A new startup could revolutionize computing forever with technology that uses light instead of charge.
Lightmatter, a three-year-old startup, is seeking to upend AI computing with photonic chips. The technology has existed for years now but it is only now that a company has come up with a way to implement it at scale. And investors are queuing up. Lightmatter has successfully raised $80 million in a round led by Viking Global Investors, with participation from HP Enterprise, Lockheed Martin, SIP Global Partners, and previous investors GV, Matrix Partners and Spark Capital. It brings Lightmatter’s total raised to about $113 million which includes the initial $11 million A round that the founders received right after graduating from MIT, then GV investing a $22 million A-1, and this $80 million.
The technology that Lightmatter is building focuses on ultra-fast photonic chips specialising in AI work. In short, Lightmatter’s chips perform complex calculations fundamental to machine learning. But instead of using logic gates and transistors to record and manipulate data, they use photonic circuits that perform the same calculations by altering the path of light.
The company has expanded to more than 70 employees in Mountain View and Boston and is looking at expanding even further as it gets close to launching a product. The Lightmatter Envise, is what the company calls a “general-purpose photonic AI accelerator” and could be that product. It is a server unit designed to find itself in normal data center racks but is equipped with multiple photonic computing units.
A quick look at the website tells us that the Envise is 5x faster than an Nvidia A100 unit on a large transformer model like BERT, while only accounting for about 15% of the energy consumption. Such performance allied with efficiency makes the Envise very interesting to big tech and social media companies who are constantly after more computing power and are stuck with ever increasing energy bills.
The company intends ro have Envise units available to some early customers by the end of this year followed by a large scale rollout later. The founders are quick to dismiss talk of photonic computers reaching consumer grade computers any time soon. But the fact that 10 percent of the world’s energy consumption will be from server farms by 2030 means that photonic computing could soon be a reality there.