Semiconductors Carver Mead Gordon and Betty Moore Professor of Engineering and Applied Science, California Institute of Technology ACM ACM 97 THE NEXT 50 YEARS OF COMPUTING ACM ACM 97 THE NEXT 50 YEARS OF COMPUTING Copyright 1997 ACM, Association for Computing The files on this disk or server have been provided by ACM. Copyright and all rights therein are maintained by ACM. It is understood that all persons copying this information will adhere to the terms and constraints invoked by ACM’s copyright. These works may not be reposted without the explicit permission of ACM. Reuse and/or reposting for noncommercial classroom use is permitted. Questions regarding usage rights and permissions may be addressed to: permissions@acm.org ACM CARVER MEAD ACM Computation Church-Turing Thesis: Any computable function can be computed on a Turing machine. Mead's Thesis: The Class of Computable Functions is defined by Algorithms that Execute Efficiently on Commercial Machines. ACM Computation Complexity Theory: Time: The number of steps required to execute a program Space: The number of memory locations required by a program. Assumption: The computation done by different machines (per step or per unit hardware complexity) differs by at most a polynomial factor. ACM What if We Could Build a Computing Structure With a Computational Capability That Increased Exponentially with Its Size It Would Completely Change the Game! Candidate Structures: – Ultra-Parallel Digital VLSI Structures – Neural Computing Structures – Quantum Computing Structures ACM Ultra-Parallel Digital VLSI Structures The Good News: Remarkable Speedup Can De Achieved for Many Functions The Bad News: Speedup is No More than Polynomial ACM What is Going On ? Digital Systems: Represent information by a finite set of Discrete Symbols Digital Representation Permits Information to be – Transmitted through Space – Stored through Time without Loss Signal Representing the Symbol is Restored to the Nearest Symbol by a Contractive Mapping Precision of the Representation is exponential in the Number of Symbols per Value ACM ACM ACM Digital System Limitations Have No Natural Representation for Time All Continuous Variables Must Be Represented by Finite Strings of Discrete Symbols Process Information in Discrete Chunks Have No Notion of Locality or Continuity ACM Digital System Limitations When Used to Simulate Continuous Non-Linear Systems No General Criterion is Known for Numerical Stability Most Computations Dominated by Aliasing Artifacts Alternative Hypotheses Have Only Discrete Representation Exponential Alternatives Require Exponential Resources Quantizes After Every (Very Simple) Operation ACM ACM Neural Computation Signals Transmitted over Distance: Represented as Events – Discrete in Amplitude – Discrete in Time – Continuous Arrival Time Variable Local Signals: Electrical Potential Continuous in Amplitude Chemical Concentration Continuous in Time Information is encoded in Spatio-Temporal Structure of Signals }-{ ACM ACM Neural Computation Signals are Decoded in Highly Branched Dendrite Structure – Arrival Time Aligned by Propagation Delay – Temporal Integrity Maintained – Continuous Amplitude – Active Amplification – NonLinear Interaction – Adaptive Control keeps Structure Stable – Positive and Negative Feedback Keeps Many Combinatorial Possibilities Active in Same Structure Avoids Pre-Quantization ACM ACM Neuromorphic VLSI Systems Silicon and Wetware Have Similar Physics Basic Continuous Representation – Continuous Time – Limited Precision – Many Operations Come Free From Physics Interconnection Limits Complexity – Energy is Precious – Stability is a major Concern Natural World as Source of Information – Real-Time Systems – Adaptation to Environment – Learning vs programming – Time Evolution as Source of Learning ACM Many Functional Sub-Systems Have Been Built Vision Systems – – – Auditory Systems – – – Retinas Motion Perception Stereo Matching Cochleas Auditory Feature Extraction Stereo Localization In Situ Learning – Floating Silicon Gates – Autonomous On-Chip Operation – Weight Modification when In Use Has Not Yet All Come Together ACM Quantum Computation Information Encoded in Spatio- Temporal Structure of Many-Body Wave Function Phase Space for Coupled Many-Body System is Cartesian Product of Individual Phase Spaces – Greatly Enlarged Dimensionality Time Evolution of Coupled Many-Body System Explores Volume in Phase Space that is Exponential in the Number of Dimensions – All in Parallel Physical Size of System is Linear in the Number of Dimensions ACM Quantum Computation Theory: New Model of Computation Experiment: – Many Delightful Model Systems – Working Understanding of Quantum Mechanics Problems: Unwanted Coupling to rest of Universe – DePhasing of the Wave Function – Limits Computational Possibilities ACM ACM