AI has propelled chip architecture towards a tighter bond with software

In the 2000s, software companies like Google, Microsoft and Meta were content with the incremental processing gains that chipmakers delivered every few years. But in the early 2010s, Google realised that enthusiasm for artificial intelligence (ai) applications based on machine learning could overwhelm it. If everyone with an android phone were to use its voice-control feature for three minutes a day, one executive calculated, the firm would need double its data-centre capacity. By 2015 newer machine-learning algorithms were demanding 100 times more processing power than previous versions. Because the chipmaking industry did not have any idea that would make “everything better” quickly enough, says David Patterson of the University of California Berkeley, who is also an adviser to Google, there was little alternative but specialisation, working on chips that did only a few things, but did them very much better.