Amplified training

Wireless chips can be frugal.

Whereas the 2022 version of a widely-used AI system consumed more than 1,200 megawatts during training. Far too much, according to a London-based startup founded out of a UCL research group.

Rather than relying on electric signals to communicate with AI chips, its team is using amplified laser light in the ultra-wideband range.

A network of thousands of graphic processing units could slash a comparable AI system’s energy consumption to 30 megawatts. Making it 40 times more efficient than the 2022 model.

Additionally, training could progress at 100 times the previous rate. A potentially influential development.

Particularly for AI systems focusing on climate change, like those analysing global shipping or sending signals during wildfires.