As traditional semiconductor technologies approach their physical limits, the demand for computing power continues to rise, largely driven by the rapid expansion of artificial intelligence (AI). Addressing this conundrum, Lightmatter, a company founded by three MIT alumni, has developed pioneering computing technology that harnesses the properties of light to bolster data processing and transport.
Lightmatter’s initial pair of products are a chip designed for AI tasks and an interconnect that enables data exchange between chips. By incorporating photons and electrons, these products stimulate more efficient operations. “We’re addressing two main questions–’how do chips converse?’ and ‘how are these AI computations performed?’” explains Nicholas Harris, Lightmatter co-founder and CEO.
Signifying the enormous potential of this technology, Lightmatter recently raised over $300 million at a $1.2 billion valuation in 2023. With this backing, the company aims to reduce the excessive energy demand of data centers and AI models by collaborating with several major global tech firms. “Our technology will enable use of hundreds of thousands of next-generation computing units that wouldn’t be possible with existing technology,”, Harris commented.
Previously, Harris worked for semiconductor firm Micron Technology, gaining first-hand insight into the constraints faced by conventional transistor-focused approaches to boosting computing performance. His subsequent research at MIT into quantum computing and photonics, including development of silicon-based photonic chips, formed the foundation for Lightmatter’s technology.
Harris realized that photonic quantum computing systems could be utilized for deep learning, a prominent AI technique becoming increasingly important in various fields. Consequently, instead of pursuing an academic career as originally planned, Harris decided to establish a startup, alongside Darius Bunandar and Thomas Graham, to accelerate innovation with more funding.
Lightmatter’s Envise chip blends the computational skills of electrons, such as memory, with the light-based strengths of photonics, especially performing large matrix multiplications needed for deep-learning. Furthermore, data incoming in different light colors can be processed simultaneously– a crucial feature for energy efficiency. Passage, the company’s interconnect product, exploits the low latency and high bandwidth of light to connect processors, while allowing chips as large as whole wafers to operate as single processors.
Both products aim to bring improved energy efficiency to computing, an essential development given that, by 2040, an estimated 80 percent of global energy consumption could be dedicated to data centers and computing, with AI being a significant portion. Harris comments, “AI model training deployments are headed towards consuming hundreds of megawatts – energy usage on par with urban areas”.
Lightmatter’s technology is also silicon-based and can be produced with existing processes, facilitating mass production alongside chipmakers and cloud service providers. The company’s long-term goal is to develop and replace computer components with light-centered counterparts, with computing and interconnection as its current focus. By doing so, Lightmatter plans to chart a new trajectory for the computing industry that will have profound environmental and economic implications.
