Pluggable, LRO, LPO, CPO: Where the DSP Went

Scope

This article compares the four architectures competing to carry datacenter traffic at 800G and 1.6T: the fully retimed pluggable module, LRO (transmit-retimed, linear receive), LPO (fully linear), and co-packaged optics. It covers what the module DSP does and why removing it is tempting, per-bit power with a consistent worked example, the interoperability and serviceability consequences, and the state of deployment in late 2026. Coherent modules for longer reaches are a different ladder and out of scope; the light-source side of CPO has its own article.

What the module DSP actually does

Inside a conventional pluggable module, the DSP terminates the electrical link from the host in both directions: it recovers the clock, equalizes the host-side channel, regenerates the data, and drives the optics from a clean slate; on receive it does the same for the optical channel before re-launching the signal into the host board. The module therefore presents two independently specified interfaces, electrical and optical, and conformance at each can be certified alone; the transmit eye that TDECQ scores is the DSP's product, not the host's. The cost of this tidiness is power and latency: the DSP is roughly half of a module's dissipation, and at the 1.6T generation a full-DSP module runs on the order of 20 W.

Note what the DSP does not do: the forward error correction for these links terminates in the switch ASICs at the ends, not in the module. The ladder below is about retiming and equalization, with the FEC as the common safety net under all of it.

The ladder, step by step

Pluggable, fully retimed. The incumbent. Both directions retimed in the module; any compliant host interoperates with any compliant module; the faceplate cage makes every port a field-replaceable unit. The 800G generation is mature volume product and 1.6T OSFP modules are in mass production on eight 200G lanes.

LRO, transmit retimed. The receive-side DSP is dropped; the transmit retimer stays, so the module still answers alone for its launched eye while the host serdes absorbs the received one. Roughly half the DSP saving at a fraction of the interoperability risk; the LRO entry covers the asymmetry in detail.

LPO, fully linear. No DSP at all: the module amplifies linearly in both directions and the host serdes does everything. The saving is the whole DSP; the price is that the optical link's conformance now depends on the specific host-module-host combination, which is why multi-vendor LPO interoperation is hard at 100G per lane and harder at 200G, and why deployment has concentrated in single-operator environments that qualify combinations themselves. The LPO entry covers the architecture.

CPO, co-packaged. The optics leave the faceplate entirely and sit beside the switch ASIC, shortening the electrical channel from tens of centimeters to millimeters; the channel becomes short enough that the heavy per-port electrical equalization goes away, and the lasers move out to external laser sources. This is no longer a module choice but a switch-design choice: the optics are bought with the switch, and the serviceable units are the fiber connectors and the laser shelf.

Power per bit

Consistent numbers, using the representative 1.6T figures from the Datacenter Link Budget Explorer: a full-DSP pluggable at 22 W is 13.8 pJ/bit; the LPO variant at the tool's published-figure multiplier lands near 12.6 W, or 7.9 pJ/bit; the CPO figure near 10.2 W, or 6.4 pJ/bit. LRO sits between the first two, recovering about half of the DSP's share. Vendor claims run further: Nvidia quotes about 3.5 times better power efficiency for its co-packaged switch ports against pluggables, a network-level figure that includes eliminating the host-side retimer stages, not only the module. All such numbers are architecture-and-generation specific; the stable conclusion is the ordering, and the observation that at a million ports the difference between 13.8 and 6.4 pJ/bit is several megawatts of continuous draw, which is the number procurement actually argues about.

Interoperability, serviceability, risk

The ladder trades three currencies at once. Interoperability is spent as the DSP is removed: full retiming means any-to-any; LRO preserves it on the transmit side where certification lives; LPO moves qualification onto the buyer; CPO makes the question internal to one vendor's switch. Serviceability is spent at the last step: a pluggable of any kind is a thirty-second field swap, while a co-packaged engine is not, which is why CPO designs work to make the lasers, the dominant failure item, external and swappable, and why early CPO deployments are in operators with the engineering staff to accept the model. Risk concentration moves in the same direction: a failed pluggable takes one port; a failed co-packaged engine, depending on design, can take many.

These currencies explain the deployment pattern better than the power numbers do. Hyperscalers running uniform, self-qualified networks adopt LPO and CPO first, because they can pay in qualification effort and service-model change for power they value at fleet scale. Everyone else moves a rung at a time, and LRO exists precisely to be the rung that costs least to step onto.

Who ships what, late 2026

As of late 2026: 800G full-DSP pluggables are the volume workhorse and 1.6T full-DSP modules are in mass production. LPO ships in niches, mostly intra-rack and single-vendor; LRO demonstrations at 1.6T (Eoptolink among them, from OFC 2025 onward) have matured into products positioned as the low-risk saving. In CPO, Nvidia's Quantum-X InfiniBand switches shipped in early 2026 with Spectrum-X Ethernet following, and Broadcom's Tomahawk 6 Davisson at 102.4 Tb/s is shipping; both use external laser sources. The next inflection is the 400G-per-lane serdes generation, where every step of the ladder gets harder and the linear options' interoperability problem tightens further; first-generation 400G-per-lane optical DSPs were announced at OFC 2026.

The architecture question is not which step wins but how much of the market each step holds at each generation, and for how long the FEC budget keeps absorbing what the removed DSPs used to fix. Dated claims in this article carry their dates deliberately; the physics of the trade does not expire, the market shares do.

References: Juniper Networks, "Types of optics: LPO and LRO" (800G documentation); LPO MSA materials (lpo-msa.org); OIF ELSFP Implementation Agreement (2023); Nvidia GTC 2025 co-packaged switch announcements and subsequent IEEE Spectrum coverage; Broadcom Tomahawk 6 Davisson materials (2026); OFC 2026 show reporting on 1.6T mass production and 400G-per-lane DSPs. The Datacenter Link Budget Explorer compares the architectures' power at fleet scale, and the PAM4 Eye & TDECQ Explorer shows the transmit conformance that the retimer question turns on.