Partnering with Graphcore: Building a Chip for the Age of Machine Intelligence
By Matt Miller
Published November 12, 2017
Editors note: This piece was originally posted on Matt’s LinkedIn
Machine intelligence is the most impactful technology trend of the next decade and beyond. In the past few years, we have seen a rise in new machine learning technologies that will radically transform traditional industries. From autonomous cars to new air traffic control systems and advances in medical imaging, we are surrounded by the early signs of the machine intelligence age.
This rise has been fueled by the discovery that the graphics processing unit (GPU), with its more parallel architecture, provides meaningful performance improvements for machine learning over the more general-purpose central processing unit (CPU). The result has been amazing for GPU maker NVIDIA. Its share price has grown by ~7x over the last two years, with the company adding more than $100B in market cap.
While the GPU is the best commercially available hardware for machine learning, it was not designed for this use case. To unleash the next wave of innovation, several young companies are now building specialized chips that will provide meaningful performance advancements over the GPU.
At Sequoia, we looked at this market thoroughly and met teams in the US, Europe, Israel and China. We sought input from across the machine learning community, enterprise technology landscape and the large cloud providers. Time and time again, we heard about a small company in the southwest of England: Graphcore.
Graphcore has built a new microprocessor, the intelligence processing unit (IPU), that utilizes a graph architecture with thousands of cores and memory laid out across the chip. In the first iteration of the IPU, the result is a 100x performance improvement over the GPU. The founders of Graphcore are not newcomers to the silicon world; they have built and shipped eleven microprocessors in their careers and helped build numerous successful chip companies in the past. Of the new entrants, Graphcore is the furthest along in development and has the most sophisticated customer interest.
For these reasons, we’re thrilled to announce that we’ve partnered with Graphcore and invested $50M.
More than 45 years ago, Don Valentine left National Semiconductor to found Sequoia Capital. He saw an opportunity to help high-potential chip makers build great long-term businesses. Today, we set out to do the same again with Graphcore. With the Sequoia presence in the US and China we hope to help Graphcore build out teams and infrastructure in these markets as they work to serve the demand for their planned 2018 product launch.
We believe Graphcore has the potential to be the most transformative chip company we have partnered with since we invested in NVIDIA over 20 years ago.
We welcome Graphcore to the Sequoia family.
To learn more about Graphcore, please visit: http://www.graphcore.ai
Graphcore has built a new microprocessor that utilizes a graph architecture with thousands of cores and memory laid out across the chip.