SNS silicon: a first-of-its-kind multi-bit ReRAM spiking neuron

The first chip designed and fabricated within SNS is a single spiking neuron in which the synaptic weights are stored directly as multi-bit ReRAM conductances — to our knowledge the first integration of its kind of multi-bit memristive synapses with a spiking neuron. The neuron combines an analog SpikingGamma synapse, whose temporal kernels map naturally onto the dynamics of a differential-pair-integrator (DPI) circuit, with a novel sigma–delta neuron that reproduces the full SpikingGamma dynamics that conventional analog spiking neurons cannot provide. The design also incorporates multiple such synapses and the supporting circuitry to form, set and reset the ReRAM devices.

Beyond characterising the technology and the circuit behaviour, the chip lays the groundwork for self-healing directly in hardware. A measurement block in the forming circuit can verify whether a device has reached its target state, and the forming circuitry can be reused to attempt reformation of devices affected by stuck-at-0 faults. Looking ahead, this enables local-learning mechanisms that let the network compensate for non-idealities in the ReRAM itself. Initial simulations indicate the approach faithfully reproduces the SpikingGamma synaptic behaviour, and the fabricated silicon is now being brought up and characterised.