I’ve been speculating that AI (artificial intelligence, but don’t mistake it for self-agency) would suck down enough power to make them too pricey to use for piddling tasks, but now comes this, from NewScientist (21 October 2023, short version, paywall):
A reconfigurable transistor can run AI processes using 100 times less electricity than the standard transistors found in silicon-based chips. It could help spur development of a new generation of smartwatches or other wearable devices capable of using powerful AI technology – something that is impractical today because many AI algorithms would rapidly drain the batteries of wearables built with ordinary transistors.
The new transistors are made of molybdenum disulphide and carbon nanotubes. They can be continuously reconfigured by electric fields to almost instantaneously handle multiple steps in AI-driven processes. In contrast, silicon-based transistors – which act as tiny on-or-off electronic switches – can only perform one step at a time. This means an AI task that might otherwise require 100 silicon-based transistors could instead use just one reconfigurable transistor, thereby reducing energy consumption.
Performance is not mentioned, unfortunately, and I’m not quite sure what it means for something to be continuously reconfigured – the frustrations of being obsolete, eh?
But it remains fascinating.
