Lightmatter

Lightmatter is a photonic computing startup that offers a full stack of photonics-enabled hardware and software solutions that claims to simultaneously reduce power consumption and increase performance.  

The company offers the following three products:

Passage: This is a wafer-scale, programmable photonic interconnect device that enables arrays of heterogeneous chips to communicate and share data using photonics. This allows computer chips to share data with high bandwidth and extremely low latency that can be scaled up at low cost. It is claimed that Lightmatter has scaled up to supercomputing scales and provides more than 100x more bandwidth than existing chip-to-chip interconnect solutions. 

Envise: A general-purpose machine-learning accelerator that combines photonics and transistor-based systems. Envise is offered as a server blade form factor that offers up to 6.4 TB per second optical data interconnect for scaling up, one terabyte of DDR4 DRAM, and three terabytes of solid-state memory. 

Idiom: A software that allows developers to interface with standard deep-learning frameworks and AI model formats, including PyTorch, TensorFlow, or ONNX, to automate and optimize the deployment of models  on Lightmatter hardware. The software features ML libraries, graph compilers, and debugging and profiling tools to fix and optimize model performance.   

The uniqueness of Lightmatter’s chips is that they typically use optical flow to solve computational processes like matrix-vector products and use an array of microscopic optical waveguides to let light pass through them to perform logic operations. This form of optimal computing can increase the power of the chip by using more than one color at once while reducing costs by not requiring expensive fiber-to-chip attachments. As a result, the firm claims its solutions are 5x more energy efficient and 200x more bandwidth, which reduces operating costs.

Its pilot products are in the beta phase, and Lightmatter plans to begin mass production in 2024.