Bristol chip designers turn to artificial intelligence
GraphCore in Bristol has been spun out of chip designer XMOS by former chief executive Nigel Toon and former chief technology officer Simon Knowles to develop an entirely new device called an Intelligent Processing Unit (IPU). Toon remains chairman of XMOS after stepping down as CEO.
The IPU is a completely new type of processor that has been designed to accelerate the next generation of intelligent machine learning by working alongside a traditional computer processor. Many products and services currently use machine learning in the cloud and the IPU technology allows more intelligent products to be created and delivered quickly.
The two founders of Graphcore were part of the original team at Icera in Bristol before it was bought by US chip giant NVIDIA. The other co-founders of Icera – Stan Boland and Steve Allpress – have also launched a startup in Bristol called FiveAI developing artificial intelligence software for autonomous systems such as driverless cars.
GraphCore’s IPU design has been optimised to work efficiently on the extremely complex high-dimensional models that machine learning and deep learning require. It emphasises massively parallel, low-precision floating-point compute and provides much higher compute density than other solutions.
Chip acts like a human brain
Like a human brain, the IPU holds the complete machine learning model inside the processor and so has a hundred times the memory capability of traditional devices. This gives lower power consumption and much higher performance.
At its labs in Bristol and Silicon Valley, Graphcore is developing an ‘IPU-Accelerator’ that can be easily plugged into a server to accelerate machine-learning applications which it says will be ready soon. A stand-alone ‘ IPU-Appliance’ would provide much higher performance to support many more users or to provide much faster training.
To help develop machine learning applications using the IPU, Graphcore has built a complete set of software tools and libraries that integrate seamlessly with the leading machine learning frameworks.
You can find out more at the Graphcore website.