May 20

Réunion Comics (& Nanoarchi) 20/11/2014

Réunion Comics (& Nanoarchi) 20/11/2014

Présents : Michele Amato, Arnaud Bournel, Laurie Calvet, Deng Erya, Philippe Dollfus, Joseph Friedman, Nicolas Locatelli, Nguyen Mai Chung, Nguyen Viet Hung, Damien Querlioz, Jérôme Saint Martin, Su Li, Tran Van Truong, Adrien Vincent, Damir Vodenicarevic, xx.

Adrien, MTJ as stochastic synapses, a good idea ? How to use them smartly ?

First talk with Sozi-made slides : Presentation_COMICS_2014-11-20_full_chemins

Computer (till 200 kW, 3,9 m²) vs. brain (20 W, 1,3 l).

Spike Timing-Dependent Plasticity : learning rule to be implemented in a MTJ. Thanks to their tunable stochasticity (switching probability vs. current pulse).

Physical macrospin model → Monte Carlo simulation of a stochastic LLG eq (programmed in Python), from which an analytical model is derived.

System level simulation… Cross bar architecture for A Si retina. 600 000 neurons to simulate. 1 bit per synapse. C++ program, time step based simulator.

Choice of the metric… Best neuron on each lane, instead of average on the selected neurons. For focusing on the device performances without aggregating.

Result. Average recognition rate as a function of the synaptic variability and for different current regimes. Low current regime not interested, the other are more resilient to variability. Very small advantage to high current regime.

Choice of the threshold for optimizing the performance.

Programming power consumption. Depends heavily on the chosen current regime. If low current regime is used, longer pulses are necessary → 420 µW @ 0.30 V vs. 220 nW @ 0.40 V for the intermediate regime and 200 nW @0.60 V for the high current regime. Disappointing…

Optimizing power consumption thanks to the analytical model. Last (preliminary) result obtained yesterday 😉

Perspectives. Both at the circuit level (cf. Nicolas VerilogA model + …) and at the system level (several ideas with Damien).

Leave a Reply