Artificial intelligence is rapidly gaining traction, and R&D professionals from science and industry alike are innovating solutions to a major challenge: the need for chips capable of completing a mind-boggling number of computations without sending energy consumption through the roof.

One strategy that has been garnering increasing attention is to integrate memory and processing as seamlessly as possible. Leti, a CEA Tech institute, recently built a new and very promising chip that does just that.

Leti researchers found inspiration in biological neural networks. They utilized analog pulse-coded neurons, which are more energy efficient than conventional neurons due to the fact that the operations they use are simpler. For the first time ever, they combined pulse-coded neurons with OxRAM-type resistive non-volatile memory (the synapses) on the same circuit. A large number of synapses were integrated very close to the chip's processing capabilities to keep the movement of data to a strict minimum.

The circuit, which they named SPIRIT, uses around five times less energy than an equivalent circuit with conventional formal coding. A proof-of-concept prototype was presented at the IDEM trade show in December 2019. The chip was developed to complete a simple task, the recognition and classification of hand-written numbers. But this is only the beginning: It will now be possible to develop specific chips for much more complicated tasks.