ROLLS - INI

The ROLLS chip implements key requirements for online learning and adaptive behavior in neuromorphic systems, with highly flexible digital configuration options.

Overview

• Presented as a “Reconfigurable On-line Learning Spiking Neuromorphic Processor”
• 256 silicon neurons, 128K plastic synapses, 256K programmable synapses
• Mixed-signal analog/digital implementation
• 180nm CMOS process, 51.4 mm2 die area
• Modeling of neural systems that adapt in real-time through on-chip learning

Silicon Neurons

• Implement adaptive exponential Integrate-and-Fire (I&F) model
• Exhibit various spiking behaviors (frequency adaptation, refractory period, threshold adaptation, bursting)
• Highly configurable through 13 tunable bias parameters
• Low mismatch between neurons (9.4% avg)
• On-chip learning circuits to evaluate spike-timing-based plasticity rules

Synapses

• Two core arrays:
  • Long-term plasticity (LTP) synapses with bi-stable weights using drift-diffusion mechanism
  • Short-term plasticity (STP) synapses with 4-level programmable weights and short-term depression
• LTP array models NMDA receptors with longer time constants
• STP array models AMPA/GABA receptors with shorter time constants
• Sparse digital logic for memory and configuration in every synapse
• Digital configuration of synapse types and connectivity

Architecture

• Virtual “diffuser” synapses to combine signals from groups of synapses
• Synapse demultiplexing to allow flexible neuron-synapse mapping
• Asynchronous peripheral digital circuits for input/output
• Bias generator for tuning analog parameters
• ADC readout circuit for monitoring internal signals

Applications

• Hardware attractor networks to model working memory and decision making
• Real-time image classification demonstration with spiking vision sensor

Related publications