> ## Content Index > Fetch the complete content index at: https://www.insidedeeptech.com/llms.txt > Use this file to discover other available public pages before exploring further. # Linear Pluggable Optics (LPO): A Full Guide to DSP-Less Transceivers When CPO Is Too Early - URL: https://www.insidedeeptech.com/linear-pluggable-optics-lpo-full-guide/ - Published: 2026-09-15T06:18:47.000Z - Updated: 2026-09-15T06:18:46.000Z - Description: LPO strips the module DSP and keeps the pluggable cage: LPO MSA 100G-DR-LPO for 800G over 500 m SMF, OIF linear interfaces, Cisco and Marvell shipping signals, and when DSP optics or CPO still win. - Author: Irina Peskisheva - Tags: AI, Photonics, Hardware Every watt spent retiming light inside a pluggable cage is a watt that never reaches a GPU. At 800G and the early 1.6T generation, that DSP tax is large enough that operators are buying a middle path: keep the QSFP-DD or OSFP form factor, strip the module DSP, and push equalization back onto the host SerDes. That path is **linear pluggable optics** (**LPO**). This guide covers what LPO actually is, how it differs from DSP pluggables, LRO, CPO, AOC, and AEC, where it sits in 800G and 1.6T AI scale-out, who ships and which standards matter, and the honest limits that keep DSP modules and CPO on the same BOM spreadsheet. ## Key Takeaways - LPO removes the module DSP; the host SerDes equalizes the analog link. - LPO MSA 100G-DR-LPO (March 2025) covers up to 800G over 500 m SMF. - OIF cites up to \~50% module power savings versus retimed optics. - Cisco's OFC 2024 dual-switch demo showed \~700 W system savings with LPO. - Use LPO when host SerDes, channel loss, and pair qualification are ready; keep DSP or wait for CPO otherwise. 📌 Scope note: LPO here means linear-drive (DSP-less or reduced-DSP) pluggable optics in QSFP / QSFP-DD / OSFP cages. It is not co-packaged optics and not an active electrical cable. Claims below are dated as of September 2026 and cite the LPO MSA 100G-DR-LPO specification, OIF CEI-112G-Linear demos, Cisco Silicon One, and Marvell primary product materials. ## What LPO actually is A conventional 800G pluggable module contains lasers or modulators, photodiodes, drivers, TIAs, **and** a DSP that retimes, equalizes, and often converts between host and optical domains. That DSP is the reliability and interoperability engine. It is also a power, latency, and cost line item that scales with port count. LPO removes that retimer from the module. The host ASIC SerDes drives a linear electrical interface into the module. The module amplifies and converts with linear driver and TIA paths. Equalization, FEC encode/decode, and retiming stay on the host. The [LPO MSA 100G-DR-LPO specification](https://www.lpo-msa.org/files/live/sites/lpomsa/files/specs/LPO%5FMSA%5FSpecification%5Fv1p2%5Ffinal.pdf?ref=insidedeeptech.com) (Revision 1.0, 19 March 2025) states this cleanly: the module conveys analog signals between host and fiber; the host performs FEC, retiming, and DAC/ADC functions. Optical interfaces run 53.125 GBd PAM4 per lane on single-mode fiber with a required operating range of 0.5 m to 500 m. Lane counts of 1, 2, 4, or 8 give 100G, 200G, 400G, and 800G Ethernet-class links in form factors such as QSFP, QSFP-DD, and OSFP. Here's why that matters. You keep the operational habit of a field-replaceable pluggable. You discard a large share of module watts and DSP latency. You accept a harder host and channel problem in exchange. ## How LPO differs from DSP pluggables, LRO, CPO, AOC, and AEC Do not collapse every "lower power optics" slide into one column. | Option | Where DSP lives | Form factor | Typical job | Honest catch | | ------------- | --------------------- | -------------------- | --------------------------------- | ------------------------------------- | | DSP pluggable | Inside module | QSFP-DD / OSFP | Broad multi-vendor 800G/1.6T | Highest module power | | LPO | Host SerDes only | Same cages | Short–mid AI scale-out | Host/channel pair-sensitive | | LRO / TRO | TX retimed; RX linear | Same cages | Interop compromise | Half the savings; mixed BER tradeoffs | | CPO | Package / engine path | On-package / NPO | Switch density & power | Serviceability and supply early | | AOC | Usually in ends | Fixed cable | Rack/row optics | Not a field-swappable PMD SKU | | AEC | Retimer in copper | OSFP / QSFP-DD cable | Short copper when optics overkill | Meters, not halls | Cisco's optics sessions describe **LRO** (linear receive optics, also called TRO) as a half-linear design: a DSP on transmit, linear path on receive. It is pitched as an interoperability and thermal compromise. Full LPO is more aggressive on module power and more demanding on host SerDes quality end to end. Co-packaged optics moves the optical engine onto the switch or accelerator package. That is a different BOM fight from LPO. LPO keeps the cage. CPO removes or shrinks the cage path. Inside Deep Tech's [CPO full guide](https://www.insidedeeptech.com/co-packaged-optics-cpo-full-guide/) covers that package transition; this article stays on the pluggable linear path. Active optical cables and [active electrical cables](https://www.insidedeeptech.com/active-electrical-cables-aec-ai-data-centers-full-guide/) solve adjacent reach problems. AOCs are usually fixed assemblies with electronics in the ends. AECs put retimers in copper for short AI scale-out when fiber is overkill. Neither is LPO. 💡 Inside Deep Tech's take: treat LPO as a host-qualified media type, not as a drop-in SKU swap for every DSP 800G port. If the SerDes, PCB loss, and CMIS calibration path are not designed for linear drive, buying "LPO" modules will not buy you DSP-class multi-vendor luck. ## Standards stack: OIF, IEEE, and the LPO MSA Three layers matter. First, [OIF CEI-112G-Linear](https://www.oiforum.com/technical-work/hot-topics/common-electrical-interface-cei-112g-2/?ref=insidedeeptech.com) defines the chip-to-module linear electrical interface for \~112G PAM4 without a DSP/SerDes inside the optical module. OIF's public interoperability materials (including the ECOC 2024 CEI-112G-Linear demo package) list advantages operators care about: lower module power (cited up to about 50% versus traditional retimed modules), lower latency, protocol-agnostic data path, retained sideband manageability, and demonstrated reaches including 100 m MMF and 500 m SMF across multi-vendor LPO, RTLR, and DSP mixes. Second, IEEE 802.3 supplies the Ethernet framing, FEC, and related PMD heritage. The LPO MSA specification explicitly builds on IEEE Std 802.3-2022, 802.3ck-2022, 802.3df-2024, and draft 802.3dj work, plus OIF CEI-112G-LINEAR-PAM4. Third, the [LPO MSA](https://www.lpo-msa.org/news/lpo-msa-announces-release-of-specification-for-linear-pluggable-optica?ref=insidedeeptech.com) published the completed 100G-DR-LPO single-mode specification at OFC 2025 (announced 25 March 2025). The group reported 50-plus member companies and a February 2025 interoperability event with margins above required link performance across switches, NICs, and modules. Co-chair Andreas Bechtolsheim stated the goal plainly: let large datacenter operators deploy 800G LPO modules in volume with confidence. The MSA's next stated target is 200G-per-lane linear work with OIF and IEEE. On the electrical budget, 100G-DR-LPO extends maximum host-to-module channel loss versus the base OIF Linear IA: from 13 dB to 16 dB at 26.56 GHz (half baud). Hosts must implement RS(544,514) FEC. Modules and hosts exchange insertion-loss and CTLE calibration data over CMIS Versatile Control Set so the linear EQ can track real port loss. That is not optional trivia. It is how a DSP-less module stays honest across a chassis. [Co-Packaged Optics: A Full Guide to CPO, LPO, and Why AI Switches Are Absorbing the TransceiverCPO versus LPO versus pluggables, Broadcom Bailly, NVIDIA Quantum-X and Spectrum-X, and an honest timeline for 1.6T and 3.2T.![](https://www.insidedeeptech.com/favicon.ico)Inside Deep Tech](https://www.insidedeeptech.com/co-packaged-optics-cpo-full-guide/) ## Where LPO sits in 800G and 1.6T AI scale-out AI clusters burn interconnect power in three places: scale-up GPU fabrics, scale-out Ethernet or InfiniBand leaves and spines, and the optics that stitch them. LPO is aimed at the pluggable Ethernet (and Ethernet-like) scale-out and short-reach fabric ports where DSP pluggables dominate today and CPO is still early. At 100G per lane, an 800G OSFP or QSFP-DD LPO module is the practical object. At 200G per lane, the same architecture targets 1.6T modules, but the MSA's 100G-per-lane specification is the completed, interoperability-tested baseline as of March 2025\. 200G-per-lane linear remains an active standards and silicon push, not a finished twin of 100G-DR-LPO. Reach is deliberately short-to-medium by DSP DR/FR standards: 500 m SMF is the 100G-DR-LPO required operating range. That fits leaf-to-spine and many AI pod geometries. It does not replace campus DCI or long FR/LR DSP optics. Power is the buying argument. OIF materials cite up to roughly 50% module power reduction versus retimed modules. Cisco published a concrete system comparison at OFC 2024: two identical Silicon One 51.2 Tbps, 64-port G200-based switches, one fully populated with retimed optics and one with LPO, both running full traffic, with an overall power reduction of about 700 W for the LPO system. That is a hyperscaler-grade existence proof, not a universal calculator for every leaf. Latency falls because the module DSP pipeline disappears. For tightly synchronized training collectives, that is a secondary but real benefit next to watts. ## Who ships: hosts, silicon, and module ecosystem Hosts need SerDes that can drive linear optics. [Cisco Silicon One G200](https://www.cisco.com/c/en/us/solutions/collateral/silicon-one/silicon-one-g200-ds.html?ref=insidedeeptech.com) data sheets explicitly list Linear Pluggable Optics among supported physical connectivity options on its 512×112G SerDes. Cisco's broader Silicon One messaging ties the same 112G ADC SerDes generation to LPO and CPO channel classes. Switch ASICs with strong 100G PAM4 SerDes (for example Broadcom's Tomahawk 5 class devices with dense 106G PAM4 I/O) are the other side of that host story. The point is not a single SKU endorsement. It is that LPO only works when the host electrical eye and EQ were designed for a linear module load. On the module silicon side, [Marvell announced general availability](https://www.marvell.com/company/newsroom/marvell-introduces-1-6-tbps-lpo-chipset.html?ref=insidedeeptech.com) on 10 December 2024 of a 200G-per-lane TIA and laser-driver chipset aimed at 800G and 1.6T linear-drive pluggable optics for short-reach scale-up fabrics. That is the complementary bet to Marvell's DSP portfolio: linear analog front ends when the host already owns equalization. Module vendors in the published LPO MSA member roster span the AI optics supply base (examples include AOI, Coherent, Hisense, Lumentum, Source Photonics, Semtech, and MACOM, among fifty-plus members). Multi-vendor demos at OIF events matter more than logos: the ECOC 2024 materials explicitly call out multi-vendor LPO modules, LPO-to-RTLR, and LPO-to-DSP interoperability cases. Silicon photonics shows up as one implementation path inside many of those engines, not as a synonym for LPO. See the [silicon photonics full guide](https://www.insidedeeptech.com/silicon-photonics-ai-data-centers-full-guide/) for the chip stack; LPO is the module architecture that may or may not use SiPh. [Optical Circuit Switching (OCS): A Full Guide to Photonic Fabrics When Packet Switches Hit the WallMEMS and related OCS for spines and AI pods: the reconfigurable photonic fabric layer after electrical packet spines.![](https://www.insidedeeptech.com/favicon.ico)Inside Deep Tech](https://www.insidedeeptech.com/optical-circuit-switching-ocs-ai-data-centers-full-guide/) ## Honest limits: when DSP pluggables or CPO still win LPO fails the wrong buyer for predictable reasons. - Host SerDes and PCB loss are not linear-ready. If die-to-module loss and EQ calibration are outside the MSA envelope, margins collapse. - You need casual multi-vendor mix-and-match without pairwise qualification. DSP pluggables still win that operational model. - Reach or optical budget exceeds \~500 m SMF class needs, or you need FR/LR/ZR coherent spans. - Telemetry, loopbacks, and FEC partitioning that lived inside the module DSP are still mandatory for your ops model. - Your real density bottleneck is package-level SerDes power on a 51.2T/102.4T switch, where CPO (not LPO) is the architecture bet. - 200G-per-lane linear is still harder than 100G-per-lane. Do not treat 1.6T LPO as a carbon copy of shipping 800G LPO practice. On the flip side, DSP pluggables still win when you are filling heterogeneous fleets, buying through distributors that must mix vendors, or running links where host channels were designed only for retimed modules. CPO still wins when the switch or GPU package itself is the thermal and SerDes wall, and when you accept early supply, service, and fiber-attach complexity for package-level watts. LPO also is not a substitute for [NVLink, InfiniBand, or UALink](https://www.insidedeeptech.com/nvlink-infiniband-ualink-ai-gpu-interconnect-full-guide/) scale-up fabrics, and it is not optical circuit switching. [OCS](https://www.insidedeeptech.com/optical-circuit-switching-ocs-ai-data-centers-full-guide/) rewires photonic paths between systems. LPO is still a point-to-point pluggable PMD. ⚖️ Inside Deep Tech's take: LPO is the commercial bridge while CPO ramps. Buy it where host silicon already advertises linear optics, cable plants fit 500 m SMF class, and you can qualify host-module pairs the way hyperscalers already qualify optics. Do not use LPO as a slogan to defer either DSP reliability engineering or a real CPO roadmap. ## What to do Monday If you run or buy AI networking gear: - Ask switch and NIC vendors for explicit LPO / linear-drive SerDes support, host loss budgets, and CMIS-VCS calibration behavior, not just "800G optics." - Read the LPO MSA 100G-DR-LPO PDF and the relevant OIF CEI-112G-Linear materials before accepting a power-savings slide. - Qualify modules as host-paired SKUs. Budget lab time for BER/FEC margin across temperature and worst-case ports. - Keep DSP pluggables on the BOM for long reach, mixed fleets, and ports that were never designed for linear drive. - Track 200G-per-lane linear and CPO as separate roadmaps. They answer different questions on the same rack drawing. The industry did not invent LPO because DSP optics failed. It invented LPO because DSP watts became visible at AI port counts. The operators who win will treat linear pluggables as a qualified media class with clear failure modes, not as a free lunch. [Read the State of Data Centers 2026 report](https://www.insidedeeptech.com/the-state-of-data-centers-2026/) --- ## FAQ #### What is linear pluggable optics (LPO)? LPO is a pluggable optical module architecture that removes the module DSP/retimer and uses a linear electrical interface to the host. The host SerDes and FEC do the equalization and error correction while the module handles linear electro-optic conversion in QSFP, QSFP-DD, or OSFP form factors. #### How is LPO different from CPO? LPO keeps a field-replaceable pluggable cage and strips the module DSP. CPO moves optical engines onto the switch or accelerator package to cut package-level SerDes and pluggable power. Both chase watts; only CPO changes the service and packaging model. #### How far can 800G LPO reach? The LPO MSA 100G-DR-LPO specification requires 0.5 m to 500 m over single-mode fiber for 100G-per-lane links up to 800G (8 lanes). Longer campus or DCI spans still belong to DSP FR/LR or coherent optics. #### How much power does LPO save versus DSP pluggables? OIF interoperability materials cite up to about 50% module power savings versus traditional retimed modules. Cisco's OFC 2024 comparison of two Silicon One G200 51.2T switches reported roughly 700 W lower system power when the switch was fully populated with LPO instead of retimed optics. Treat both as published benchmarks, not a universal calculator. #### What is LRO or TRO compared with LPO? LRO (also called TRO) keeps a retimer on the transmit path and a linear receive path. It is a half-linear compromise aimed at interoperability and thermal tradeoffs. Full LPO is linear in both directions and more dependent on host SerDes quality. #### Which standards define LPO? OIF CEI-112G-Linear defines the linear chip-to-module electrical interface. The LPO MSA 100G-DR-LPO specification (March 2025) defines interoperable 100G-per-lane single-mode optical and electrical requirements up to 800G, building on IEEE 802.3 and OIF work. #### When should a cluster keep DSP pluggables instead of LPO? Keep DSP modules when you need broad mix-and-match interoperability, longer reach, ports without linear-ready SerDes, or ops workflows that still depend on module-side DSP telemetry and loopbacks. --- [Explore more Inside Deep Tech AI infrastructure guides](https://www.insidedeeptech.com/)