A transistor inspired by human synapses

搜索最新SoC解决方案
关于汽车电子...物联网...安全...音频...视频

CEA-Leti has developed a transistor inspired by human synapses – an innovation that is quite possibly the closest we've come to imitating how the brain's neurons connect. This transistor boasts many similarities with human synapses: how it operates, its ultra-low energy consumption and a similar level of miniaturization. It opens the door to a future with more powerful circuits that will meet the needs of artificial intelligence.

www.leti-cea.com, Mar. 01, 2023 – 

The transistor presented by CEA-Leti is analog. In other words, instead of having two possible states, open & closed, it has around 50 possible states. This enables it to reproduce how synapses function. Synapses act as the connection between brain neurons, which are activated when their synapses have recorded a certain number of electrical impulses. A transistor with several dozen states can imitate this behavior.

Energy consumption to equal a synapse

Another similarity between this new transistor and a synapse is that both are ionic. The transistor takes advantage of the same electrochemical reaction found in a synapse. In the transistor's case, its channel is made up of titanium oxide through which lithium ions flow. According to their number, they modulate the channel's electronic conductivity.

This bio-inspired function enables the transistor to set new records in terms of energy efficiency: the transistor consumes 1 femtoJoul/square micron, the same amount as a synapse. Its energy consumption is up to 100 times less than other components such as resistive memory, which are currently used for the same applications

200 mm wafers & CMOS compatibility

This new transistor layer is ultra-thin (200 nm thick) and durable (more than 100,000 cycles). When integrated in a neuromorphic circuit, it achieved excellent results during an MNIST image recognition benchmark test. CEA-Leti designed the transistor with mass-production in mind: it is manufactured on 200 mm wafers with CMOS compatible processes.

"All of these elements are encouraging, but we're only at the first stages of the evaluation process. We have to continue bringing the transistor to maturity and ensure an in-depth evaluation of its durability and reliability," explains Sami Oukassi from CEA-Leti.

Targeting deep neuronal networks

Once the transistor is fully evaluated and confirmed, its low energy consumption would enable it to be integrated into neuromorphic circuits. These circuits are dedicated to image and voice recognition. During learning phases, these circuits are used intensively and any savings in terms of energy consumption are particularly valuable at this moment. CEA-Leti has filled three patent requests to protect this promising technology.

click here to read more...

 Back