Silicon photonics integrates optical and electronic devices on silicon to enable efficient transmission and processing of signals in the optical domain. Compared with traditional electrical interconnects, it offers ultra-high bandwidth, ultra-low power, and compact footprint—a key technology for removing data-transfer bottlenecks in data centers, high-performance computing, and AI clusters.
Beeplux's in-house Optical Circuit Switching (OCS) technology enables high-speed, low-latency data transport with all-optical paths and supports dynamic topology reconfiguration. Optimized for large "elephant flows," it is well suited for AI training and other high-performance workloads. We have mass-produced 8×8 OCS chips and are scaling toward 128×128 to build next-generation, high-port-count, low-power optical interconnect architectures.
Beeplux provides leading high-speed optical interconnect technologies spanning 400G to 1.6T optical engines, with advanced Co-Packaged Optics (CPO) capabilities and ongoing development of Optical I/O (OIO) chiplets. Our interconnect solutions enable high-bandwidth, low-power communication between switching silicon and compute silicon, compatible with PCIe, Ethernet, and various proprietary protocols—significantly increasing bandwidth between chips and boards to accelerate data centers, AI clusters, and cloud-scale systems.
Packaging options: Supports Near-Packaged Optics (NPO) and Co-Packaged Optics (CPO) advanced packaging.
We own key processes and proprietary IP spanning design through manufacturing:
We deliver a portfolio of products built on leading silicon photonics, designed to overcome conventional interconnect bottlenecks and meet next-generation data communication needs—helping customers build efficient and reliable optical interconnect infrastructure.
We offer OCS products from board-level 8×8 up to 128×128 optical cross-connects to enable high-speed GPU-to-GPU interconnects, with a roadmap toward 1024×1024 architectures. Key advantages include temperature insensitivity, polarization insensitivity, low insertion loss, and fast switching. Built on a non-blocking architecture and supporting hot-swap, our systems are designed for stable operation in complex data-center environments and meet the stringent bandwidth and reliability requirements of hyperscale AI clusters.
We provide 400G, 800G, 1.6T, and higher-rate optical engines and modules, compatible with mainstream interfaces such as PCIe and Ethernet. Core offerings include multi-protocol, high-density, low-power optical I/O chiplets supporting NPO/CPO advanced packaging—delivering flexible and efficient optical interfaces for compute and networking equipment.
Built on silicon photonics, this product replaces the complex fiber structures found in traditional FMCW systems to achieve high integration and significant cost reduction. It supports high-precision ranging within 100 meters and can be customized for scanning methods (e.g., semi-solid-state scanning), ranging accuracy (down to sub-millimeter), and target statistics.
By introducing a hybrid optical-electrical switching architecture, OCS handles sustained high-bandwidth "elephant flows" while traditional electrical switching handles bursty, small "mice flows." This approach can significantly reduce data-center hardware counts (e.g., Spine switches and optical modules), lowering overall power consumption and latency and improving total cost of ownership.
Benefits: Compared with all-electrical architectures, it can reduce electrical switch counts by up to 87%, cut power by ~40%, and reduce latency by ~10% (based on Google Apollo OCS case study).
To address communication bottlenecks in large-scale GPU/accelerator clusters for AI training and HPC, our OCS solutions provide a dynamically reconfigurable, high-speed direct-connect network. Across tensor parallel (TP), pipeline parallel (PP), and data parallel (DP) training, OCS optimizes communication paths, reduces cross-node latency, and significantly improves training efficiency and compute utilization.
Use cases: Parallel training optimization for GPT-3, BERT, Switch Transformer, long-context models, and more.
For sensing in complex environments, we provide modular, high-precision sensor solutions covering wide-range wind measurement, millimeter-level ranging beyond 100 meters, in-field speed measurement, and MHz-level picometer vibration detection. Our products emphasize high reliability and customization for a broad range of intelligent systems and industrial applications.
Applications: Smart manufacturing, unmanned systems, smart cities, environmental monitoring, structural health monitoring, and more.
Shanghai Bopu Semiconductor possesses a rare full-scale industrial chain capability for mass-produced silicon photonics in China, with its core team having over 15 years of experience in the silicon photonics industry. Leveraging vertically integrated capabilities in design, manufacturing, packaging, and testing, the company has achieved large-scale production of 100G/lane and 200G/lane silicon photonic PIC chips, completed process and mass production verification for domestic 12-inch CMOS silicon photonics production lines, continuously advanced NPO/CPO technology implementation, and independently developed dynamically reconfigurable silicon photonic OCS solutions to address the issues of insufficient bandwidth, high power consumption, and scalability challenges in AI computing clusters. The company adopts the "mass-produced silicon photonics IP + customized development" model, providing layered technology delivery and joint customization for optical modules, switches, and cloud providers, helping clients rapidly build silicon photonics capabilities using their own resources.
We look forward to connecting with you and exploring the potential of silicon photonics in data centers, AI, and high-performance computing. Please reach out for inquiries and collaboration.