Event-Driven QDI Circuits Nabil Imam 1 , Filipp Akopyan 2 , John - PowerPoint PPT Presentation

A Digital Neurosynaptic Core Using Event-Driven QDI Circuits Nabil Imam 1 , Filipp Akopyan 2 , John Arthur 2 , Paul Merolla 2 , Rajit Manohar 1 , Dharmendra Modha 2 1 Cornell University 2 IBM Research Almaden

Neuromorphic VLSI • Biological nerve cell models in silicon • Real-time, low-power operation - Commercial and medical applications - Tools for brain science 05/7/2012 – 2/16

Spiking Neural Networks • Neuron • Dendrites (inputs) • Axon (output) • Synapses (connections) TIME 05/7/2012 – 3/16

Models in Silicon Analog Circuits Differential Equations (e.g. Hodgkin-Huxley, Integrate-and-Fire, etc) • Computatation - Intrinsic properties - Network connectivity • Efficiency - ~100 9 neurons and ~100 12 synapses in the human brain - 20 W Power - 2 L Volume 05/7/2012 – 4/16

Implementation Challenges AER Packet Communication Computation Memory Off-chip Memory For Synaptic Information 05/7/2012 – 5/16

Crossbar Synapses • Integrated Computation and Memory • Crossbar – Rows: Axons – Columns: Dendrites – Junctions: Synapses • Flexibility • Efficiency: Large fanout in one operation 05/7/2012 – 6/16

Scalable Neurosynaptic Core • Crossbar Synapses • Digital Circuits – Deterministic – Scalable (45nm SOI) • Asynchronous design – Minimal power dissipation when idle – Fast when active – Robust • Scalable to a multi-core/multi-chip system 05/7/2012 – 7/16

System Architecture 04/11/2012 – 8/14

Neuron Model Leaky Integrate-And-Fire k  V ( t  1)  V ( t )  Leak  [ A j ( t )  W j  S j ] j  1 if V i ( t )   Spike  1 V i ( t  1)  0 ฀ 05/7/2012 – 9/16 ฀

Communication Infrastructure Scheduler for synchronization AER transmitter Routing fabric and receiver for for multi-core fan-out communication 05/7/2012 – 10/16

Chip Test 05/7/2012 – 11/16

Example Application: Sound Localization 05/7/2012 – 12/16

Example Application: Odor Recognition Neuron Index [C] 05/7/2012 – 13/16

Example Applications Motor Control Visual Recognition Pong! Auto-associative Memory 05/7/2012 – 14/16

Conclusion • Neurosynaptic Core – Digital Asynchronous Circuits – Crossbar synapses – Scalable to multiple cores and multiple chips • Applications – 1-1 hardware-software correspondence – Sensory-motor activity – Bio-inspired solutions to technological problems 05/7/2012 – 15/16

A Digital Neurosynaptic Core Using Event-Driven QDI Circuits Nabil Imam 1 , Filipp Akopyan 2 , John Arthur 2 , Paul Merolla 2 , Rajit Manohar 1 , Dharmendra Modha 2 1 Cornell University 2 IBM Research Almaden 05/7/2012 – 16/16