Graphcore said on Monday that it had raised $50 million from Sequoia Capital to put the finishing touches on its machine learning chip for data centers. It is aiming to dethrone graphics chips from Nvidia, which Sequoia also invested in over twenty years ago.
Graphcore is wrestling with rivals like Wave Computing and Cerebras Systems to usher custom chips into data centers where software runs to grasp human speech or predict drilling spots for natural gas. But it is one of the closest to releasing its first product, the intelligence processing unit, which can be used to train algorithms to make inferences without explicit programming.
The IPU design excels at handling computational graphs, which can represent correlations within large amounts of data. The company, started by Icera Semiconductor's founders, also makes software that compiles code into a format digestible by Graphcore’s IPUs. Its first product, Colossus, will chew through deep learning tasks with over a thousand cores.
“Machine intelligence will cause an explosion of new applications and services that will transform every industry,” said Matt Miller, a Sequoia partner, who will join Graphcore’s board, in a statement. He added that the “unique nature of this workload and scale of this opportunity creates an opening where a valuable new chip company can be created.”
Last month, Graphcore released preliminary benchmarks that claim its custom silicon handles training and inferencing tasks an order of magnitude faster than accelerators based on Nvidia’s previous Pascal architecture. Nvidia says that its latest architecture Volta performs around 12 times faster than Pascal, training models in days rather than weeks.
The company, based in Bristol, United Kingdom, has now raised $110 million since it was founded in 2015. Its other investors include Dell Technologies as well as Demis Hassabis, one of the founders of Alphabet’s artificial intelligence unit DeepMind, Uber’s chief scientist Zoubin Ghahramani, and the founders of the OpenAI Laboratory.
The funding helps it to keep pace with the spending power of Nvidia, which has poured billions into customizing its graphics chips for deep learning while preserving its Cuda software stack. Last year, Intel spent around $400 million to acquire Nervana Systems, whose custom neural network processor will be released before the end of the year.
Graphcore also plans to hire new employees with the funding. “We have 75 people in the company and we’ll double that in the next two years,” Nigel Toon, the company's chief executive and one of its founders, told Forbes.
Graphcore previously said that it would start serving IPU cards to its first customer later in the year, but it seems to have recently revised that schedule. A spokeswoman said that the company plans to start shipping its first product to customers early next year with “wider availability” later in 2018.