Teal overlay

AI computational power leaves Moore’s Law in the dust

Moore’s Law has held up pretty well over the past few decades.

The observational ‘law’ was based on the rate at which the number of microcomponents in a microchip or integrated circuit increased.

Essentially, this predicted that available computational power or processing speeds would double every 18 months.

The rate held more or less steady from 1975 until 2012, but now the advent of artificial intelligence (AI) has seen an increasingly rapid acceleration in processing power.

According to Stanford University’s 2019 AI Index, the speed at which computational power is doubling has increased massively.

The report, which was produced in partnership with the likes of Google, McKinsey & Company, PwC, Genpact, OpenAI and AI21Labs, said that before 2012, AI results tracked Moore’s Law closely, with compute doubling every two years, but after 2012, compute has doubled every 3.4 months.

Study looked at AI’s ability to decipher images and actions

The report also examined how AI algorithms were improving in a number of different areas.

One is image recognition based on supervised machine learning, which is when the algorithm is trained on a labelled dataset with both input and output parameters.

The study author said that this area gave a good indication as to the maturity of AI infrastructure in general, including advances in both hardware and software.

It found that the time required to train a cloud infrastructure network to do this fell from three hours in October 2017 to just 88 seconds in July this year.

Associated costs had also fallen for completing the same or similar tasks.

The study found that it would have cost $2,323 to train the network to an accuracy level of 93% in 2017.

By September 2018, it would have cost just over $12 to achieve the same result.

The study also looked at AI’s ability to recognise human activities from watching videos.

It said that today’s sophisticated algorithms were able to identify hundreds of different and often complex human behaviours in real time.

There was still plenty of work to be done, however.

The study noted: ‘After organising the International Activity Recognition Challenge (ActivityNet) for the last four years, we observe that more research is needed to develop methods that can reliably discriminate activities, which involve fine-grained motions and/or subtle patterns in motion cues, objects and human-object interactions.’

Today’s news was brought to you by TD SYNNEX – the UK’s number one distributor.

Back to Top