Two-tier AI system-on-a-chip can run on tiny solar cell

While AI in its wildly varying forms is becoming ubiquitous, AI-based technology comes with the downsides of generally requiring a lot of power to run and needing to be permanently connected to the cloud, raising data protection, energy efficiency and security concerns.

The team of CSEM engineers have taken a step towards resolving these issues through the development of a system-on-a-chip which consumes relatively low levels of power and executes AI operations at the edge (on the chip rather than in the cloud).

The system uses an entirely new signal processing architecture which minimises power consumption. It consists of an ASIC chip with a RISC-V processor (another CSEM innovation), and two tightly coupled machine-learning accelerators: one for facial detection, for instance, and the other for classification. The first is a binary decision tree engine for performing simple tasks but not recognition operations.

“When our system is used in facial-recognition applications, for example, the first accelerator will answer preliminary questions like: 'Are there people in the images? And if so, are their faces visible?'” said Stéphane Emery, CSEM’s head of system-on-chip research. “If our system is used in voice recognition, the first accelerator will determine whether noise is present and if that noise corresponds to human voices. But it can't make out specific voices or words; that's where the second accelerator comes in.”

The second machine-learning accelerator is a convolutional neural network engine for performing more complex tasks, such as recognising individual faces or specific words. This accelerator – which consumes much more energy – only runs when necessary. This two-tiered data processing approach reduces the average power consumption of the system, enabling them to run from a tiny battery or solar cell.

As part of their study, the engineers next enhanced the performance of the accelerators themselves. They rendered the system highly versatile; it is modular, allowing it to be tailored to any application for which real-time signal and image processing is required.

“Our system works in basically the same way regardless of the application,” said Emery. “We just have to reconfigure the various layers of our CNN engine.”

The new system-on-a-chip architecture could open the door to a wholly new generation of devices with processors which run independently for over a year. It could also slash the installation and maintenance costs for this class of devices and enable them to be used in locations which would make it impractical to regularly change the battery.