Just when massive machine learning models like GPT-3 and BERT had us all feeling like we needed a nuclear reaction to generate the 1.21 gigawatts of electricity required for processing, the crew over at the Swiss research and technology organization CSEM designed an

 AI system-on-chip (SoC) that runs on a much tighter energy budget. In fact, this SoC can run on either a tiny coin cell battery or even solar power.

The Sunny Future of AI

This may sound like a gimmick, but the device is actually able to run real-time signal and image processing algorithms — nothing is sent to the cloud for processing. Being fully modular, the SoC can be specifically tailored to different applications.

Just when massive machine learning models like GPT-3 and BERT had us all feeling like we needed a nuclear reaction to generate the 1.21 gigawatts of electricity required for processing, the crew over at the Swiss research and technology organization CSEM designed an

AI system-on-chip (SoC) that runs on a much tighter energy budget. In fact, this SoC can run on either a tiny coin cell battery or even solar power.

This may sound like a gimmick, but the device is actually able to run real-time signal and image processing algorithms — nothing is sent to the cloud for processing. Being fully modular, the SoC can be specifically tailored to different applications.

Accelerator chip (📷: CSEM)

The system consists of an ASIC chip with a RISC-V processor and a pair of machine learning accelerators. CSEM demonstrated a device built for facial recognition. In this case, the first accelerator is a binary decision tree (BDT) that can only carry out the relatively simple operation of detecting whether or not a face is present in an image. The BDT cannot provide any more detailed information (such as who the detected face belongs to), but it can operate with very little power.

When a face has been detected, then the second accelerator comes into play. This one runs a convolutional neural network that can perform more complicated tasks — like facial recognition. This accelerator does consume more power, but by using a two-tiered approach that only triggers this module when really necessary, the system’s overall power requirement is drastically reduced. The demonstrated device also made use of an e-ink display for output.

The Sunny Future of AI

It is possible to reconfigure the accelerators to handle other machine learning tasks that slowly sip energy. For example, a voice recognition system could be built by reconfiguring the first accelerator to detect the presence of sound. The layers of the neural network in the second accelerator can likewise be adjusted to recognize specific words.

Devices powered by this innovation can run independently for over a year. That has the potential to reduce installation and maintenance costs, as well as allow for installations in locations that would previously have been impractical