Key Findings
A recent Pandaily article concludes that Co-Packaged Optics (CPO) will serve as a core solution for next-generation AI data centers, effectively alleviating GPU communication bottlenecks caused by increasing data center traffic. CPO is positioned as a pivotal technology to achieve the “ultimate goal” of low latency and low power consumption, essential for enhancing data center performance and energy efficiency. Specifically, an explosive adoption of Near-Packaged Optics (NPO) is anticipated between 2026 and 2027, followed by a gradual transition to CPO, outlining a clear roadmap for optical integration.
Technical / Clinical Details
The article elaborates on a three-stage roadmap for optoelectronic integration, illustrating the evolutionary path of optical technology in AI data centers:
- Stage 1: Pluggable Optical Modules: This widely used format converts electrical signals into optical signals for data transmission. However, it faces limitations in meeting the ultra-high bandwidth and low power consumption demands of AI workloads.
- Stage 2: Near-Packaged Optics (NPO): Expected to see explosive adoption from 2026 to 2027, NPO involves placing optical engines in close proximity to the ASIC package. This shortens electrical trace lengths, reducing power consumption and latency, thus offering superior performance over traditional pluggable modules.
- Stage 3: Co-Packaged Optics (CPO): Considered the “ultimate goal” for low latency and low power consumption, CPO integrates optical engines directly within the same package as ASICs like CPUs or GPUs. This minimizes electrical signal transmission distances, leading to dramatic performance improvements and optimized energy efficiency. Such integration significantly expands communication bandwidth between AI chips, enabling the scaling of high-density AI clusters.
These technologies are particularly critical for addressing the growing computational demands in AI training and inference. CPO resolves AI data center bottlenecks by substantially reducing power consumption per bit and enabling terabit-level data transfer speeds.
Background & Context
The explosive growth of AI workloads places unprecedented demands on data center interconnects. Traditional copper-based interconnects are approaching their limits in terms of bandwidth, power consumption, and thermal management. Consequently, major industry players, including NVIDIA, are making a significant shift towards optical interconnects. NVIDIA’s plan to fully adopt silicon photonics and CPO in its AI data center platforms by 2026 strongly supports this trend. CPO is specifically recognized as an indispensable technology for enabling high-speed and efficient communication between GPUs, especially with the evolution of large language models (LLMs) and generative AI.
Strategic Significance & Outlook
The roadmap presented in the Pandaily article provides insight into the future evolution of AI data centers over the coming years. Following the rapid introduction of NPO, the eventual standardization of CPO will dramatically enhance AI computing capabilities and efficiency. This technological evolution will enable further advancements in AI, opening new horizons for data-intensive applications such as autonomous driving, HPC, and cloud services. The widespread adoption of co-packaged optics signifies a fundamental paradigm shift in data center design, contributing to the realization of more sustainable and high-performance AI infrastructure.
Source: https://pandaily.com/co-packaged-optics-ai-data-centers-jun2026
Comments