As artificial intelligence workloads expand, conventional silicon based processors are facing increasing pressure from rising energy demands and performance limits.
Lumai has unveiled what it describes as the world’s first optical computing system capable of running a billion parameter large language model in real time, while significantly reducing energy consumption. The system, called Iris Nova, is designed to support AI inference workloads with up to 90 percent lower power usage compared to traditional GPU based architectures.
The development comes as data center power consumption continues to rise sharply with the growth of generative AI services. Traditional silicon processors deliver higher performance through successive generations, but often at exponentially increasing energy costs, creating scalability challenges for large scale AI deployment.
Iris Nova addresses this by shifting computation from electrons to photons. Instead of relying solely on two dimensional silicon chips, the system uses light based processing to perform large scale mathematical operations in parallel. This allows multiple computations to be executed simultaneously within a three dimensional optical volume.
At the core of the system is a hybrid architecture. An optical tensor engine handles mathematical computation using light based processing, while a digital control layer manages system operations and orchestration. This combination allows the system to integrate with existing data center infrastructure while offloading intensive AI workloads to optical hardware.
The system has demonstrated real time inference on models such as Llama 8B and 70B, showing compatibility with modern large language model workloads. It is designed for deployment across hyperscalers, enterprise environments, and research institutions.
According to Suraj Bramhavar, alternative computing approaches are becoming increasingly important as AI systems scale beyond current silicon efficiency limits. He noted that optical computing could represent a viable path beyond traditional digital architectures.
Xianxin Guo described the development as part of a broader transition toward the post silicon era of computing, where new hardware paradigms will be required to sustain inference driven AI growth.
Click here for the official announcement.