One of Bristol’s leading AI startups plans to ship its first chip by the end of the year.

Graphcore has taken a different approach to designing a chip to handle AI and machine learning algorithms, says Simon Knowles, chief technology officer (above). Knowles was previously responsible for chip designs at Bristol startups Icera Semiconductor (which became part of NVIDIA) and Element14 which became part of Broadcom.

The Colossus chip is designed for both training and deployment (also called the inference engine). It is being built on a 16nm semiconductor process with over 1000 totally independent processors linked by a non-blocking interconnect (the latest server processors are built on 14nm).

It uses mixed precision floating point operations, with 16bit multipliers and 32bit accumulators. It can be programmed in the popular TensorFlow machine learning language or with GraphCore’s Poplar development tool for ‘bare metal’ development.

“Inventing a new processor architecture and rendering it as a leading-edge chip is difficult and expensive,” said Knowles, speaking at the recent 3rd Research and Applied AI Summit (RAAIS) in London. “Building all the software on top to make it useful and easy to use is similarly difficult and even more expensive, so unless it’s going to be useful for 20 years or more, don’t bother.”

New kind of chip

One of the ways to do this effectively is to keep all the processing on the chip with lots of memory that can handle both the training of machine networks and the inference.

“We should not need two types of machine architecture”


“Any intelligence computer requires just two components: an inference engine and a knowledge model,” he said. “Learning is a form of inference, in which we are trying to condense experience (data) into the knowledge model. We do this by inferring the parameters and perhaps the structure of the model from the training data.

“We can then use the learned model to infer some unknowns from some new input data, such as what motor actions to take in response to incoming video frames. We should not need two types of machine architecture. There might be large machines and small machines, but they can have the same architecture and run the same software efficiently.”

Chips and appliances will be available for early adopter customers by the end of the year, with more general availability next year. The company has raised over $30m in funding so far.

You can see images of how GraphCore’s Poplar software handles neural networks here