Today, selling custom chips for artificial intelligence is still a small business. The current market is worth around $2.5 billion, one half of one percent of global semiconductor sale estimates this year, according to Intel. But startups are still raising billions of dollars of venture capital to plant a stake in the market, which could be worth $20 billion in 2021, according to market researcher Morningstar.
Graphcore, which has designed a custom chip and the accompanying programming tools, said on Tuesday that it had raised another $200 million. The funding brings the company's total to $310 million since it was founded in 2016. The investment gives Graphcore more financial firepower as it moves to ramp up production of its chips, which are designed for use in data centers.
The funding adds to the $50 million it raised from Sequoia Capital last November to fund the development of its intelligence processing unit, or IPU. Graphcore's IPU is designed to handle both training - teaching algorithms the differences between sounds, objects or other data - and inferencing - introducing the software to new data and interpreting it. The company said that it runs 10 to 100 times faster than today's chips.
Graphcore and other startups are raising billions of dollars as the amount of compute required for training and inferencing continues to grow. Last month, Wave Computing raised more than $85 million to bring the company's total amount raised to $205 million since it was founded in 2009. Last year, Cerebras Systems was reportedly valued at around $1 billion. Graphcore's latest funding increases its valuation to $1.7 billion.
"Machine intelligence marks the start of a new age of computing which needs a radically different type of processor and software tools,” Nigel Toon, Graphcore's chief executive and co-founder, said in a statement. “This new, fast growing market creates the opportunity for Graphcore to build a major global technology company,” he added. Graphcore has shipped its first generation of data center hardware to early customers.
Graphcore also has a direct line to many leading artificial intelligence researchers. The company is financed in part by Uber's chief scientist Zoubin Ghahramani, DeepMind's founder Demis Hassabis, and the founders of the OpenAI foundation - including Greg Brockman, Ilya Sutskever, Pieter Abbeel and Scott Gray. Google's Geoff Hinton, considered the godfather of deep learning, has stashed one of the company's chips in his wallet.
"Training machine intelligence models in minutes rather than days or weeks will profoundly transform how developers work, how they experiment and the results they will see," Brockman said in a statement. "Being able to experiment across a much broader front, at a much faster pace will create new breakthroughs." Ghahramani added in a statement: "Current hardware is holding us back from exploring these different approaches."
Graphcore is trying to dethrone Nvidia, which has jumped ahead in the market for chips capable of training machine learning software. The startup is also fighting Intel, which has taken the lead supplying chips that can handle inferencing. Intel has taken advantage of its dominant share of the data center market. Google and Amazon have carved into the company's total addressable market by renting out - or planning to start renting out - custom chips over the cloud.
Graphcore's latest funding was led by venture capital firm Atomico and investment firm Sofina. It was also supported by Sequoia Capital, Microsoft, Robert Bosch, BMW, Dell Technologies, Samsung, and others. The company said that the $200 million would be used to expand its global operations and build new generations of chips, including for the automotive market. "We had to tell lots of potential investors that we didn’t have room for them," Toon said.