipolar   >10 7   <100  ns;  <100  ns   >10 9             Others         a-­‐C   Unipolar,  bipolar   >3  x  10 2   <50  ns;  <10  ns   >10 3   a-­‐Si   Bipolar   >10 7   <5  ns;  <10  ns   >10 8   AgI   Bipolar   >10 6   <50  ns;  <150  ns   >4  x  10 5     Table  6.  Comprehensive  list  of  relevant  properties  for  interesting  materials.  The  operation  speed  is   written  as  ‘set  (write)  speed;  reset  (erase)  speed’.  The  symbol  ‘-­‐‑‘  means  that  no  data  concerning  that   characteristic  is  found.  (after  F.  Pan,  S.  Gao,  C.  Chen,  C.  Song,  F.  Zeng,  Materials  Science  and  Engineering  R  83   (2014)  1–59.)           Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     34   WORKSHOP  LOGISTICS   Roundtable  Participants     Co-­‐‑chairs:     Ivan  K.  Schuller   University  of  California,  San  Diego   Rick  Stevens   Argonne  National  Laboratory  and  University  of  Chicago     Participants  and  Observers:     Nathan  Baker   Pacific  Northwest  National  Laboratory   Jim  Brase   Lawrence  Livermore  National  Laboratory   Hans  Christen   Oak  Ridge  National  Laboratory   Mike  Davies   Intel  Corporation   Massimiliano  Di  Ventra   University  of  California,  San  Diego   Supratik  Guha   Argonne  National  Laboratory   Helen  Li   University  of  Pittsburgh   Wei  Lu   University  of  Michigan   Robert  Lucas   University  of  Southern  California   Matt  Marinella   Sandia  National  Laboratories   Jeff  Neaton   Lawrence  Berkeley  National  Laboratory   Stuart  Parkin   IBM  Research  –  Almaden   Thomas  Potok   Oak  Ridge  National  Laboratory   John  Sarrao   Los  Alamos  National  Laboratory   Katie  Schuman   Oak  Ridge  National  Laboratory   Narayan  Srinivasa   HRL  Laboratories  LLC   Stan  Williams   Hewlett-­‐‑Packard   Philip  Wong   Stanford  University           Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     35   Roundtable  Summary   Roundtable  on  Neuromorphic  Computing:     From  Materials  to  Systems  Architecture     Co-­‐‑chairs:     Ivan  K.  Schuller   University  of  California,  San  Diego   Rick  Stevens   Argonne  National  Laboratory  and  University  of  Chicago     DOE  Contacts:     Robinson  Pino   Advanced  Scientific  Computing  Research   Michael  Pechan   Basic  Energy  Sciences     Purpose   The  Neuromorphic  Computing:  From  Materials  to  Systems  Architecture  Study  Group   convened  national  laboratory,  university,  and  industry  experts  to  explore  the  status  of  the   field  and  present  future  research  opportunities  involving  research  challenges  from   materials  to  computing,  including  materials  science  showstoppers  and  scientific   opportunities.  The  output  goal  of  the  roundtable  is  a  symbiotic  report  between  systems,   devices  and  materials  that  would  inform  future  ASCR/BES  research  directions.     Logistics   Gaithersburg,  MD,  Montgomery  Ballroom   Thursday,  October  29,  2015  (5:00pm  –  8:00pm)   Friday,  October  30,  2015  (8:00am  –  5:00pm)     Participants   Participation  and  observation,  by  invitation  only,  was  approximately  20  external  scientists   (DOE  laboratories,  university  and  industry).  Two  co-­‐‑chairs  helped  select  participants  and   helped  lead  the  discussion.  Several—approximately  10—Federal  Program  Managers  from   DOE  attended  as  observers.  The  total  meeting  size  was  approximately  30.     Agenda   The  agenda  comprised  two  days  and  included  a  keynote  address,  overview  talks,  discussion   sessions,  breakout  sessions,  and  a  closing  summary.     Roundtable  Report   A  draft  will  be  delivered  by  December  18,  2015.         Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     36   Roundtable  Agenda   Roundtable  on  Neuromorphic  Computing:   From  Materials  to  Systems  Architecture       Co-­‐‑chairs:     Ivan  K.  Schuller   University  of  California,  San  Diego   Rick  Stevens   Argonne  National  Laboratory  and  University  of  Chicago     DOE  Contacts:     Robinson  Pino   Advanced  Scientific  Computing  Research   Michael  Pechan   Basic  Energy  Sciences     Agenda:     Thursday,  October  29,  2015   5:00pm   Registration   6:00pm     Working  Dinner  with  Keynote  Speaker  Stanley  Williams   8:00pm   Adjourn     Friday,  October  30,  2015   8:00am   Continental  Breakfast  and  Registration   8:30am   Morning  Session:  Overview  talks   10:30am   Break   10:45am   Guided  Discussion  Session   12:00pm   Working  Lunch   1:00pm   Breakout  Sessions   3:00pm   Reports-­‐‑outs   4:00pm   Closing  Summary   5:00pm   Adjourn           Neuromorphic  Computing:  From  Materials  to  Systems  Architecture     37   Disclaimer   This   report   was   prepared   as   an   account   of   work   sponsored   by   an   agency   of   the   United   States  Government.  Neither  the  United  States  Government  nor  any  agency  thereof,  nor  any   of  their  employees,  makes  any  warranty,  express  or  implied,  or  assumes  any  legal  liability   or   responsibility   for   the   accuracy,   completeness,   or   usefulness   of   any   information,   apparatus,   product,   or   process   disclosed,   or   represents   that   its   use   would   not   infringe   privately  owned  rights.  Reference  herein  to  any  specific  commercial  product,  process,  or   service   by   trade   name,   trademark,   manufacturer,   or   otherwise,   does   not   necessarily   constitute   or   imply   its   endorsement,   recommendation,   or   favoring   by   the   United   States   Government  or  any  agency  thereof.  The  views  and  opinions  of  authors  expressed  herein  do   not  necessarily  state  or  reflect  those  of  the  United  States  Government  or  any  agency  thereof.                             This page intentionally left blank  01010011010101101001011010011001101001101 01010011001100101010011001100101010101010 10101010101011110100101001010011100100101 01001101010101011001100110011001010011001 01001010010100101010101001010101010101100 10100110011001100101001100101001100101001 10011001100111111011011110110111011101101 11011011010010010010011101010100101010101 0101010101010101010100101010111011011011 0001001101110110110110101001110011000101 0101010101110100100101010100111001101101 10110101011011011101110111011101110110101 01011011010001110010101011001100110010100 11001010011001010011001010010100101110011 00110010101011100010001001001000100000001 01111110010101010110101010101010101010010 10101111010010100101001110010010101001101 01010101100110011001100101001100101001010 01010010101010100101010101010110010100110 01100110010100110010100110010100110011001 10011111101101111011011101110110111011011 0100100100100111010101001010101010101010 10101010101010010101011101100101011110100 1010010100111001001010100110101010101100 1100110011001010011001010010111000000000 0100101001010101010010101010100010110010 1001100110011001010011001010011001010011