I've just had an interesting conversation with Nigel Toon, CEO of Graphcore, a (relatively) new hardware startup that aims to shake up the world of hardware accelerated machine learning. Based in Bristol, England, the company's engineering team is led by Simon Knowles and based on a group of experienced engineers who have produced a succession of succesful custom processors at STMicroelectronics, Element 14 and Icera. That's an impressive track record so I am inclined to take their efforts very seriously.
Graphcore's plan is to reverse the current trend of shoehorning neural networks onto available computational hardware, typically a GPU or DSP, by creating an architecture specifically to handle them more efficiently than either of those architectures can. Since neural networks (and AI problems in general) are best expressed by graphs where compute is represented by vertices and connections by edges, then the most natural and scalable way to accelerate that class of problems is to closely replicate the graph structure in the accelerator.
While this approach is not unique -there are other companies such as Wave Computing that are claiming a similar approach-Toon claims that none of the competitors he is aware of have started with the neural network as the problem statement and built a computer from scratch to solve it.
Research into the new architecture was started in 2012, soon after Icera was bought by Nvidia, and although Toon is not making any announcements, the proposed introduction date of late 2017 for PCIx cards bearing the accelerator strongly indicates that they already have hardware in house. As a first stage in the deployment, Graphcore will sell the addin cards alongside a custom server. The company is claiming wide scalability for the architecture so we could see later versions addressing more power constrained markets in the future.
The lack of detailed data on the chip makes it impossible to review the technology at present but Toon is promising more information closer to their release date so I'll be keeping an eye on their progress and you'll see a full review in a future edition of The VPU Report