Chipmakers size up the AI opportunity

Why is there an accelerating arms race in processors designed for AI and machine learning and what does this mean for businesses developing AI/ML to solve their own problems?

In August, chip giant Intel revealed details of a completely new architecture. Its Nervana Neural Network Processor (NNP), named after the company it acquired in 2016, promises features needed to execute large deep learning models, without the overhead needed to support legacy technology, the company said. The announcement would have been biggest hardware story of the week, but instead it was overshadowed by a Californian start-up which paraded the world's largest processor, also designer for AI, in front of the world's press.

Cerebras Systems hit the media spotlight by producing a chip the size of a dinner plate containing 400,000 programmable cores supported by 18 gigabytes of on-chip memory. Coming in the same week, the two announcements signal a growing arms race in AI-specific processors.

The market already hosts Nvidia, which made its fortune producing graphics processors, Qualcomm, Google and UK start-up Graphcore.

Processor makers are keen on this market because practical AI workloads are currently limited by hardware design. The prize is solving high-value problems in pharmaceuticals, finance and engineering which could justify the expense of developing high-cost processors at low volume, says IDC European AI systems senior analyst Jack Vernon.

To date, data scientists have approached AI with a mix of graphics processor units (GPUs), conventional CPUs and field-programmable gate arrays, the architecture for which can be altered according to the task.

Beyond von Neumann

CPUs, which run nearly all the world's PCs, servers and smartphones, are not suited to AI tasks, Vernon says. They rely on von Neumann architecture, an approach to solving problems sequentially which dates back to the 1940s. But neural networks which support deep learning are not sequential.

To continue reading this article register now