Going neuro down in data centros

The idea of processors acting like brains is as old as computers, but when it comes to energy efficiency they have a long way to go

- Dick Pountain

I've written many sceptical words about AI in this column, railing against overconfidence and hype, hubristic pursuit of AGI, deepfakery and content pillage. Nevertheless I do believe that AI - once we've civilised it - is going to be hugely important to science, economics, robotics, control systems, transport and everyday life. This assumes that, through political will, public concern about misinformation, invasion of privacy and theft of artistic data can be regulated away. Even then, one colossal stumbling block will remain: energy consumption.

When AI firms consider purchasing mothballed nuclear reactors to power their compute-servers, the absurdity of AI's current direction ought to be visible to everyone.

Current GPT-based AI systems depend on supercomputers that can execute quintillions of simple tensor arithmetic operations per second to compare multiple layers of vast matrices holding encoded parameters.

Currently all this grunt is supplied using the same CMOS semiconductor process technologies that gave us the PC, the smartphone and computer games - the Nvidia chips that drive most AI servers are descendants of those originally developed for 3D games. The latest state-of-the-art GPUs have a watts/cm² power density around the same as an electric hob, and the power consumption of AI server farms scales exponentially to the square of the number employed (order O(N²) in complexity theory jargon).