Startup Graphcore takes on Nvidia with latest A.I. chip

British startup Graphcore is taking on semiconductor titan Nvidia with a new computer chip designed specifically for running cutting-edge artificial intelligence algorithms.

Graphcore, which is based in the English city of Bristol, unveiled a new computer chip Wednesday that packs a remarkable 59.4 billion transistors and almost 1,500 processing units into a single silicon wafer.

The company said that in benchmark tests, its chips—which are sold in a set of four designed to work together—performed up to 16-times faster than those from Nvidia.

Nvidia’s chips, which were originally designed to handle the intensive computing work needed for computer graphics, are currently the market leaders for machine-learning applications.

“Nvidia is the 200-pound gorilla but we see ourselves as the main challenger,” Nigel Toon, Graphcore’s chief executive, said.

Among those trying out Graphcore’s new technology are J.P. Morgan Chase, Oxford University, and the U.S. Laurence Berkeley National Laboratory. Another is Atos, the French supercomputing vendor that is helping a number of European research labs access the new chips.

Graphcore is among a growing number of startups, as well as some large tech firms, that have created chips designed for A.I. Google was one of the first to create such chips for use in its own data centers, but it does not sell them commercially.

Intel, which has long dominated the market for general computing chips and data center servers, has bought several startups working on A.I.-specific chips, including, most recently, Israel’s Habana Labs in December. But so far, it has struggled to make inroads in the new market.

Graphcore, which was founded in 2016 by a team that had sold a previous hardware company to Nvidia, debuted its first generation of A.I.-specific chips in 2018 and they are already in use with a number of customers, including a number of hedge funds and banks, and, most notably, Microsoft’s Azure data centers.

The second generation of A.I. Chips, which Graphcore unveiled Wednesday, are called the Mk2 IPU (short for intelligence processing unit). They are designed specifically to handle the very large machine-learning models that are being used for breakthroughs in image processing, natural language processing, and other fields. For instance, San Francisco A.I. Research company OpenAI’s latest language model, called GPT-3, takes in 175 billion different variables.

To address that need, Graphcore’s chips are designed to work together in large high-performance computing clusters. The company is marketing them alongside its own proprietary software that will coordinate those clusters, with up to 64,000 of its latest-generation chips, allowing them to process huge amounts of data in parallel.

“There is definitely a requirement from customers to run these large models,” Nigel Toon, Graphcore’s chief executive, said. “With our system, you can develop them and run them at scale and train them at a sensible timeline.”

Last year, Nvidia paid $6.9 billion to acquire Mellanox, a company that specialized in software for coordinating these kinds of computing clusters to up its own capabilities in this area.

Toon says Graphcore’s own tests show how the speedup offered by its latest-generation chips can save companies money. On a state-of-the-art image classification benchmark test, for instance, it would take just eight of Graphcore’s new four-chip clusters, working together, to train an algorithm, at a cost of $259,000. To achieve the same result using Nvidia’s DGX, a system which combines 8 of its top-of-the-line chips, would require 16 of these clusters and cost about $3 million.

Graphcore, which employs about 450 people globally, including a team in Oslo, Norway that has worked on the systems to run very large computing clusters, has received more than $450 million in venture capital funding to date.