Why HPC Chip Designers Are Turning to Linear Pluggable Optics
The rapid evolution of high-performance computing (HPC) and hyperscale data centers has pushed interconnect technologies to their limits. As AI-driven workloads, large-scale machine learning models, and GPU-accelerated computing demand higher bandwidth, lower latency, and greater power efficiency, traditional electrical and optical connections are struggling to keep up.
To address these challenges, chip designers and network architects are exploring new approaches to data transmission. One technology gaining traction is Linear Pluggable Optics (LPO)—a simplified optical module design that shifts certain digital processing functions from the module to the host ASIC. This approach offers a promising balance between power savings, performance, and scalability, making it an attractive alternative to traditional optical transceivers and even co-packaged optics (CPO) in some cases.
The Growing Need for Better Interconnects
Hyperscale data centers, which power everything from cloud computing to AI training, are experiencing a surge in bandwidth demands. The next generation of switch chips is moving from 25.6 Tb/s to 51.2 Tb/s, with 800G and 1.6T Ethernet ports becoming the new standard. Meanwhile, PCIe and CXL-based accelerators require ultra-fast interconnects within HPC clusters.
While electrical copper interconnects have been the backbone of these systems, they are quickly becoming infeasible at higher speeds. Copper traces suffer from increased signal loss, power consumption, and limited reach beyond a few inches at 112 Gbps per lane. This is where optical interconnects come in—offering the ability to carry high-speed signals over long distances without excessive power dissipation.
However, the challenge with traditional pluggable optical transceivers is that they incorporate Digital Signal Processors (DSPs) for retiming and signal conditioning. These DSPs consume significant power and add extra latency to the data path, which becomes problematic for AI and HPC workloads that demand ultra-low-latency communication.
This is where Linear Pluggable Optics (LPO) offers a breakthrough—by eliminating the DSP within the optical module and instead relying on the host ASIC’s built-in SerDes (Serializer/Deserializer) processing, LPO reduces both power consumption and latency, making it an ideal choice for next-generation data centers and HPC clusters.
What is Linear Pluggable Optics (LPO)?
Linear Pluggable Optics (LPO) is a form of optical transceiver that eliminates the internal DSP chip, instead relying on the switch ASIC’s analog signal processing. Unlike traditional pluggable optics, where an embedded DSP cleans up the signal before sending it to the host, LPO directly transmits the high-speed signal with minimal processing in the module itself.
Key Advantages of LPO:
- Lower Power Consumption – By removing the DSP from the module, LPO reduces power per bit, cutting optical transceiver energy use by 30-50% compared to traditional optics.
- Reduced Latency – The absence of extra retiming stages in the module minimizes serialization delays, which is critical for HPC, AI, and latency-sensitive workloads.
- Higher Bandwidth Density – LPO allows for simpler module designs that support 800G and future 1.6T links, helping to scale hyperscale networking infrastructure.
- Compatibility with Existing Systems – LPO can be deployed within standard pluggable module slots (e.g., QSFP-DD, OSFP), enabling gradual adoption without requiring a full redesign of networking equipment.
How LPO Works in a Data Center Network
A standard leaf-spine topology in a data center connects Top-of-Rack (ToR) switches to spine switches using optical transceivers. Traditionally, these transceivers contain both optics and a DSP for signal retiming. In an LPO-based system, however, the host switch ASIC performs equalization and error correction, eliminating the DSP in the module.
This results in a simpler, more efficient architecture with fewer electrical-to-optical conversions, leading to lower power consumption and faster signal transmission.
LPO vs. Co-Packaged Optics (CPO): Competing Approaches
While Co-Packaged Optics (CPO)—where optical components are directly integrated inside the switch or HPC chip package—has been touted as the future of ultra-high-speed networking, LPO offers a more immediate and practical solution for many applications.
| Feature | Linear Pluggable Optics (LPO) | Co-Packaged Optics (CPO) |
|---|---|---|
| Power Efficiency | ~30-50% savings over standard pluggables | Even greater savings (but complex cooling) |
| Latency | Lower than DSP-based optics | Minimal latency (shorter electrical paths) |
| Deployment Complexity | Simple drop-in replacement | Requires full switch redesign |
| Cost & Flexibility | Lower cost, retains modular, replaceable optics | Higher initial cost, non-replaceable optics |
| Availability | Near-term (2025-2026 adoption) | Longer-term (post-2027) |
While CPO is ideal for ultra-dense, high-performance AI supercomputing environments, LPO is a much easier upgrade path for existing data centers, making it the preferred option for switch-to-switch and rack-level connectivity in hyperscale clouds and HPC clusters.
Challenges in Implementing LPO
Despite its advantages, LPO is not without challenges.
- Signal Integrity & Reach – Without a DSP in the module, the burden of signal processing shifts entirely to the host ASIC, making PCB trace design and optical channel equalization more critical.
- Standards & Interoperability – Because LPO lacks a retimer, industry-standard signal specifications are still evolving, requiring collaboration between switch vendors, optics manufacturers, and hyperscalers.
- Scaling to 224G SerDes – As the industry moves to 224 Gbps per lane signaling, maintaining linear transmission without performance degradation will be a key engineering challenge.
Despite these hurdles, major industry players like Broadcom, Marvell, Microsoft, Meta, and Google are actively exploring LPO solutions for their next-generation data center infrastructure.
The Future of LPO in AI and HPC Systems
AI Training Clusters and GPU Networking
The rise of AI superclusters, which rely on thousands of interconnected GPUs, demands ultra-fast, low-latency communication between nodes. LPO reduces power draw in InfiniBand and Ethernet interconnects, making it a compelling choice for AI workloads where energy efficiency is critical.
PCIe and CXL Over Optical Links
Beyond Ethernet, LPO is being investigated for PCIe and CXL interconnects, which are fundamental to CPU-to-GPU, memory pooling, and high-speed storage architectures. By extending PCIe lanes over linear optical connections, LPO could help enable disaggregated HPC systems, allowing for flexible, modular compute architectures.
A Stepping Stone to Co-Packaged Optics
While CPO is still several years from mainstream adoption, LPO serves as a critical transition technology—delivering many of the same benefits without requiring a complete overhaul of switch and server designs.
Conclusion: LPO as the Next-Generation Optical Solution
As data center bandwidth demands continue to grow, Linear Pluggable Optics (LPO) offers an efficient, practical path forward for HPC chip designers. By eliminating redundant DSPs, reducing power and latency, and retaining the flexibility of modular optics, LPO is positioned to play a crucial role in next-generation networking and AI computing systems.
While co-packaged optics (CPO) may be the long-term vision, LPO is here now—providing a powerful, scalable, and cost-effective solution for the growing interconnect challenges in AI, HPC, and cloud computing.
