A 0.086-mm ² 12.7-pJ/SOP 64k-Synapse 256-Neuron Online-Learning Digital Spiking Neuromorphic Processor in 28-nm CMOS.

Area of Science:

• Neuromorphic Engineering
• Computer Architecture
• Artificial Intelligence

Background:

• Traditional von Neumann architectures face limitations in low-power sensory data processing.
• Spiking neural networks (SNNs) offer a promising alternative for event-based, energy-efficient computation.
• Online learning is crucial for SNNs to adapt in real-world, uncontrolled environments but faces complexity challenges.

Purpose of the Study:

• To develop a compact, low-power digital neuromorphic processor enabling online learning for SNNs.
• To address the complexity and area overheads associated with embedding online learning capabilities in SNNs.
• To present ODIN, a versatile platform for cognitive neuromorphic device research.

Main Methods:

• Design and fabrication of ODIN, a 0.086 mm² 64k-synapse 256-neuron processor using 28-nm FDSOI CMOS technology.
• Efficient implementation of the spike-driven synaptic plasticity (SDSP) learning rule for high-density embedded online learning.
• Configurable neuron models supporting both leaky integrate-and-fire and Izhikevich behaviors.

Main Results:

• Achieved a minimum energy per synaptic operation (SOP) of 12.7 pJ with a synapse area of 0.68 μm².
• Demonstrated 84.5% classification accuracy on MNIST using a single-layer network with on-chip SDSP learning.
• Attained an energy consumption of 15 nJ/inference at 0.55 V using rank order coding.

Conclusions:

• ODIN provides an efficient hardware platform for online learning in SNNs, overcoming previous complexity barriers.
• The processor enables adaptive, low-power, and cost-effective processing for cognitive neuromorphic applications.
• This work facilitates advancements in developing intelligent, adaptable systems for sensory data processing.