« Previously: Getting computers to write programs  

On the same day Gamalon that came out of stealth mode, has Graphcore also described the software behind its AI processor. Its so-called Poplar framework is being designed to map on to one or many of its chips an entire deep-learning model, called a graph, giving it a performance edge over today’s CPUs and GPUs.

Graphcore is one of a handful of closely watched AI processor start-ups including Corenami and Wave Computing, developing new microprocessors specifically for deep learning. In a blog posted today, Graphcore described Poplar as a C++ framework for mapping on to its processor the primitives called graphs created in deep learning neural networks.

The start-up described graphs as vertices and connecting edges in a neural network. It also showed striking visualisations of example graphs its chip can display (below).

Graphcore model 01 (cr) Figure 1: An image of a deep-learning model for astrophysics, presumably displayed using a Graphcore processor.

Graphcore said it is “building a comprehensive, open source set of Poplar graph libraries” for its processor that can be used in existing deep-learning frameworks including TensorFlow and MXNet. The Graphcore libraries aim to drive deep learning beyond today’s “tensor-centric” models, it said.

Poplar’s graph compiler aims to translate the underlying operations used by today’s machine-learning frameworks into optimised application code for its chip. “The compiler builds up an intermediate representation of the computational graph to be scheduled and deployed across one or many” of its chips, the blog said.

One of the interesting capabilities of the Graphcore compiler is it can display a graph, so an application written at the level of a machine-learning framework reveals an image of the graph running on the start-up’s chip. The start-up also is developing debug and analysis tools as well as C++ and Python interfaces for Poplar.

“Machine learning is the future of computing and a graph processor like [the Graphcore chip] is the architecture that will carry this next wave of computing forward,” the blog said.

 
« Previously: Getting computers to write programs