A single 800G DSP-based optical transceiver can draw up to 17 watts. Fill a 64-port 51.2T switch with them, and the optics alone consume over 1,000W before the switch ASIC contributes a single watt.
For hyperscale and AI data centers, optical power consumption is becoming an increasingly important design consideration. As network speeds continue to increase, 800G LPO (Linear Pluggable Optics) has emerged as one approach to reducing power and latency.
By removing the DSP-based retiming function from the optical module and relying more heavily on the host switch ASIC’s SerDes, LPO can significantly reduce module power and processing latency. However, these benefits come with trade-offs in reach, signal integrity, and interoperability.
Linear Pluggable Optics (LPO) is a non-retimed optical module architecture designed to reduce power and latency by removing the DSP-based retiming function from the module.
In a conventional retimed optical module, the electrical signal from the host switch is processed by a DSP before driving the optical components. The DSP can perform functions such as equalization, signal conditioning, retiming, and gearbox processing, depending on the module architecture.
In an LPO design, much of this signal conditioning is handled by the host switch ASIC’s SerDes. The optical module uses relatively simple linear electrical components, typically including a linear driver on the transmit side and a TIA on the receive side.
This architecture reduces the amount of active digital processing inside the module. As a result, LPO can offer lower power consumption and lower module latency, but it also places greater demands on the host SerDes and the overall link design
The clearest way to understand 800G LPO is to measure it against the DSP-based modules it replaces. Three numbers tell most of the story.
Depending on the implementation, conventional 800G DSP-based modules can operate in the mid-teens of watts, while LPO designs may reduce module power to around 8W or below in some implementations.
The savings compound at the switch level. Where 64 retimed modules drew over 1,000W, the same switch populated with LPO modules draws closer to 500W. Cisco demonstrated a roughly 700W system-level reduction on a 51.2T Silicon One switch comparing fully retimed and fully LPO ports. Across a data center with hundreds of switches, that translates directly into lower cooling loads and a smaller power draw.
Every DSP processing cycle adds fixed delay. Retimed optics add roughly 100 nanoseconds of latency through the module. LPO removes that step entirely, reducing module latency to under 10ns in many implementations, with some sources citing up to a 90% reduction.
Retimed optical modules can introduce additional processing latency, often on the order of tens of nanoseconds depending on the DSP architecture. Because LPO eliminates module-level DSP retiming, its module contribution to latency can be significantly lower, with some implementations achieving single-digit-nanosecond latency.
For AI training and inference workloads where latency directly affects job completion time, this is a meaningful gain. It is not a rounding error.
The DSP is expensive silicon. Removing it cuts module cost even though the linear driver and TIA require somewhat higher-grade analog components. The net effect is a lower bill of materials and a lower cost per bit, which matters when you are buying modules by the thousand.
| Attribute | DSP-based 800G | LPO-based 800G |
| Module power | Typically higher | Typically lower |
| Module latency | Higher | Lower |
| Reach | From short reach to long reach, depending on optics | Primarily short-reach applications |
| Interoperability | Generally more mature | More platform-dependent |
| Module cost | Generally higher | Potentially lower |
The trade-off is real. LPO is not a free lunch. It buys lower power, lower latency, and lower cost at the price of shorter reach and tighter host dependency, which is why it is not a universal replacement for retimed optics.
LPO sits at one end of a spectrum of low-power optics. Three related terms come up constantly, and confusing them leads to poor decisions.
LRO(Linear Receive Optics) is commonly used to describe an architecture in which the transmit path retains DSP or retiming while the receive path uses a more linear architecture. Some vendors use terms such as half-retimed or HALO for related architectures, but terminology can vary.
LRO offers better link stability and lower host dependency than LPO, at the cost of giving back some of the power savings, roughly half of what LPO achieves. It is often described as the pragmatic middle ground.
CPO co-packages the optical engine directly with the switch ASIC, removing the pluggable module entirely. NPO (Near Package Optics) places the optics adjacent to the ASIC but still on the board. They address the same broader challenge—reducing the power and electrical-loss penalty between high-speed switch ASICs and optical engines—but use fundamentally different packaging architectures.
HYBRID approaches retain partial DSP functionality to strike a balance between power and signal integrity. Vendors describe these slightly differently, but the theme is the same: keep just enough retiming to improve interoperability without paying the full power cost.
The key takeaway is that LPO is one point on a deliberate trade-off curve, not a category of its own.
| Variant | DSP / Retiming | Power | Reach | Interoperability |
| LPO | Fully removed | Lowest | Shortest | Most challenging |
| LRO / Half-retimed | Tx only | Medium | Medium | Better |
| HYBRID | Partial | Medium | Medium | Better |
| CPO / NPO | Varies | Lowest | Short | Non-pluggable |
For most operators evaluating 800G today, the real decision is LPO versus LRO versus staying fully retimed. CPO remains a longer-term architectural bet.
800G LPO does not invent a new form factor. It rides on the same pluggable packaging already deployed in high-speed networks, which is a significant part of its appeal.
800G LPO can be implemented in established pluggable form factors rather than requiring a completely new module package.
OSFP and QSFP-DD800 are two important form factors for 800G networking. OSFP provides a larger thermal envelope and is widely used in high-density switch and AI networking platforms. QSFP-DD800 offers a more compact form factor and can be attractive where backward compatibility and port density are important.
The choice between OSFP and QSFP-DD800 is primarily platform-dependent. Network designers should consider switch or accelerator compatibility, thermal requirements, connector configuration, power limits, and the specific LPO implementation rather than treating one form factor as universally superior.
LPO is generally best suited to short-reach optical links. Without a module-level DSP to perform retiming and advanced signal conditioning, the link depends more heavily on the host SerDes and the characteristics of the optical channel.
Actual reach depends on the optical configuration, fiber type, host SerDes capability, FEC, and module implementation. Common short-reach 800G LPO configurations may include SR8, DR8, and 2×FR4 architectures, with some implementations targeting approximately 100 m for multimode links and several hundred meters for single-mode links.
LPO should therefore not be defined by a single maximum distance. For longer-reach applications, especially multi-kilometer links, DSP-based optics generally remain more practical because the DSP can provide the signal processing and compensation required for demanding optical channels.
A useful rule of thumb is to consider LPO primarily for short-reach AI and data-center fabrics, while DSP-based optics remain the safer choice for longer-distance, metro, and DCI applications. LRO can provide an intermediate option where lower power is desired without moving completely to a linear architecture.
The hardest part of adopting 800G LPO is not the hardware. It is proving the links work.
Because LPO removes the module’s DSP, performance depends heavily on the switch ASIC’s SerDes. A fully linear LPO implementation pairs best with advanced SerDes silicon such as Broadcom’s Tomahawk 5 class, while a segmented-compensation approach suits mid-range platforms like Marvell’s XLS800. The module and the switch are now a matched pair, not independent commodities.
This is a genuine change for network teams. With retimed optics, you could swap modules across vendors freely. With LPO, each switch-module combination should be qualified together.
The OIF’s CEI-112G-Linear-PAM4 specification, which underpins linear-mode operation, is not yet fully finalized, and vendor CTLE and equalization implementations differ. The LPO MSA is working to close these gaps, but multi-vendor interoperability still requires validation through programs like UNH-IOL.
Before deploying 800G LPO, network operators should qualify the complete link rather than evaluating the optical module in isolation.
At minimum, testing should cover:
Because LPO shifts more signal-processing responsibility to the host, interoperability testing should be performed using the actual switch and module combinations planned for deployment.
LPO is not a universal upgrade. It is a deliberate trade-off that fits some environments and not others.
LPO earns its place in short-reach, high-density fabrics where power and latency dominate. AI and HPC clusters, GPU interconnects, and top-of-rack to leaf-spine links are the natural home. When a single rack row holds thousands of links, halving module power is a strategic advantage.
The market is moving. 800G LPO shipments entered early commercial deployment in 2026, with hyperscalers such as Meta and Amazon placing orders and the broader 800G module market projected to exceed $4.8 billion in 2026, driven by AI cluster interconnect.
Stay fully retimed for any link beyond about 500 meters, for multi-vendor environments where you cannot qualify every pairing, and for long-haul or DCI where dispersion management is non-negotiable. If your team lacks the cycles for a formal qualification phase, LPO’s interoperability overhead may outweigh its power savings.
Looking ahead, the industry is already signaling that at 200G-per-lane and 1.6T speeds, LRO may be more deployable than LPO. Nearly every vendor showing 1.6T LPO at recent industry events also showed an LRO variant. The pattern to watch is that as speeds climb, the half-retimed middle ground gains ground. At 1.6T and 200G-per-lane speeds, the trade-off between power efficiency, signal integrity, and interoperability becomes even more challenging. As a result, LRO and other partially retimed architectures are attracting attention as potential alternatives to fully linear designs.
800G LPO offers a compelling way to reduce optical-module power and latency in high-density AI and data-center networks. By removing module-level DSP retiming and relying more heavily on the host SerDes, LPO can deliver significant efficiency improvements.
However, LPO is not a universal replacement for DSP-based optics. Its benefits come with trade-offs in reach, signal integrity, and interoperability. For short-reach AI fabrics, LPO can be an attractive option, while longer-distance and more complex links may still benefit from fully retimed DSP optics. LRO can provide a practical middle ground when lower power and greater interoperability are both important.
The right choice ultimately depends on the network topology, host ASIC, reach requirements, power budget, and interoperability targets. Rather than asking whether LPO is better than DSP, network designers should determine where each architecture fits best within the overall network.
800G LPO (Linear Pluggable Optics) is an optical module architecture that removes module-level DSP retiming and relies more heavily on the host switch ASIC’s SerDes for signal conditioning. This can reduce power consumption and latency.
Power consumption depends on the module design and implementation. Many 800G LPO modules target significantly lower power than conventional DSP-based modules, with some designs operating around 8W.
The main difference is signal processing. DSP-based modules perform retiming and signal conditioning inside the optical module, while LPO shifts more of this processing to the host switch ASIC. LPO can reduce power and latency but has greater host and link-design dependencies.
800G LPO is primarily designed for short-reach applications. Actual reach depends on the optical configuration, fiber type, host SerDes capability, FEC, and module implementation. Some implementations target approximately 100 m for multimode links and several hundred meters for single-mode links.
No. LPO interoperability is more platform-dependent than conventional retimed optics because the host switch SerDes performs more of the signal conditioning. The specific module, switch ASIC, SerDes configuration, and firmware should be qualified together.
LPO removes module-level DSP retiming, while LRO (Linear Receive Optics) typically retains more processing on the transmit side and uses a more linear receive path. LRO can provide a compromise between LPO’s low power and the interoperability of fully retimed optics.
Yes. LPO is particularly attractive for short-reach, high-density AI and HPC networks where optical power, thermal management, and latency are important. It is commonly considered for GPU interconnects and high-bandwidth data-center fabrics.
Not completely. LPO is well suited to certain short-reach, power-sensitive applications, but DSP-based optics remain important for longer-reach and more demanding links. The choice depends on reach, host platform, power budget, signal integrity, and interoperability requirements.
