Neuromorphic Hardware

We categorize systems with the following criteria:

• Digital, Analogue or Mixed-mode. How are models for neurons or synapses implemented? How are spikes sent from one neuron to another?
• Scalability. Is it possible to interconnect instances of a platform to make a larger system?
• Neuron and synapse models. Which models have been implemented?
• Synaptic plasticity. Does the system support it?
• Precision. How many values can a weight have? What's the precision of the update rule in plasticity?
• Simulation time. Biological real time, faster than real time?
• Energy consumption.

---

The systems we reviewed are (in order of appearance):

1. HI-CANN:

• Mixed-mode (analogue neurons and digital communications)
• Scalable
• Configurable parameters of neuron models
• Fixed plasticity rule
• 16 different values for synapses
• Faster than real time
• 7.41 nJ/SE

2. HiAER-IFAT:

• Mixed-mode
• Scalable
• Configurable parameters of neuron models
• No plasticity
• Analogue neurons and synapses
• Real time
• 22 pJ/SE

3. Spinnaker:

• Digital (software neuron and synapse models, digital communications)
• Scalable
• Programmable neuron, synapse and axonal delays.
• Programmable plasticity rule
• 11- to 14-bit synapses
• Real time
• 8 nJ/SE

4. Neurogrid:

• Mixed-mode
• Scalable
• Configurable parameters of neuron models
• Fixed plasticity rule
• 13-bit shared synapses
• Real time
• 941 pJ/SE

5. TrueNorth:

• Digital (ASIC implemented neuron models, digital communications)
• Scalable
• Configurable parameters of neuron models
• No plasticity
• 122 bits for parameters and states. Synapses can have one of four possible signed integers and an ON-OFF switch (we consider this 4-bit precision)
• Real time
• 26 pJ/SE