Backers include Samsung, Bosch and Hermann Hauser.
Graphcore is developing processors specifically for machine learning – an approach which, it says, will deliver up to 100x more performance than off-the-shelf GPUs. Graphcore calls its processors IPUs.
‘The IPU is a completely new type of processor designed to help customers accelerate the development of current and next generation machine intelligence products and services, says Graphcore, ‘the IPU has been optimized to work efficiently on the extremely complex high-dimensional models that machine intelligence requires. It emphasizes massively parallel, low-precision floating-point compute and provides much higher compute density than other solutions. The IPU holds the complete machine learning model inside the processor and has over 100x more memory bandwidth than other solutions. This results in both lower power consumption ‘
Graphcore’s IPUs will support TensorFlow and MXNet backed by a open source software framework of tools, drivers and application libraries, called Poplar. With a C++ and Python interface, Poplar is a graph programming framework designed to allow developers to modify and extend our wide set of libraries, making our IPU systems quick and easy to use. ‘
Graphcore!s IPU-Appliance (coming in 2017) is designed to lower the cost of accelerating AI applications in cloud and enterprise datacenters.
Graphcore’s IPU-Accelerator is a PCIe card which can be plugged into a server to accelerate machine learning applications.