Lightmatter has raised $154 million as it prepares to introduce a system that uses light to perform calculations central to artificial intelligence. The company is testing its hardware and software in beta pilots, with mass production planned for 2024.
Why Lightmatter is turning to light
Many AI workloads rely on matrix vector products, calculations typically handled by GPUs and TPUs. Those processors use silicon gates and transistors, but the source article describes mounting challenges in increasing their speed and density while managing power use and heat.
Large computing systems used to train models such as GPT-4 can consume substantial electricity and generate significant waste heat. Lightmatter argues that these pressures could make it harder to scale AI using conventional hardware alone.
The company’s approach uses arrays of microscopic optical waveguides. Light passing through these structures can perform logic operations, creating a system the article describes as an analog-digital hybrid. Because the waveguides are passive, Lightmatter says the main power needs are producing the light and handling the output.
More computation through multiple wavelengths
Lightmatter says its optical arrays can carry out different operations using multiple colors of light at once. In practical terms, the system uses different wavelengths: one operation might use 800 nanometers, while another uses 820.
The company calls these parallel operations “virtual chips.” The concept is to increase the computation performed by an array by assigning work to multiple wavelengths. Implementing that approach is not simple, but it is one way Lightmatter aims to expand the capacity of its photonic hardware.
The article notes that Lightmatter declined to provide specific performance or efficiency figures. Comparisons with conventional chips are also difficult because the company is using a different computing architecture. The company’s claims therefore remain to be assessed through pilots and deployment.
A full stack for AI workloads
Lightmatter’s offering combines three products: Envise computing hardware, Passage interconnect technology for linking larger computing operations, and Idiom software. The interconnect matters because processing capacity on a single board would not be enough for larger systems if it could not be connected effectively.
Lightmatter says Idiom integrates with PyTorch and TensorFlow, two applications used to build machine learning systems. Developers can import the company’s libraries and run their neural networks on Envise, according to CEO and founder Nick Harris.
This setup is intended to let developers work with familiar tools while the underlying calculations run on photonic hardware. Whether that transition works smoothly in practice is among the questions the beta pilots can help answer.
Funding and the path to deployment
Lightmatter grew out of optical computing research by Harris and his team at MIT, which is licensing relevant patents to the company. Lightmatter raised an $11 million seed round in 2018 and later received a further $80 million from investors. The new $154 million round is intended to support its next stage as it prepares for a commercial debut.
The funding came from SIP Global, Fidelity Management & Research Company, Viking Global Investors, GV, HPE Pathfinder and existing investors. The company’s pilots are in beta, and it planned mass production for 2024, with data center deployment anticipated as the technology and feedback matured.
Lightmatter’s chips are designed for AI tasks rather than general-purpose uses such as powering a laptop. The company is betting that specialized computing can help address the cost and complexity of AI systems at scale. The pilots will be an important step in showing how its photonic hardware, interconnect and software work together outside the lab.