June 18, 2021 | Artificial intelligence and energy frugality are not incompatible

18 juin 2021 | Quand intelligence artificielle rime avec frugalité énergétique
© AdobeStock_willyam
CEA-List recently developed an image and word recognition solution that is both powerful and energy efficient. This advance marks a step toward integrating artificial intelligence into IoT devices.

​When it comes to the kinds of complex capabilities artificial intelligence enables, high power consumption is no longer a given. CEA-List and fellow CEA Tech institute CEA-Leti combined an ultra-low-power processor with the PNeuro® neural accelerator in a demonstrator system that can support application scenarios representative of today’s IoT devices. The demonstrator, which goes by the name of SamurAI, was able to recognize images and keywords in natural language.

The architecture that underpins SamurAI hinges on a very-low-power fast-wakeup function and a very energy-efficient “on-demand” processing unit that is activated only when complex processing is needed. SamurAI also leverages the PNeuro® neural accelerator, which allows word and image recognition applications to be rolled out very rapidly using N2D2, a code generator developed by CEA-List Carnot institute.

For keyword spotting in natural language, SamurAI is 300 times faster and more energy efficient than conventional processors.

This level of performance and energy efficiency make SamurAI a solution of choice for AI for IoT. With SamurAI, near-sensor processing is possible, for even lower-power and lower-latency that also protects data privacy.


Read article at