Semtech backs linear pluggable optics for AI data centres

Semtech has outlined its position on linear pluggable optics for AI and data centre interconnects, describing an approach that removes the DSP retimer from optical modules.

The explanation sets out how linear pluggable optics, or LPO, differs from the conventional retimed optical links used in data centres. In Semtech's description, a traditional module relies on a digital signal processor inside the transceiver to retime, reshape and re-amplify signals on both transmit and receive paths before handing traffic back to the host switch.

That architecture has helped modules cope with degraded electrical signals and reflections across the link, but it comes with a power cost. According to Semtech, a 1.6T DR-8 retimed module in current data centre deployments typically consumes 23 to 25 watts.

Under an LPO design, the DSP retimer is removed from the module. Instead, the module uses analogue elements including a continuous time linear equaliser on the transmit side, together with a driver and modulator, while the receive side uses a transimpedance amplifier and, in some cases, a linear equaliser.

This shifts signal conditioning. Rather than the module handling the main equalisation work, the host switch SerDes takes on that task, while the transceiver focuses on analogue amplification and optical conversion.

Power trade-off

Semtech argued that the main attraction is lower power use in dense AI and data centre networks. It said LPO modules for 200G links are targeting about 10 watts, less than half the draw of today's retimed modules and below the roughly 16 watts associated with Retimed Transmitter Linear Receiver designs, also known as Linear Receive Optics.

The company also offered a system-level illustration. In a 512-port configuration, which it linked to the current Tomahawk-6 generation, LPO could save about 500W at module level and as much as 1kW across the system when cooling is included. At 1024 ports, those savings would double.

The trade-off is that LPO is not intended as a universal replacement for retimed optics. Deployments with high-loss electrical channels, significant reflections or a need for maximum transmit equalisation flexibility may still be better suited to RTLR or fully retimed modules.

Not passive

Semtech also sought to address what it described as a common source of confusion around the term. LPO is not a passive optical module because it still includes active analogue components and requires configuration and calibration, even though it does not contain a DSP retimer with clock recovery.

That distinction matters as network operators weigh how much complexity should sit in the transceiver and how much should remain on the switch. By shifting more of the equalisation burden back to the host SerDes, LPO places greater emphasis on calibration and link design at system level.

This does not necessarily mean operators need new switch ASICs. Instead, Semtech argued, the host SerDes must be calibrated correctly to provide the equalisation that LPO modules expect, making the issue as much about configuration and methodology as hardware replacement.

Standards work

The standards picture remains in development. Semtech said the electrical side is being shaped by the OIF CEI 224 linear specification, which targets 22 dB die-to-die port loss, while IEEE sets the optical performance requirements for 200G links.

It added that the Linear Pluggable Optics Multi-Source Agreement Group is bringing those strands together into a single framework. Semtech is a founding member of the group.

The debate over LPO comes as operators look for ways to contain power use in AI infrastructure while increasing bandwidth between switches, accelerators and storage. Optical interconnect choices have become a more visible design issue as rack-scale and cluster-scale systems grow in port count and cooling demands rise alongside them.

Semtech's account suggests LPO is moving from an experimental concept to a design option for deployment planning, even though standards work is still active. The company argued that the standards framework and calibration methodology are sufficiently mature to support deployment planning today.