Understand how Linear Pluggable Optics (LPO) reduce 800G optical power and latency by removing the retiming DSP, and what that change means for host SerDes, signal integrity, interoperability, thermals and deployment.
800G optical links are becoming an increasingly important part of AI and high-performance computing networks, but increasing bandwidth also creates pressure on power, cooling and front-panel thermal density.
Traditional DSP-based optical transceivers perform significant signal conditioning and retiming inside the module. Linear Pluggable Optics (LPO) take a different approach. By removing the retiming DSP, LPO shifts more responsibility for signal integrity to the switch ASIC, host SerDes and equalization engines.
That architectural change can significantly reduce module power, heat and optical processing latency. The tradeoff is that electrical channel quality, host SerDes performance, firmware, FEC behavior and platform interoperability become more important to the performance of the complete link.
For an example of how Axiom implements LPO as an integrated 800G system, review the Axiom 800G Switch + LPO Solution .
What is Linear Pluggable Optics (LPO)?
Linear Pluggable Optics remove the retiming DSP found in conventional high-speed optical modules and rely more heavily on the switch and NIC SerDes to maintain signal integrity.
Removing that processing stage can lower module power, heat and optical processing latency. The tradeoff is greater dependence on the electrical channel, host SerDes, equalization, firmware, FEC and complete platform design.
What Changes When the Retiming DSP Is Removed?
The fundamental difference between LPO and traditional retimed optics is where signal conditioning occurs.
Traditional DSP-Based Optics
DSP-based modules perform digital equalization and signal conditioning inside the transceiver, helping compensate for channel loss, reflections, crosstalk, host impairments and PCB losses.
Linear Pluggable Optics (LPO)
LPO removes the retiming DSP and shifts more signal-processing responsibility to the switch SerDes, NIC SerDes and ASIC equalization engines.
The Engineering Consequence
With LPO, the transceiver can no longer be evaluated as independently from the host platform as a DSP-retimed optic. Switch ASIC capability, SerDes performance, PCB channel quality, equalization, NIC behavior, FEC and firmware become more important parts of the link.
Why LPO Is Emerging With 800G AI Networks
AI infrastructure creates a strong use case for lower-power optical architectures because a large fabric can contain hundreds of switches and thousands of optical links.
More GPU Traffic
Distributed AI training moves large amounts of east-west traffic between accelerator nodes and across the leaf-spine fabric.
More Optical Links
Higher port counts increase the number of modules packed into the switch front panel, making per-module power increasingly important.
More Heat to Remove
Optical-module power becomes heat at the front panel, where airflow and heat-sink space are constrained by high port density.
Repeated GPU Communication
Operations such as All Reduce, All Gather, Broadcast and Reduce Scatter repeatedly move data across the network during distributed training.
Where the LPO Power Savings Come From
Removing the retiming DSP can substantially reduce the power consumed by each 800G optical module.
Axiom engineering modeled an 800G DSP optic at 18W and an Axiom 800G LPO at 9W. That represents a 9W reduction per module, or approximately 50 percent in the modeled comparison.
These figures are modeled examples based on an 18W DSP module and 9W Axiom LPO. Actual module power and savings depend on the specific optic, reach, host platform, operating conditions and deployment.
Why Scale Changes the Power Conversation
A few watts per module may appear small when evaluating a single transceiver. Across thousands of continuously operating links, that difference can become tens of kilowatts of direct electrical load while also affecting cooling requirements and front-panel thermal density.
Review the complete cluster-scale power and TCO model in the Axiom 800G Switch + LPO Solution .
LPO vs. DSP-Based 800G Optics
LPO is not simply a lower-power version of the same architecture. Removing the retiming DSP changes where equalization happens and shifts more responsibility to the complete system.
| Factor | Linear Pluggable Optics (LPO) | DSP / Retimed Optics |
|---|---|---|
| Retiming DSP | No | Yes |
| Module Power | Lower | Higher |
| Optical Processing Latency | Lower | Higher due to DSP processing |
| Heat Dissipation | Lower | Higher |
| Host SerDes Dependence | Higher | Lower |
| Electrical Channel Sensitivity | Higher | Lower |
| Channel Margin | Tighter | Greater |
| Interoperability Requirements | More platform-dependent | More mature across varied hosts |
| Tolerance to Poor Channel Quality | Lower | Higher |
| Supply Dependency | No retiming DSP dependency | Requires retiming DSP |
| Best Fit | Controlled, high-density AI and HPC environments | Broad or heterogeneous deployments |
The Tradeoff
LPO trades some of the channel tolerance and host independence of DSP-retimed optics for lower power, reduced heat and lower optical processing latency. That makes system design and validation more important, not less.
LPO Changes the System Boundary
With a DSP-retimed module, the transceiver performs a significant amount of signal conditioning internally. With LPO, link performance depends more heavily on the complete electrical and optical path.
If an LPO link becomes marginal or unstable, engineers may need to investigate more than the optical module itself. Potential contributors include:
- Switch SerDes behavior
- TX and RX equalization
- Lane margin
- PCB and connector channel loss
- Host firmware
- FEC configuration and behavior
- NIC SerDes performance
- Optical power
- Module thermals
- Fiber-path conditions
Where LPO Makes the Most Sense
LPO is strongest when the host platform, optics, electrical path and validation process can be engineered together.
AI Training Fabrics
High-density GPU clusters with sustained east-west traffic where optical power, latency and rack density are major design considerations.
HPC Networks
Scientific computing and message-passing environments where low communication latency and power-efficient high-speed links are important.
Dense RoCEv2 Fabrics
Controlled leaf-spine environments with high port counts, modern SerDes, predictable topology and strong telemetry.
Where DSP-Based Optics May Still Be the Better Fit
- Heterogeneous brownfield environments
- Platforms with uncertain SerDes capability
- Electrical channels with limited margin
- Deployments requiring broad third-party interoperability
- Applications where maximum reach flexibility is more important than power
- Environments without system-level qualification capability
What Engineers Must Validate Before Deploying LPO
Link-up is only the beginning. LPO qualification should prove the complete host-to-optic path.
OMA
TDECQ
Equalization
Lane margin
Post-FEC
Error trends
Traffic stability
Latency
ECN
RDMA
GPU traffic
Stability
Optical & Electrical Performance
Evaluate TX and RX optical power, OMA, TDECQ, receiver sensitivity, jitter, PAM4 behavior, equalization and electrical lane margin.
Error Performance
Axiom's qualification framework targets pre-FEC BER better than 1E-4, post-FEC BER better than 1E-12, and zero uncorrectable errors during long-duration testing.
SerDes & Firmware
Verify switch ASIC, NIC, firmware, equalization settings, FEC behavior, port configuration and electrical-channel compatibility.
Real Network Load
Test IMIX, large-message traffic, storage traffic, sustained Ethernet load and representative AI traffic rather than relying on basic link-up.
Congestion Behavior
Validate PFC, ECN, RDMA throughput, buffers, queue stability, retransmissions, congestion recovery and deadlock behavior.
Collective Communications
Qualification should include representative collective operations such as All Reduce, All Gather, Broadcast and Reduce Scatter.
Long-Duration Testing Matters
Axiom's engineering framework includes BER and stability testing across 24-hour, 72-hour, 7-day and 30-day periods. Extended testing helps expose marginal electrical channels, thermal behavior and intermittent errors that may not appear during a short link-up test.
Telemetry Becomes More Important With LPO
Because LPO shifts more signal-integrity responsibility toward the host, operators need visibility beyond whether the interface is simply up or down.
Physical-Layer Telemetry
BER, FEC behavior, optical power, temperature, lane margin, PAM4 performance and link stability help identify marginal conditions before a hard failure occurs.
Congestion Telemetry
ECN marking rates, PFC pause statistics, buffer occupancy, port utilization and congestion hot spots show how the broader Ethernet fabric is behaving.
Application Behavior
RDMA throughput, GPU communication behavior, retransmissions and workload stability help connect network conditions to actual AI application performance.
Troubleshooting LPO Requires a System-Level Approach
With LPO, an unstable link does not automatically mean the optical module has failed. The root cause may sit elsewhere in the host-to-host path.
Engineers may need to isolate the switch SerDes, equalization settings, electrical channel, optics, fiber path, remote host, NIC, firmware and FEC behavior one variable at a time.
An LPO Link Has a Larger Failure Domain
The more signal-conditioning responsibility that moves out of the module, the more important it becomes to diagnose the complete system rather than immediately replacing the optic.
For a production troubleshooting workflow, review the Optical Transceiver Failure Analysis Guide .
How Axiom Approaches Linear Pluggable Optics
Axiom approaches LPO as a system architecture rather than a standalone optical-module replacement.
Axiom's 800G implementation pairs Axiom Linear Pluggable Optics with an Axiom Broadcom Tomahawk 6-based network switch. The Axiom LPO is equalized for the Axiom switch platform, allowing the host electrical path and optical module to be engineered together.
The switch platform provides 102.4 Tbps of switching capacity, 64 x 800G OSFP front-panel interfaces, RoCEv2 support, PFC, ECN, MMU optimization and network telemetry for high-density AI and HPC Ethernet fabrics.
See How Axiom Combines the Switch and LPO as One 800G Architecture
Review the Axiom Broadcom Tomahawk 6-based switch, Axiom 800G LPO, modeled power savings, RoCEv2 capabilities and integrated AI networking architecture.
Related Technical Resources
Axiom 800G Switch + LPO
See how Axiom combines a TH6-based switch with 800G LPO equalized for the platform.
800G Optics Guide
Review 800G form factors, optical architectures, use cases and deployment considerations.
800G Transceiver Validation Guide
Review FEC, optical power, thermals, traffic stability and production qualification.
AI Networking Infrastructure Guide
Plan AI networking across switching, optics, NICs, interconnects, power and migration requirements.
Linear Pluggable Optics FAQs
Linear Pluggable Optics, or LPO, remove the retiming DSP found in conventional high-speed optical modules and rely more heavily on switch and NIC SerDes capabilities, host equalization and the overall electrical channel to maintain signal integrity.
A major source of LPO power reduction is the removal of the retiming DSP from the optical module. In Axiom's modeled comparison, an 18W DSP-based 800G optic is compared with a 9W Axiom 800G LPO, representing approximately 50 percent lower optical-module power in that scenario.
LPO removes the retiming DSP processing stage from the optical module. This reduces optical signal-processing delay compared with architectures that perform digital retiming within the transceiver.
LPO reduces power and processing latency but becomes more dependent on the host platform. Switch SerDes performance, electrical channel quality, equalization, NIC behavior, FEC, firmware and interoperability become more important to overall link performance.
LPO requires a host platform with sufficiently capable SerDes and equalization for the target electrical channel and optical module. The complete switch, optics, firmware and NIC path should be qualified together rather than assuming any 800G host will behave identically.
LPO is well suited to controlled, high-density AI and HPC networks where low optical power, lower processing latency, strong host SerDes and system-level validation are priorities. It is particularly relevant to large 800G leaf-spine and RoCEv2 fabrics.
DSP-retimed optics may remain a better fit in heterogeneous environments, electrical channels with limited margin, deployments requiring broad host interoperability, or applications where maximum channel tolerance and reach flexibility are more important than module power.
Engineers should evaluate optical power, PAM4 signal integrity, host SerDes, equalization, lane margin, BER, FEC, firmware, Ethernet traffic, RoCEv2 behavior, PFC, ECN, thermals, interoperability, telemetry, AI workloads and long-duration stability.
Axiom pairs its 800G Linear Pluggable Optics with an Axiom Broadcom Tomahawk 6-based network switch. The Axiom LPO is equalized for the Axiom switch platform, allowing the host electrical path and optical module to be engineered as an integrated 800G solution.
Evaluating LPO for an 800G AI or HPC Network?
Talk with Axiom about Linear Pluggable Optics, host SerDes requirements, 800G switching, power modeling, RoCEv2, interoperability and system-level qualification for your target environment.
