Transcript Presentation - UCSD VLSI CAD Laboratory

Embrace the BRAIN Century:
EDA Challenges in Neuromorphic
Computing
Hai Li and Yiran Chen
Evolutionary Intelligence Lab (EI-Lab)
Electrical and Computer Engineering
University of Pittsburgh
1
What is Neuromorphic Computing?
• An interdisciplinary technology that was inspired from biology,
physics, mathematics, computer science, and electronic
engineering to design artificial neural systems. (Wikipedia)
• It is supposed to fulfill the weakness of
von Neumann architecture in processing
cognitive applications.
• The relevant research has been well
funded by all major funding agencies:
• And supported in many countries:
2
Question I – Understanding?
• Unfortunately we still do not know much about human brains.
• The artificial neural network models also evolves over years.
– Representation of neuron: 1943, McCulloch (Pitt)
– The 1st learning rule: 1949, Hebb
– Neuron nets: 1955, Dartmouth Summer Research Project on AI
– STDP (Spike-timing-dependent plasticity): 1973, Taylor
– CNN (Convolutional neural networks): 1989, LeCun
• Do we really need to understand brains before designing
a useful N.C. system?
– No. Many useful systems have been prototyped, e.g., IBM TrueNorth.
– The debates on “Emulative vs. Simulative”.
3
Question II – Platform?
AD
AD
ASIC
SC
PE
PR
PO
FPGA
SC
PE
PR
PO
Application Specific IC
Programmable Hardware
Misra et al, Neurocomputing, 2010
Graf et al, NIPS, 2009
Misra et al, Neurocomputing, 2010
CPU
AD
Adaptivity
(AD)
Performance
(PE)
Power Efficiency
(PO)
Programmability
(PR)
Scalability
(SC)
AD
NCA
GPGPU
SC
PE
PR
PO
General Purpose Platform
P. J. Fox, Tech. Report, 2013
Graf et al, NIPS, 2009
SC
PE
PR
PO
Memristor Based
Reconfigurable Design
H. Li, HPEC, 2010 4, DAC, 2015
4
Question III – Technologies?
• Are conventional CMOS and EDA technologies capable to
support long-term research and development of N.C.
systems?
– Debates
• Analog or Digital?
Qualcomm, Zeroth
Custom hybrid
Spike neurons on chip
Synapse off chip
• Spiking-based or level-based?
• Synchronous or asynchronous?
• CMOS or Post-Silicon?
– Other Challenges
• Programmability
J. Hsu, IEEE Spectrum, 2014
J. Gehlhaar, ASPLOS, 2014
HP, memristor X-bar
Analog computing
Dense connection
• Security
Stanford, Brain in Silicon
Mixed-signal VLSI
1M neurons/16 chips
1B synapse/16 chips
B. Benjamin, Neurogrid, 2014
D.B. Strukov, Nature, 2014
• Reliability
• Scalability
IBM, TrueNorth
SRAM synapse
Digital spike
1M neurons/chip
256M synapse/chip
HBP
Analog VLSI
64 neurons/chip
1024 synapses/chip
S. Miller, ESANN, 2012
Micron, Automata
Massively parallel
Memory driven
Non-von Neumann
XML-based language
F. Samarrai, UVAToday, 2014
5
Acknowledgement
• Dr. Daniel Hammerstrom, Program manager, DARPA
• Dr. Robinson Pino, Program manager, DOE
• Dr. Dharmendra S. Modha, IBM Fellow and IBM Chief Scientist
for Brain-inspired Computers
• Dr. Mark Barnell, Senior computer scientist and program
manager, US AFRL
• Dr. H.-S. Philip Wong, Willard R. and Inez Kerr Bell Professor,
Stanford University
Q & A?
6