Workflow
光跃LightSphere X分布式OCS全光互连芯片及超节点解决方案
icon
Search documents
曦智科技沈亦晨:3D CPO有望在五年内实现
Core Insights - NVIDIA has introduced two silicon photonics CPO switches at the GTC conference to enhance the interconnect speed and energy efficiency of GPU clusters, making CPO a focal point in the industry [1] - The evolution of optical interconnect technology is crucial, with a roadmap from pluggable optical modules to 3D CPO, significantly increasing single-chip bandwidth [1][3] - The demand for computing power is growing globally, necessitating advancements in optical interconnect products to address this challenge [1][2] Group 1: Optical Interconnect Technology - The transition from traditional network interconnects to NVIDIA's GB200 NVL72 supernode can increase throughput by over three times compared to conventional methods [2] - Domestic AI chip and server manufacturers are increasingly adopting the supernode concept, indicating a shift in industry trends [2] - The two main paths for expanding supernode scale are using high-density cabinets or multiple cabinets with direct optical interconnect capabilities [2][3] Group 2: Challenges and Solutions - Current solutions face bandwidth and resource wastage issues, leading to network congestion, highlighting the need for revolutionary interconnect systems [3] - The proposed evolution of optical interconnect technology includes a shift towards 3D co-packaged optics, which could enhance interconnect bandwidth by 1-2 orders of magnitude within five years [3] - The complexity of connecting multiple GPUs necessitates advanced scheduling systems for efficient network management [3] Group 3: Innovations and Developments - At the 2025 WAIC, the company launched the LightSphere X distributed OCS all-optical interconnect chip and supernode solution, demonstrating its application with partners [4] - LightSphere X is recognized as the first domestic solution for optical interconnect and GPU supernodes, winning the SAIL Award for its innovation [4][5] - The technology allows for flexible scaling of supernodes, reducing deployment costs and enabling dynamic adjustments based on computing needs [5] - Performance metrics indicate that the unit interconnect cost is only 31% of that of the NVL72, with a significant increase in model computing efficiency [5]