News

March 3, 2022

A supercomputer more powerful than the human brain may be in reach

Graphcore plans to get there in two years


Maija Palmer

4 min read

Building a supercomputer as powerful as the human brain has long been a stretch goal for the tech industry.

But now we may be getting closer to it. Graphcore, the UK AI-chip designer and one of the country's most valuable private companies, today announced plans for an ultra-intelligent AI computer that could surpass the parameters of the human brain.

Despite the promise of supercomputers, they still aren’t as fast or as powerful as the grey matter in your skull. The human brain is able to handle more than 100tn parameters — or pieces of data — which is a level of computing power that hasn’t been matched by any silicon computer.

Advertisement
This is not a one-off. It is a product we intend to deliver in volume because it is what our customers have been asking for

Graphcore’s computer would be ready to ship in two years and cost just $120m, the kind of price that makes it accessible for individual companies and universities. For context, $120m is a fraction of the $1bn it cost to build the world’s fastest supercomputer, Japan’s Fugaku.

“This is not a one-off. It is a product we intend to deliver in volume because it is what our customers have been asking for,” says Simon Knowles, cofounder and CTO at Graphcore.

The computer uses a new type of silicon processor that Graphcore has developed, perhaps another stroke in Graphcore’s favour as it aims to overtake rival Nvidia. US-based Nvidia dominates the market for AI chips, but challengers like Graphcore, Cerebras and SambaNova have emerged with chips specifically designed for AI calculations.

Graphcore's Bow 2000 machine
Graphcore's Bow 2000 machine

Five times bigger and 10 times faster than the human brain?

The new computer that Graphcore is proposing — based on a recent breakthrough in its AI chip technology — would be able to support AIs with 500tn parameters. It will also be able to calculate at more than 10 exaflops — or 10 billion billion — calculations per second. It has been estimated that the human brain calculates at 1 exaflop.

Graphcore is calling this the Good Computer, named after Jack Good, an early computer pioneer who worked on developing the world’s first electronic computer, Colossus, at Bletchley Park, the centre of the UK’s codebreaking operations during the Second World War. In the 1960s, he was Stanley Kubrick’s science adviser for 2001: A Space Odyssey.

Whether or not Graphcore’s Good Computer will really surpass the human brain in a meaningful sense isn’t certain. The brain is still hugely more efficient than any machine, with cells that are able to simultaneously perceive, interpret, store, analyse and distribute information with an elegance that machines have yet to come close to emulating. And the brain has features such as plasticity, which allows it to morph its physical structure if it finds a more efficient way to compute something. Plus the brain does all this with a tiny amount of space and power.

The new computer that Graphcore is proposing would be able to support AIs with 500tn parameters — five times the human brain

Other projects like Manchester University’s SpiNNaker supercomputer have been trying to learn how to emulate the human brain, not just in terms of size but in the way it can communicate large amounts of information very efficiently.

But the advances demonstrated by Graphcore show that some of the computer power constraints that have kept artificial intelligence at a relatively rudimentary level so far are starting to be overcome.

The Good Computer is based on a new type of silicon processor that Graphcore has unveiled, a 3D chip in which two pieces of silicon wafer are joined together. Graphcore says this gives the processor 40% higher performance and 16% more power efficiency than the company’s previous models.