Lightmatter builds a purely optical interconnect layer that lets hundreds to thousands of GPUs work synchronously in specially designed racks. Its photonic chips and on-rack optical wiring can enable 1,024 GPUs to operate together and the current interconnect delivers about 30 terabits, with 100 terabit capability on the roadmap. The company has been developing photonic chips since 2018 and is also working on new chip substrates to enable tighter networking tasks using light. Lightmatter positions itself as a foundry-like platform for hyperscalers, and several major datacenter companies are already customers though unnamed. Management expects power-per-chip and wafer-scale approaches to be key future differentiators and is preparing technology roadmaps accordingly. The company’s tech targets the interconnect bottleneck that limits scaling beyond single racks and aims to dramatically reduce latency and power overhead in large clusters. Lightmatter develops photonic AI hardware that uses arrays of microscopic optical waveguides to perform matrix-vector computations with light, targeting large AI workloads. Its full-stack offering includes Envise (computing hardware), Passage (interconnect) and Idiom (software) that integrates with PyTorch and TensorFlow. The company licenses relevant patents from MIT and highlights the ability to create “virtual chips” by using multiple wavelengths to multiply computational density. Lightmatter says this analog-digital hybrid approach can substantially reduce energy consumption compared with traditional GPU/TPU architectures. The company is running several pilots in beta and plans mass production in 2024. Financially, Lightmatter has been scaling since an $11M seed in 2018 and additional capital in 2021, and most recently closed a $154M Series C. Lightmatter develops photonic computing hardware and systems — including the Envise photonic AI accelerator — that perform certain neural-network linear algebra operations by steering light through optical circuits rather than using transistors. Its chips are currently optimized for high-throughput linear algebra (inference) and the company says its platform can be ~5x faster than an Nvidia A100 on large transformer models while using roughly 15% of the energy. The firm has built server-sized units that fit into standard racks and is working on a wafer-scale optical interconnect called Passage for future generations to improve chip-to-chip communication. Lightmatter has offices and more than 70 employees in Mountain View and Boston and is preparing an early access program to deliver test units to hyperscale customers. The company is hiring to support hardware production, go-to-market, service/support and the software stack required to operate the systems. Lightmatter develops ultra-efficient light-powered AI chips built on photonic processor technology. Its optical computer chips leverage the properties of light rather than electrical signals to enable fast and efficient inference and training engines. The company intends to use the newly raised funds to continue development of its light-powered computing platform. Financially, it announced $22M in a Series A-1 round, bringing its total Series A to $33M. The round was led by GV with participation from Matrix Partners and Spark Capital; GV partner Tyson Clark joined Lightmatter’s board. The funding supports continued R&D to advance photonics-based inference and training capabilities. Lightmatter develops photonic chips that perform key AI matrix-vector operations by routing light through configurable optical elements to compute results at the speed of light with far lower power and latency than conventional silicon chips. The technology originated from research by CEO Nick Harris at MIT; MIT owns and licenses some of the related patents to the company. Lightmatter has demonstrated a prototype chip with 32 “neurons” and is working toward chips with hundreds of neurons to increase density and parallelism. The chips are intended for specialty AI hardware used by developers and cloud providers rather than consumer devices. The company says its approach is compatible with existing CMOS fabrication methods, avoiding the need for new fabs. The recent financing is intended to build the team that will take the technology from prototype to product and commercial deployment.

Total raised

$667M

Funding rounds

5

Latest round

Series D

Latest activity

Oct 2024