You are here : Home > INDUSTRIAL INNOVATION > DEMOS > Spiking neural networks enabling massively parallel, low-power & low-latency computation

Articles & files | Focus | Article



Published on 26 January 2021

Spiking neural networks enabling massively parallel, low-power & low-latency computation


   What is Spirit?

  • CEA-Leti introduces SPIRIT, the world-first fully integrated neural network on-chip with non-volatile resistive memory. So far, memories were placed outside of chips leading to high energy consumption. With this co- integration in the same die of analog spiking neurons and resistive synapses leveraging resistive random access memory cells (RRAM), CEA-Leti enables the push for distributed computing devices to support artificial intelligence at the edge. These spiking neural networks are designed by CEA-Leti and the RRAM are fabricated in a post-processat CEA-Leti on CMOS-based wafers.


  • SPIRIT allows massively parallel, low-power and lowlatency computation. It is a perfect candidate for embedded classification applications required in Lidars or IoT devices. The technology also finds applications in event-based sensors packed with rich temporal content.

   What's new?

  • Spiking neural networks are considered as the third generation of neuralnetworks. SPIRIT enables non-volatile, high density and computational memories on the chip. This has been made possible thanks to the Back-End- Of-Line (BEOL), CMOS compatible, resistive memories developed and fabricated at CEA-Leti. The demonstration features a neural network—perceptron—trained offline for handwritten digits classification. The circuit operates live classification of digits drawn on a touch screen. Spike coding reduces activity translating into less power dissipation, with less than one spike generated per synaptic connection.

  What's next?

  • CEA-Leti keeps working on improving the resistive RAM technology, especially the integration density. Work is ongoing on the development of a BEOL selector to enable mega-bit scale crossbars, which are the ultimate structure density-wise. CEA-Leti is also pursuing the implementation of mutiple-level cells, i.e. several bits per cell. This demonstration represents the first milestone of CEA-Leti’s roadmap for large spiking neural network accelerators leveraging resistive memories. The next step will consist in interfacing accelerators with event-based vision sensors and LIDARs to enable pose estimation, simultaneous localization and mapping (SLAM), and object classificationA prototype will be available in 2021.


 3.6pJ per synaptic event (at neuron-level)

 High RRAM synapse robustness to spiking events (1 billion spikes reached without any impact)

 3 patents pending

  •       FLYER


© Leti – Crédits photos :
La Chouette Compagnie