TECHNOLOGY EXPLAINER
NPO and LPO
— before getting rid of the box: empty it out, or move it closer
CPO, which puts the optics inside the same package as the chip, is not the only way to cut the power of optical modules. LPO keeps the pluggable form exactly as it is but takes the DSP (digital signal processing chip) out of the module. NPO mounts the optical engine right next to the chip in a socket, so that it can be removed if it fails. Both save power by shifting the burden somewhere else. Where that burden lands is where the materials problems are.
- What NPO and LPO are (the short version)
- Four approaches side by side — where the DSP sits and how far the signal travels as electricity
- Our calculation: how much of a switch's power the optical modules were taking
- Inside LPO — what happens when you take the DSP out
- A materials engineer's view (1): in LPO, board loss becomes part of the optical link
- Inside NPO — optical engines mounted in sockets
- A materials engineer's view (2): solder or socket — optical parts that cannot take reflow
- How it all compares with CPO, on one page
- What is still hard
- Glossary / References / Claim-to-source audit
Sourced = stated in a specification or a public document from a standards body (link given)
Our calculation = a figure this article derived, with the assumptions spelled out
Not yet confirmed = still in development or planned, with nothing yet settled or delivered
Structural readings and materials-design interpretations are marked separately as Commentary.
1. What NPO and LPO are (the short version)
- LPO (Linear Pluggable Optics): the LPO MSA defines LPO as "a low-power pluggable module that does not incorporate a DSP chip"Sourced. Signal processing is left to the chip in the switch or NIC
- NPO (Near-Package Optics): an OIF document uses the term "socketed, “near-package optics” (NPO)" for an arrangement in which a packaged ASIC and optical engines are mounted in sockets on a common substrate, so they can be removed during assembly or reworkSourced
- The shared aim: both set out to lower the power, cost and latency of optical modules (optical engines), but they go about it differently. LPO removes parts; NPO shortens the distance the signal travels electrically (our framing)
The box that plugs into the front panel is covered in our explainer on optical transceivers, and putting the optical engine in the same package as the ASIC in our Co-packaged Optics explainer. This article deals with the two options in between. Why the loss of electrical wiring grows with distance is covered in our explainer on optical interconnects and SerDes, and modulation formats such as PAM4 in our explainer on optical modulation and modulation formats.
2. Four approaches side by side — where the DSP sits and how far the signal travels as electricity
The differences between pluggable, LPO, NPO and CPO come down to two questions: (1) where is the DSP that cleans up the signal, and (2) how far does the signal travel as electricity between the chip and the point where it becomes light (our framing).
Seen this way, LPO keeps the distance and removes parts, NPO keeps the parts and shortens the distance (while staying removable), and CPO makes the distance as short as possible and gives up removability. None of them is simply the right answer; the choice depends on whether power, serviceability or compatibility with existing equipment comes first (our commentary).
3. Our calculation: how much of a switch's power the optical modules were taking
Why has the power of optical modules become such an issue? The OIF Co-Packaging Framework Document (2022) tabulates the power of what was then a typical 1U data-centre switch (12.8 Tb/s capacity, 32 ports of 400G QSFP-DD)Sourced.
| Item (as listed in the OIF table) | Power |
|---|---|
| 12.8 Tb/s switch chip (NPU) | 432 W |
| 32 ports of 400G optical modules (12 W each) | 384 W |
| Typical power consumption (including fans, power supplies and typical optical modules) | 900 W |
| Maximum power consumption (maximum traffic, highest ambient temperature, maximum fan speed) | 1,500 W |
All Sourced (OIF "Co-Packaging Framework Document", Table 5 [Ref. 5]). These were given as a "typical" example as of 2022 and are not measurements of a particular product.
The same document expects the energy efficiency of dense co-packaged optical engines to be around 5 to 15 pJ/bit, and lists ≤15 pJ/b in the requirements column for data-centre switch applicationsSourced. Set against the 30 pJ/bit above (our calculation), the problem being posed is clear: cut the energy spent on optical conversion to half or less. Note, though, that the OIF figure comes with conditions — it includes the engine-side electrical interface, CDR, optical components and light source but excludes the switch side — so its accounting boundary does not match that of the calculation in this section (our commentary).
4. Inside LPO — what happens when you take the DSP out
In March 2025 the LPO MSA published "100G-DR-LPO", a specification for 100 Gb/s per laneSourced. Its main points are as follows.
| Item | What the 100G-DR-LPO specification (Rev 1.0) says |
|---|---|
| Optical signal | 100 Gb/s per lane, 53.125 GBd PAM4, nominal wavelength 1310 nm, 0.5 to 500 m over single-mode fibre |
| Link configurations | 1, 2, 4 or 8 lanes (n00G-DRn-LPO). QSFP, QSFP-DD and OSFP are given as examples of form factor (the specification is form-factor agnostic) |
| Division of roles | The module transmits and receives the analogue signal as is. Error correction (FEC), retiming and D/A and A/D conversion are done on the host side |
| Error rate | Raw bit error ratio of 2.0 × 10−4 or better, assumed to be corrected to 10−13 or better with RS(544,514) FEC |
| Electrical loss budget | Extends the maximum channel loss of OIF CEI-112G-LINEAR-PAM4 by 3 dB, from 13 dB to 16 dB (measured at 26.56 GHz). Insertion loss from ASIC die to module die is assumed to be 20 dB or less |
| Compensation | The host may include non-linear compensation (NLC) to linearise the optical transmit waveform |
All Sourced (LPO MSA "100G-DR-LPO Specification Revision 1.0" [Ref. 1]). The optical interface is described as "similar but not identical" to 400GBASE-DR4.
The LPO MSA FAQ explains the idea like this. Modern switch and NIC ASICs have "extremely capable transmitters and receivers", but when they are connected to DSP-based modules that capability goes unused. With LPO, the whole link becomes one continuous linear channel, so the ASIC's full equalisation capability can be brought to bearSourced.
Of this "linear amplified" type, OIF points out that the power saved by omitting the DSP comes at a price: with modulation such as PAM4, the drive signal must be linear, without amplitude compression, and the ASIC's transmitters and receivers must equalise the whole link, from SerDes transmitter to receiverSourced. In November 2025 OIF also published an implementation agreement for a "half-retimed" type (EEI-112G-RTLR), which retimes only the transmit side and keeps the receive side linear, explaining that it removes the need for a receive DSP in the optical moduleSourced.
5. A materials engineer's view (1): in LPO, board loss becomes part of the optical link
In a DSP-based module, an electrical signal degraded by the host board is retimed once at the module's input. Board loss was a self-contained problem confined to the short stretch between host and module.
Not so with LPO. As the LPO MSA FAQ puts it, the link becomes "a seamless linear channel"Sourced. The transmitting ASIC's package, the host board, the module board, the optical section, and the far-end board and package — all of their degradation adds up and arrives at the receiving ASIC (our commentary).
That is why the specification recommends keeping the losses of the ASIC package, the host board and the module board (including the AC-coupling capacitors) within the ranges shown in its figureSourced. The LPO MSA FAQ likewise says that systems supporting LPO need "capable ASICs and well-designed transmission lines"Sourced.
- Dielectric loss in the board: the lower the dissipation factor of the resin, the lower the loss at high frequency
- Conductor loss: the smoother the copper foil, the lower the loss from the skin effect
- Variation: with no DSP to retime the signal, lot-to-lot variation in loss eats straight into link margin
In LPO, low-loss board materials and smooth copper foil become materials for cutting optical-module power. The work of the DSP taken out of the module is being shared between the ASIC and the board material — that, seen from the materials side, is the essence of this approach (our commentary). One more thing: the non-linear compensation (NLC) in the specification is a mechanism premised on the optical transmitter's output waveform not being linearSourced. That means the linearity of the modulator or laser, a physical property in its own right, also governs link performance (our commentary).
6. Inside NPO — optical engines mounted in sockets
The OIF Co-Packaging Framework Document calls an arrangement of an ASIC and optical engines (or electrical engines) on a high-performance substrate a Co-Packaging Assembly (CPA), and envisages parts being either soldered or socketed onto itSourced. NPO is the case where a packaged ASIC and optical engines are socketed onto a common substrateSourced.
For socketed parts, OIF says "a mechanical compression mechanism" is needed to press the socket interposer into electrical contact, and that this mechanism commonly takes the form of bolts at the four corners of the socketSourced. It also notes that the clamping hardware consumes board area and can limit packing densitySourced.
OIF says that a 3.2 Tb/s optical engine at 10 pJ/bit would consume 32 W, and that if it measures around 20 × 20 mm² the heat density would be 8 W/cm², "much higher than what the optical industry is accustomed to"Sourced. Check: 3.2 × 1012 bit/s × 10 × 10−12 J/bit = 32 W, and 32 W / (2 cm × 2 cm) = 8 W/cm²Our calculation. Whether NPO or CPO, heat sources at this density end up lined up right next to the chip (our commentary).
7. A materials engineer's view (2): solder or socket — optical parts that cannot take reflow
OIF sets out in a table the trade-offs between fixing optical engines by solder reflow and mounting them in socketsSourced.
| Aspect | Solder reflow | Socket |
|---|---|---|
| Demands on the parts | Parts must survive reflow. Makes surface mounting of optical engines possible | Simply placed. Can extend to applications that cannot go through reflow |
| Electrical performance | Can be close to optimal | Can be very good |
| Footprint | Densest | Needs clamping hardware and a retention mechanism |
| Rework | Limited, and becomes yield loss | Possible, though access in the field is limited |
| Many engines / complex assemblies | Integration yield loss when many are mounted | Makes complex configurations possible |
| Thermal | Conventional methods can be used | Mounted together with the retention hardware |
All Sourced (summarised by this article from Table 4 of the OIF Co-Packaging Framework Document [Ref. 5]).
The first row of the table is the heaviest. If an optical engine is to be soldered, it has to go through the temperature of a reflow oven. Yet an optical engine contains optical fibre, the adhesive that positions and holds the fibre, lenses, light-carrying polymers and more. As noted in our Co-packaged Optics explainer, AGC lists reflow compatibility among the strengths of its optical waveguides [Ref. 8]Sourced.
- Adhesives: if they soften or expand in the heat of reflow, the fibre shifts by micrometres and coupling loss rises
- Light-carrying polymers: if heat yellows or degrades them, transmittance falls
- Many parts: OIF writes that mounting many engines leads to integration yield lossSourced
A socket is a way of sidestepping this problem through process rather than solving it through materials. The optical engine is added afterwards without going through reflow, and can be removed if it is bad (our commentary).
But that creates a different materials problem. A compression contact clamped by bolts at four corners has to keep pressing hundreds or thousands of contacts with even force. Board warpage, stress relaxation in the springs, wear of the contact plating — problems long familiar from semiconductor test sockets and CPU sockets are carried straight over into mounting optical engines. And right beside the clamp sit a heat source of the 8 W/cm² class and a fibre exit that hates contamination (our commentary).
OIF also notes that for optical connectors behind the front panel, in a "single-mate" situation where they are only inspected after assembly, a larger loss allocation must be budgeted than for ordinary connectors that can be cleaned and re-matedSourced. A removable structure only makes sense together with materials and processes that keep things clean (our commentary).
8. How it all compares with CPO, on one page
| Aspect | Conventional pluggable | LPO | NPO | CPO |
|---|---|---|---|---|
| Where the optical engine sits | Front panel | Front panel | Next to the ASIC (shared board) | In the same package as the ASIC |
| DSP in the module | Yes | No (the ASIC compensates) | Depends on configuration | Depends on configuration |
| Replacement | Unplug and replace | Unplug and replace | Can be removed from the socket | Difficult, except for the laser |
| Electrical run | Long | Long | Short | Shortest |
| Standardisation (within what this article confirmed) | QSFP-DD and OSFP MSAs | LPO MSA specification (100 Gb/s per lane) | Defined in an OIF document. No standalone specification confirmed | OIF implementation agreements (see our Co-packaged Optics explainer) |
"Where the optical engine sits", "Replacement" and "DSP" are this article's framing based on the LPO MSA [Refs. 1 and 2], OIF [Ref. 5] and our Co-packaged Optics explainer. Whether NPO and CPO carry a DSP depends on the electrical interface type (Fig. 3), hence "depends on configuration". "Long" and "short" for the electrical run are relative comparisons, not numbers.
9. What is still hard
(1) LPO is choosy about its partner
On interconnecting LPO modules with DSP-based modules, the LPO MSA FAQ says it is possible, but is regarded as an "engineered link" and falls outside the scope of the MSA specificationSourced. LPO also presupposes ASICs with strong equalisation and well-designed transmission linesSourced. It is not a matter of plugging it into existing equipment and saving power (our commentary).
(2) LPO at 200 Gb/s per lane is still to come
Having published 100G-DR-LPO at 100 Gb/s per lane (March 2025) and 400G-FR4-LPO (September 2025, at least 500 m over one pair of SMF), the LPO MSA has announced that it has begun work on linear implementations at 200 Gb/s per laneSourced. Doubling the lane rate makes the board loss discussed in Section 5 harsher still. Whether LPO will work at 200 Gb/s per lane was not settled as of this article's researchNot yet confirmed.
(3) No standalone NPO standard could be found
NPO is defined in the OIF Co-Packaging Framework Document as one form of arrangementSourced, but an NPO-specific implementation agreement defining socket geometry or optical-engine outline could not be confirmed within the scope of this article's research. Nor is the state of NPO commercialisation or volume production described here, since nothing could be confirmed from primary sourcesNot yet confirmed.
(4) Read power-saving figures together with the basis of comparison
Press coverage and commentary cite all sorts of percentages for the power savings of LPO, but the LPO MSA's own published material stops at the qualitative "Significantly lower power consumption"Sourced. This article does not give reduction percentages whose basis of comparison cannot be confirmed.
- LPO is a pluggable optical module with no built-in DSP. Compensation is handled by the switch or NIC ASICSourced
- NPO is an arrangement in which the ASIC and optical engines are socketed onto a shared board. They can be removedSourced
- In OIF's 2022 example, optics accounted for about 47% of the combined power of chip and optical modules, and a 400G module worked out at about 30 pJ/bitOur calculation
- LPO's electrical loss ceiling is 16 dB. At 20 dB from die to die, amplitude falls to about a tenthOur calculation
- In LPO, board-material loss becomes part of the optical link; in NPO, avoiding reflow puts compression contacts and cleanliness to the test (our commentary)
- LPO at 200 Gb/s per lane and a standalone NPO standard could not be confirmed as of this article's researchNot yet confirmed
10. Glossary
- LPO
- Linear Pluggable Optics. A pluggable optical module with no built-in DSP that converts the signal to light linearly.
- LRO
- Linear Receive Optics. A half-retimed configuration in which only the receive side is linear.
- NPO
- Near-Package Optics. An arrangement in which the ASIC and optical engines are socketed onto a shared board.
- CPO
- Co-Packaged Optics. Mounting the optical engines in the same package as the ASIC.
- DSP
- Digital signal processor. Digitises the electrical signal to correct its waveform and retime it.
- Retiming
- Recovering the timing of a degraded signal and sending it on as a clean signal.
- Equalisation
- Processing that restores the high-frequency content lost in the channel, bringing the waveform back towards its original shape.
- SerDes
- A circuit that converts a chip's parallel data into high-speed serial signals for transmission and reception.
- Insertion loss (dB)
- How much a signal weakens on its way through a channel. At 20 dB, amplitude falls to a tenth.
- FEC
- Forward error correction. Redundant code is added so that errors can be corrected at the receiver.
- CPA
- Co-Packaging Assembly. An assembly of the ASIC and engines on a shared substrate (OIF's term).
- Reflow
- The heating process that melts solder to join parts to a board.
- LGA
- Land grid array. A form of connection that presses flat pads together instead of using solder balls.
11. References (primary sources)
- LPO MSA "100G-DR-LPO Revision 1.0 — Specification for 100 Gb/s per Lane Linear Pluggable Optics Single-Mode Optical Fiber Transmission", March 2025 (PDF) — lpo-msa.org
- LPO MSA "Frequently Asked Questions" — lpo-msa.org
- LPO MSA "LPO MSA Announces Release of Specification for Linear Pluggable Optical Modules", 25 March 2025 — lpo-msa.org
- LPO MSA "LPO MSA Announces Release of 400G-FR4-LPO Specification for Linear Pluggable Optical Modules", 25 September 2025 — lpo-msa.org
- OIF "Co-Packaging Framework Document (OIF-Co-Packaging-FD-01.0)", 3 February 2022 (PDF) — oiforum.com
- OIF "OIF Publishes Implementation Agreement for 112 Gb/s Retimed Transmitter Linear Receiver (RTLR) Electrical and Optical Interface Advancing Energy Efficiency", 18 November 2025 — oiforum.com
- LPO MSA Official website (Overview / LPO Benefits) — lpo-msa.org
- AGC "CES 2026: semiconductor solutions" (Japanese-language page) — agc.com
12. Claim-to-source audit
| Claim in the text | Basis | Label |
|---|---|---|
| That 100G-DR-LPO is 100 Gb/s per lane, 53.125 GBd PAM4, nominally 1310 nm, 0.5 to 500 m over SMF, with 1, 2, 4 or 8 lanes. That it is a form-factor-agnostic specification giving QSFP, QSFP-DD and OSFP as examples. That the module transmits and receives the analogue signal while FEC, retiming and D/A and A/D conversion are done by the host. That the optical interface is "similar but not identical" to 400GBASE-DR4. Raw BER 2.0 × 10−4, 10−13 after RS(544,514) FEC. That the maximum channel loss of OIF CEI-112G-LINEAR-PAM4 is extended by 3 dB from 13 dB to 16 dB (26.56 GHz). That insertion loss from ASIC die to module die is assumed to be 20 dB or less, with recommended loss ranges for each segment. The provision for host-side non-linear compensation (NLC). Revision 1.0 dated 19 March 2025 | LPO MSA 100G-DR-LPO specification Rev 1.0Reference 1 https://www.lpo-msa.org/files/live/sites/lpomsa/files/specs/LPO_MSA_Specification_v1p0_final.pdf | Sourced |
| That LPO is "a low-power pluggable module that does not incorporate a DSP chip". That ASICs have "extremely capable transmitters and receivers" and that with LPO the link becomes "a seamless linear channel" in which the ASIC's equalisation can be used. That systems need "capable ASICs and well-designed transmission lines". That interconnecting with DSP-based modules is an "engineered link" outside the scope of the MSA. That semi-linear configurations such as LRO are also covered. That it builds on OIF CEI-112G-LINEAR-PAM4 | LPO MSA FAQReference 2 https://www.lpo-msa.org/home/faqs.html | Sourced |
| That completion of the 100G-DR-LPO specification was announced on 25 March 2025. That the MSA has 50 member companies. That it is looking towards 200 Gb/s per lane | LPO MSA announcement (25 March 2025)Reference 3 https://www.lpo-msa.org/news/lpo-msa-announces-release-of-specification-for-linear-pluggable-optica | Sourced |
| That the 400G-FR4-LPO specification (at least 500 m over one pair of SMF, four wavelengths) was announced on 25 September 2025, together with the start of work on linear implementations at 200 Gb/s per lane | LPO MSA announcement (25 September 2025)Reference 4 https://www.lpo-msa.org/news/lpo-msa-announces-release-of-400g-fr4-lpo-specification-for-linear-plu | Sourced |
| That an arrangement with the ASIC and optical engines socketed onto a common substrate is called "socketed, “near-package optics” (NPO)". That both soldered and socketed mounting are envisaged for the CPA. The 12.8 Tb/s switch power in Table 5 (NPU 432 W, 32 × 12 W = 384 W for 400G optical modules, 900 W typical, 1,500 W maximum). The expected energy efficiency of 5 to 15 pJ/bit and ≤15 pJ/b for data-centre switch applications (including the engine-side electrical interface, CDR, optical components and light source, excluding the switch side). The four types (retimed, linear amplified, half-retimed, direct drive), and the need in the linear amplified type for a linear drive signal and for the ASIC to equalise the whole link. The assumption of about 50 mm on the board for the retimed type. 32 W for a 3.2T engine at 10 pJ/bit, 8 W/cm² at 20 × 20 mm², and "much higher than what the optical industry is accustomed to". The socket compression mechanism and the four-corner bolts. The solder-reflow versus socket comparison in Table 4. The need to budget a larger loss allocation for "single-mate" optical connectors | OIF Co-Packaging Framework Document (3 February 2022)Reference 5 https://www.oiforum.com/wp-content/uploads/OIF-Co-Packaging-FD-01.0.pdf | Sourced |
| That OIF published the EEI-112G-RTLR (retimed transmitter, linear receiver) implementation agreement on 18 November 2025, explaining that it removes the need for a receive DSP in optical modules | OIF announcement (18 November 2025)Reference 6 https://www.oiforum.com/oif-publishes-implementation-agreement-for-112-gb-s-retimed-transmitter-linear-receiver-rtlr-electrical-and-optical-interface-advancing-energy-efficiency/ | Sourced |
| That the LPO MSA website states the benefit qualitatively as "Significantly lower power consumption", obtained by removing DSP functions | LPO MSA websiteReference 7 https://www.lpo-msa.org/ | Sourced |
| That AGC lists high transmittance (O band and C band) and reflow compatibility and durability against high-power lasers among the strengths of its optical waveguides | AGC "CES 2026: semiconductor solutions"Reference 8 https://www.agc.com/ces/semiconductor.html | Sourced |
| 384 / (432 + 384) = 47%, 432 / 816 = 53%, 12 W / 400 Gb/s = 30 pJ/bit. Conversion of 16 dB, 13 dB and 20 dB into amplitude ratios (about 16%, about 22% and 10%) and power ratios (about 2.5% and 1%). The check 3.2 Tb/s × 10 pJ/bit = 32 W and 32 W / 4 cm² = 8 W/cm² | Our calculation. OIF's Table 5 is a "typical" example as of 2022, and fan and power-supply power are excluded. The accounting boundary differs from OIF's pJ/b figure | Our calculation |
| Whether LPO will work at 200 Gb/s per lane. The state of NPO commercialisation and volume production | The LPO MSA has only announced the start of work, so feasibility remains an outlook. For NPO, no primary source could be confirmed within the scope of this article's research | Not yet confirmed |
| The existence of an NPO-specific implementation agreement (socket geometry, engine outline) | Not stated in this article, because none could be confirmed within the scope of its research | Commentary |
| Organising the four approaches by where the DSP sits and how far the signal runs electrically. The framing of LPO as removing parts and NPO as shortening distance. Mapping the linear amplified type to LPO and the half-retimed type to LRO and RTLR. The reading that in LPO board-material loss becomes part of the optical link (dielectric loss, conductor loss, variation). The reading that sockets sidestep the reflow-resistance problem through process, and the issues of compression contacts, stress relaxation and cleanliness. The framing of the comparison table in Section 8 | This article's own framing and commentary based on public documents. Not views expressed by any of the bodies | Commentary |
| Percentages for LPO power savings cited in press coverage and elsewhere | Not stated in this article, because the basis of comparison could not be confirmed from primary sources | Commentary |
| That Figs. 1, 3 and 5 are explanatory drawings, that Figs. 2 and 4 are drawings that include our calculation, and that the hero image and Fig. 6 are AI-generated images | A note by this article | Commentary |
Last updated 25 September 2026. Sources are limited to primary material (the LPO MSA specification, website and announcements, and the OIF Framework Document and announcements). Because the article includes structural readings and materials-design interpretations, those are marked as Commentary and kept separate from sourced fact. Specific percentages for LPO power savings, an NPO-specific standard, the state of NPO commercialisation and the feasibility of LPO at 200 Gb/s per lane are not covered, because they could not be confirmed in published primary sources. The OIF Framework Document dates from 2022, and its power example gives "typical" values as of that time. All figures are for explanation. Figs. 1, 3 and 5 are vector drawings, Figs. 2 and 4 are vector drawings that include our calculation, and the hero image and Fig. 6 are AI-generated images; none of them shows the structure or dimensions of a real product.