+86 166 2042 3014 WhatsApp: +86 166 2042 3014 [email protected] Mon–Fri, 09:00–18:00 CST
Get Quote

112G PAM4 Interconnect Decoded: LPO Linear Drive vs AEC Retimer - a Pitfall Guide

As AI compute clusters march toward 800G and 1.6T, GPU Scale-Up interconnect hits a physical wall. Traditional DSP-based modules bring heavy power and cooling pressure inside dense racks, while pure copper DAC cannot reach beyond in-rack distances.

In 2026, mature LPO (Linear Pluggable Optics) and AEC (Active Electrical Cable) Retimer technologies are redefining the AI compute fabric. This guide skips the basics and compares both from physics limits and engineering pitfalls.

Pain point 1: the physical limit of 112G PAM4 and the skin effect

At 112G PAM4 per lane, insertion loss in copper explodes. Due to the skin effect, high-frequency current only flows on the conductor surface, squeezing effective passive DAC reach down to roughly 2-3 meters - which wrecks cable management and airflow inside AI racks with dense backplanes.

Breakthrough 1: AEC Retimers rescue the signal

AEC embeds low-power Retimer chips that receive, equalize and re-build the attenuated signal in real time, safely extending 112G PAM4 reach from 3m to 7m or even 10m.

Engineering pitfall: Retimer quality varies widely. Weak CDR (clock-data recovery) in low-end chips triggers BER avalanches on long runs. When selecting AECs, demand full-channel eye-diagram and BER test reports at target length.

Breakthrough 2: LPO removes the DSP entirely

LPO deletes the module's highest-power component - the DSP - keeping only the TIA and linear driver. The switch ASIC's integrated SerDes absorbs the equalization work. Power drops ~50% and latency falls from microseconds to nanoseconds, exactly what GPU collective communication needs.

Engineering pitfall: LPO shifts signal-integrity responsibility to the switch side. Only deploy LPO on switch platforms it has been jointly validated with, and keep total channel loss within the module's specified budget.

Selection cheat sheet

  • In-rack, under 3m, cost-first: passive DAC.
  • Rack-to-rack, 3-10m: AEC with quality Retimers.
  • Longer reaches or rack breakout to optics: LPO (validated platforms) or standard optics with DSP.

OTSLink supplies validated AEC/LPO-ready interconnects and pre-shipping platform compatibility testing for mainstream NVIDIA GPU and switch ecosystems.