On March 25, 2025, Lightelligence officially released its new photonic-electronic hybrid computing card, the “Lightelligence Tianshu.” Dr. Yichen Shen, Founder and CEO of Lightelligence, stated at the launch event: “The Lightelligence Tianshu marks the first application of photonic-electronic hybrid computing in complex commercial models, representing a significant breakthrough in the productization and commercialization of our technology. We firmly believe that photonic-electronic hybrid computing will drive a revolution in computational power for fields such as artificial intelligence, large language models, and smart manufacturing.”
Tianshu is a programmable photonic-electronic hybrid computing card that deeply integrates the respective advantages of optical and electronic chips, utilizing advanced 3D packaging technology. Compared to previous generations, this product achieves significant performance improvements in areas such as optoelectronic integration density, photon matrix scale, precision, and programmability. While supporting scientific computing tasks (such as Ising algorithms), Tianshu has strengthened support for commercial algorithm applications like ResNet50, significantly enhancing the product’s general adaptability while maintaining the inherent advantages of optical computing.
Lightelligence Photonic-Electronic Hybrid Computing Card “Tianshu”
In December 2021, Lightelligence first unveiled its photonic-electronic hybrid computing products, successfully demonstrating a speed advantage of hundreds of times over mainstream GPUs for specific algorithms. The release of Tianshu today signifies that photonic-electronic technology has made a massive leap in product realization.
Tianshu adopts an incoherent architecture, a design choice that not only facilitates system scalability but also offers excellent anti-interference capabilities and higher computational precision. Its core processor consists of an Optical Processing Unit (OPU) and an Electronic Application-Specific Integrated Circuit (ASIC). The optical and electronic chips work in synergy through advanced 3D packaging technology, operating at a clock speed of 1GHz with an output precision of 8-bit. The optical chip area has reached 600 square millimeters, three times that of the previous generation. It houses over 40,000 devices on the optical chip, with further reduced device dimensions and significantly improved integration density. Tianshu supports a maximum matrix scale of 128×128, four times that of its predecessor, doubling both computational power and flexibility. Users can freely configure calculation matrix coefficients via APIs, giving Tianshu greater adaptability and optimization potential.
Photonic computing is a passive operation where tasks are completed as light passes through the photonic matrix, fundamentally subverting the operational logic of traditional CMOS electronic chips. The performance improvement of photonic chips depends on parameters such as photon matrix scale, clock speed, and the number of wavelengths, rather than relying on increases in transistor density or advancements in chip manufacturing processes. At the launch event, Dr. Huaiyu Meng, Chief Technology Officer of Lightelligence, introduced for the first time a new standard called Effective Optical Processing Power (EOPP). This is a comprehensive computational power evaluation method that considers matrix scale, output precision, weight refresh speed, and other factors. Compared to current mainstream electronic chip metrics, EOPP better aligns with the principles and characteristics of photonic computing. “Photonic-electronic hybrid computing is the future trend in computational power. We hope the industry will have a more objective standard for measuring photonic computing power,” said Dr. Meng.
To achieve efficient integration between optical and electronic chips, Tianshu utilizes a photonic-electronic hybrid 3D TSV (Through Silicon Via) + FlipChip packaging technology. The use of TSV significantly reduces transmission delay between optoelectronic chips while enhancing signal integrity and thermal performance, simultaneously saving chip area and providing greater flexibility for chip design. Dr. Meng stated: “From its inception, Lightelligence recognized the importance of packaging for photonic-electronic integration and has continuously worked to overcome technical hurdles. The successful implementation of TSV resolves the bottlenecks of traditional packaging technologies, offering value and significance that benefit the entire photonic-electronic hybrid industry.”
On the software side, Tianshu is equipped with Lightelligence’s photonic-electronic hybrid computing software stack. Its operator library includes RVV (RiscV Vector) operators, electronic matrix (dMAC) acceleration operators, optical matrix (oMAC) acceleration operators, and supports CV-class and LLM-class models, as well as non-AI operators such as Ising and LineSolver. Users can flexibly utilize these operators via the Lightelligence compiler to build efficient application models. Additionally, the platform supports user-defined operators, allowing for compilation and optimization through OpenCL C/C++ languages, further expanding the flexibility of algorithm development.
Lightelligence Photonic-Electronic Hybrid Computing Card “Tianshu”
Lightelligence’s software stack is deeply integrated with mainstream frameworks such as PyTorch and ONNX. Customers can use the stack to directly accelerate and verify models and algorithms using Tianshu’s optical matrix and electronic matrix acceleration units, or compile and deploy models on edge devices for inference via the Lightelligence compilation framework. Through this software stack, Tianshu successfully ran deep convolutional neural network model ResNet50 and the AI large language model LLaMA 2, marking the first implementation of photonic-electronic hybrid computing in commercial scenarios.
Mr. Long Wang, Chief Operating Officer of Lightelligence, stated: “Behind Tianshu lies the collaborative effort of Lightelligence’s teams specializing in silicon photonics, digital systems, analog circuits, packaging, system integration, and software. We hope to attract more developers and ecosystem partners to explore broader application scenarios for photonic-electronic hybrid computing with us, moving forward together toward the commercialization of this technology.”
Lightelligence has already initiated research and development on its next-generation photonic-electronic hybrid computing products. Dr. Yichen Shen, Founder and CEO, added: “Future products will further enhance computational capabilities to support more complex commercial application scenarios, providing new types of computational power support for artificial intelligence and data centers.”