A Bristol upstart backed by Samsung and ARM daddy Hermann Hauser has raised $30m to make chips designed especially for machine learning.
Graphcore, led by CEO Nigel Toon, says it has "created a complete software and hardware system, based on our Intelligent Processing Unit (IPU), that will accelerate the full range of current and future machine intelligence methods for training, inference and prediction."
Companies working on machine learning and AI technologies use GPUs for their data-crunching. Graphcore, however, reckons its chip architecture is uniquely optimised for AI work.
"With standard chips, you end up using an awful lot of servers, so we can reduce [numbers of] machines, save a huge amount of power and make the system a lot more efficient," Toon told the Financial Times.
The Pink 'Un also mentions that Toon sold his previous company, Icera, which made wireless modems, to Nvidia for $370m in 2011 – without noting that Nvidia is shutting down Icera after failing to find a buyer for it, at a cost of up to $125m.
"One of the fundamental questions we have addressed at Graphcore is how to build highly parallel processors, complete with software tools and libraries, that are much faster, more flexible and easy to use, so that developers can explore machine intelligence approaches across a much broader front," a company press release says.
Graphcore closed its series A funding round with $30m promised to it from a variety of venture capitalists as well as Samsung’s Catalyst Fund. Its first product will come to the market next year, we are told, and Toon said that Bosch is also interested in its output. ®