Contributed Article

Provided by: Belden

For over a decade, Power Usage Effectiveness (PUE) has been the go-to metric for evaluating data center efficiency. As artificial intelligence (AI) workloads scale rapidly, however, relying so heavily on PUE has run into a problem.

Unfortunately, PUE’s facility-wide perspective provides little insight into how efficiently an AI workload is being executed. As AI workloads continue to demand growing amounts of power, the data center industry is pivoting to a more granular metric – energy per inference.

Optimizing for this metric requires scrutinizing every layer of the hardware stack. While GPUs and advanced cooling systems receive significant attention, the optical interconnects that move data within clusters are quickly becoming a strategic concern.

Why the Interconnect Layer Matters for AI Power Efficiency

In legacy data centers, the power draw of a single transceiver has always been a topic of importance, but less so than today. Modern AI fabrics have changed the conversation.

A high-performance GPU server can rely on ten or more optical transceivers for communication across spine, leaf, and server switch architectures. As data rates shift upward to 800G and 1.6T, the energy cost of moving bits begins to rival the cost of processing them. At this scale, optical design choices begin to influence energy per inference in a measurable way.

How Linear Pluggable Optics (LPO) Work

Linear Pluggable Optics (LPO) simplify the transceiver by removing the most power-intensive parts – the Digital Signal Processor (DSP) and Clock Data Recovery (CDR). In standard full re-timed optics (FRO), the DSP alone can account for roughly 40% of total optic power consumption.

Instead, LPO shifts signal conditioning responsibilities to the host switch silicon, specifically the SerDes (Serializer/Deserializer). This handoff results in a leaner optical module that slashes the power draw per 800G link from around 13W-16W to 7W-9W. Lower transceiver power means less heat generated at the switch faceplate. Less heat means reduced cooling demand across the room and less pressure on HVAC systems.

Additionally, by bypassing the DSP processing cycle, LPO reduces latency from around 100ns to less than 10ns. That not only improves the responsiveness of distributed AI training and inference but also permits infrastructure to deliver more useful compute work per watt-hour.

Crucially, LPO modules interoperate with standard FRO, allowing operators to maintain hybrid environments and pursue a phased transition within the same network fabric.

Key Technical Considerations for DSP-less Architectures

Despite the immense benefits of LPO technology, its successful deployment depends on a few, key hardware conditions.

1. Host platform dependency: Because the optic no longer carries a DSP, the host switch must do more. That means stronger SerDes performance is essential. Many early deployments are associated with newer switch platforms such as those built around Tomahawk 5-class architectures.
2. Distance and signal integrity: Without re-timing inside the optic, LPO is generally better suited to short-reach links, often under 500 meters. Additionally, LPO can be less forgiving of signal deterioration introduced by cable quality or board layout variations.

A Roadmap to 1.6T and Sustainable Growth

As AI clusters move toward 1.6T interconnect speeds, pressure to improve efficiency will continue to rise. A growing number of industry analysts see LPO as a practical near-term deployment solution that will enjoy significant, consistent growth through 2033.

In the AI era, improving energy per inference means looking beyond the GPU and examining every watt consumed across the network path. Linear Pluggable Optics are not a universal replacement for traditional re-timed modules. But in short-reach, high-density AI environments, they can play an important role in reducing power draw, lowering thermal load, and improving compute efficiency. For those building at scale, the optical layer is no longer a background consideration. It is now the next fundamental frontier in the quest for more sustainable AI.

Subscribe to the Broadband Communities newsletter!