Axiom integrates a 102.4 Tbps Tomahawk 6 switch platform with 800G LPO technology, delivering lower power consumption, reduced latency, and improved network efficiency.
The core of the solution is the pairing of two Axiom products: the Axiom 800G network switch and Axiom 800G Linear Pluggable Optics (LPO). The switch platform is based on Broadcom Tomahawk 6 and provides 102.4 Tbps of switching capacity with 64 x 800G OSFP interfaces. Axiom's 800G LPO is equalized specifically for the Axiom switch platform.
By removing the retiming DSP used in conventional 800G optical modules, the integrated switch + LPO architecture is designed to reduce optical power consumption, reduce latency, eliminate reliance on the retiming DSP supply chain, and lower acquisition and maintenance costs.
For AI and HPC environments, those advantages become increasingly important as hundreds or thousands of high-speed links are deployed across leaf-spine fabrics. Power, cooling, rack density, latency, optics availability, and operating cost all scale with the number of network connections.
For broader planning across 400G, 800G, 1.6T, optics, cables, adapters, and AI fabric design, review the AI Networking Infrastructure Guide . For additional optical deployment guidance, review the 800G Optics Guide .
Short Answer
The Axiom 800G solution is built around an Axiom switch and Axiom LPO engineered to work together.
Axiom's Broadcom Tomahawk 6-based 102.4 Tbps switch provides 64 x 800G OSFP interfaces designed for use with Axiom 800G Linear Pluggable Optics (LPO). The LPO is equalized for the Axiom switch platform, creating an integrated 800G link solution focused on lower power consumption, lower latency, reduced dependence on retiming DSP components, and lower acquisition and maintenance costs.
64 x 800G OSFP interfaces
RoCEv2 support
PFC + ECN + fabric telemetry
No retiming DSP
Lower optical power
Reduced processing latency
Key Advantages of the Axiom Switch + LPO Solution
Reduced Power Consumption
Removing the retiming DSP from the optical module can significantly reduce optical power per 800G link, lowering rack-level power and cooling requirements as deployments scale.
Reduced Latency
LPO removes DSP retiming and processing stages from the optical data path, reducing optical processing latency for latency-sensitive AI and HPC communication.
No Retiming DSP Dependency
Axiom 800G LPO does not use the retiming DSP found in conventional DSP-based optics, making the architecture immune to shortages of that DSP component.
Lower Acquisition & Maintenance Costs
The simpler optical architecture can lower acquisition cost while reduced power, cooling demand, and DSP dependency help improve operating and maintenance economics.
Axiom 800G Network Switch
The switch is an Axiom product and a core part of the integrated 800G solution.
Axiom's 800G network switch is based on the Broadcom Tomahawk 6 architecture and provides 102.4 Tbps of switching capacity. The platform includes 64 x 800G OSFP front-panel interfaces and is designed to operate with Axiom's 800G Linear Pluggable Optics (LPO).
Axiom LPO Is Equalized for the Axiom Switch Platform
This pairing is a key differentiator of the solution. LPO depends more heavily on the electrical performance and equalization capabilities of the host switch than conventional DSP-retimed optics. Axiom's 800G LPO is equalized for the Axiom Broadcom Tomahawk 6-based switch, allowing the switch and optical module to be treated as an integrated 800G link architecture.
| Platform Specification | Axiom Architecture |
|---|---|
| Product | Axiom 800G Network Switch |
| Switch ASIC | Broadcom Tomahawk 6 (TH6) |
| Switching Capacity | 102.4 Tbps |
| Front-Panel Ports | 64 x 800G OSFP |
| Front-Panel Bandwidth | 51.2 Tbps across 64 x 800G interfaces |
| Optical Pairing | Axiom 800G Linear Pluggable Optics (LPO) equalized for the Axiom switch platform |
| RDMA Transport | RoCEv2 |
| Congestion Management | PFC + ECN + MMU / shared-buffer optimization |
| Provisioning | Zero-touch provisioning capability |
| Operations | Real-time fabric telemetry and network analytics |
Note: The engineering architecture identifies 102.4 Tbps of switching capacity and 51.2 Tbps of front-panel bandwidth across the 64 x 800G OSFP interfaces. These values describe different platform characteristics.
Why 800G Matters in AI and HPC Fabrics
Distributed AI workloads create sustained communication between GPU nodes. Collective operations such as All Reduce, All Gather, Reduce Scatter, and Broadcast repeatedly move data across the fabric during model training.
As GPU count increases, network architects need more bandwidth while also managing power, thermals, latency, congestion, and rack density. The Axiom switch + LPO architecture addresses both sides of that requirement: high-density 800G switching and lower-power optical links.
GPU-to-GPU Traffic
AI training generates large amounts of east-west traffic between compute nodes, increasing the importance of fabric bandwidth and predictable communication performance.
Bandwidth Density
800G interfaces allow architects to increase aggregate fabric bandwidth without multiplying the number of physical links as quickly.
Power Per Link
At large scale, watts per optical module become a meaningful part of the rack and facility power budget. Lower-power LPO helps address that challenge directly.
Why Linear Pluggable Optics (LPO) Changes the 800G Power Equation
LPO removes the retiming DSP from the optical module and shifts more equalization responsibility to the host switch and NIC.
Traditional DSP-Based Optics
DSP
The DSP performs signal conditioning and retiming inside the optical module. This improves channel tolerance but adds power, heat, processing stages, component cost, and dependence on the DSP supply chain.
Axiom Linear Pluggable Optics (LPO)
Switch
SerDes
Driver
Axiom LPO removes the retiming DSP and is equalized for the Axiom TH6-based switch platform. This reduces module power and processing latency while eliminating reliance on the retiming DSP component.
Why Equalization Matters
LPO relies more heavily on the signal integrity, SerDes, and equalization capabilities of the host system. Pairing Axiom LPO with the Axiom switch allows the optical module to be equalized for the intended host platform instead of treating the optic as an independent component.
LPO Benefits and Engineering Considerations
| Area | Axiom LPO Advantage | Engineering Consideration |
|---|---|---|
| Power Consumption | Significantly reduced optical module power compared with a DSP-retimed architecture. | Validate actual module power, airflow, inlet temperature, and populated-port behavior. |
| Latency | Removes retiming DSP processing stages from the optical path. | Measure end-to-end latency against the target workload and platform baseline. |
| DSP Supply Dependency | No retiming DSP in the Axiom LPO module. | The architecture depends more heavily on the capabilities of the host SerDes. |
| Acquisition Cost | Simpler optical architecture can reduce component and acquisition cost. | Evaluate total link cost rather than only individual module price. |
| Maintenance Cost | Reduced power, cooling demand, and DSP dependency can improve lifecycle economics. | Include spares, energy, cooling, and service requirements in the TCO model. |
| Host SerDes | Axiom LPO is equalized for the Axiom TH6-based switch platform. | Equalization, lane margin, channel quality, PAM4 behavior, and firmware remain important. |
The Power Advantage Becomes Significant at AI Cluster Scale
Axiom engineering modeled the impact using a 2,048-GPU AI cluster with 4,096 optical modules.
The following example is a modeled scenario, not a universal savings guarantee. It compares an 18W DSP-based 800G optical module assumption with a 9W Axiom LPO assumption.
Modeled scenario assumes 4,096 optical modules, 18W DSP optics, 9W Axiom LPO, continuous operation, and PUE of 1.4. Actual results vary by hardware and facility conditions.
| Modeled Item | Assumption / Result |
|---|---|
| GPU Cluster Size | 2,048 GPUs |
| Optical Module Count | 4,096 optics |
| DSP Optical Power | 18W per module |
| Axiom LPO Optical Power | 9W per module |
| Direct Optical Power Reduction | 36.9 kW |
| Facility-Adjusted Reduction | 51.7 kW using a PUE assumption of 1.4 |
| Annual Energy Reduction | 452,892 kWh |
| Electricity Cost Assumption | $0.10/kWh |
| Modeled Annual Energy Savings | $45,289 |
| Modeled Five-Year Energy Savings | Approximately $226,445 |
Power Savings Are Only Part of the Economic Advantage
The Axiom switch + LPO architecture also targets lower optical acquisition cost, reduced cooling requirements, reduced DSP supply-chain exposure, and lower maintenance costs. At large AI cluster scale, these factors can contribute to a lower overall network TCO.
RoCEv2 for High-Performance AI Ethernet
Power-efficient 800G links still need predictable congestion behavior under AI workloads.
The Axiom switch platform supports RoCEv2 environments where RDMA traffic requires careful management of congestion, buffering, pause behavior, queue occupancy, retransmissions, and recovery.
Priority-Based Flow Control
Helps protect loss-sensitive RDMA traffic by managing pause behavior during periods of congestion.
Explicit Congestion Notification
Signals congestion before packet loss occurs and helps stabilize high-utilization Ethernet fabrics.
Buffer and Queue Management
Shared buffers, dynamic thresholds, and queue management help the fabric absorb bursts and maintain stable traffic behavior.
LPO Makes System-Level Visibility More Important
Because Linear Pluggable Optics (LPO) relies more heavily on the host-side electrical path, engineers need visibility into both optical health and fabric behavior. A successful link-up by itself is not enough.
Physical-Layer Visibility
BER trends, FEC behavior, lane margin, optical power, temperature, PAM4 behavior, and link stability.
Congestion Visibility
ECN marking rates, PFC pause statistics, buffer occupancy, congestion hot spots, queue behavior, and retransmissions.
Application Visibility
RDMA performance, GPU utilization, communication stalls, collective-operation performance, and job completion behavior.
Engineering Qualification for the Integrated Solution
The switch + LPO architecture still needs to prove performance under representative network and AI workloads.
Axiom's engineering qualification framework evaluates the physical layer, Ethernet behavior, RoCEv2 congestion controls, AI workload performance, interoperability, environmental stability, and operational visibility.
The criteria below describe Axiom's engineering qualification framework and should not be interpreted as completed results for every configuration unless supported by the applicable qualification report.
OMA
TDECQ
BER
Lane margin
Latency
Traffic stability
ECN
RDMA
GPU utilization
Job stability
Soak testing
Failure recovery
Optical & Signal Integrity
TX/RX power, OMA, TDECQ, extinction ratio, receiver sensitivity, jitter, PAM4 behavior, equalization, lane margin, BER, and FEC.
Ethernet Performance
Line-rate throughput, IMIX, large frames, storage traffic, latency, jitter, tail latency, and sustained traffic stability.
RoCEv2 Qualification
PFC, ECN, RDMA throughput, retransmissions, congestion recovery, queue stability, buffers, fairness, and deadlock behavior.
AI Workload Validation
NCCL All Reduce, All Gather, Reduce Scatter, Broadcast, distributed inference, GPU utilization, and communication behavior.
Reliability & Environment
Thermal operation, full traffic load, failure scenarios, aging, burn-in, power cycling, link stability, and long-duration testing.
Interoperability & Operations
Switch and NIC interoperability, firmware, FEC negotiation, telemetry, configuration, optical monitoring, and failure analysis.
Example Engineering Qualification Targets
| Qualification Area | Engineering Target |
|---|---|
| Ethernet Throughput | Greater than 95% of line rate |
| Packet Loss Under RoCE Traffic | Zero |
| PFC Deadlocks | Zero |
| Retransmissions | Near zero |
| Uncorrectable FEC Events | Zero |
| GPU Communication Errors | Zero |
| GPU Utilization | Greater than 90% within the qualification framework |
| Link Stability | Greater than 99.999% |
| Thermal Evaluation | Nominal, elevated, stress, and maximum operating conditions |
| Extended Stability | Up to 30-day continuous soak and traffic testing |
Qualification targets define the engineering framework. Actual qualification status should be confirmed against the applicable Axiom test report, hardware configuration, firmware, optics revision, NIC, cabling, and workload.
Where the Axiom 800G Switch + LPO Solution Fits Best
Distributed GPU Clusters
Large-scale training environments where high east-west bandwidth, low latency, optical power, cooling, and congestion behavior all influence cluster performance.
High-Performance Computing
Scientific computing, research clusters, and other environments that benefit from dense 800G connectivity and reduced optical processing latency.
Power-Constrained AI Fabrics
Large deployments where thousands of optical links make watts per module, cooling demand, rack density, and long-term network economics important design factors.
800G AI Fabric Design Checklist
Evaluate the complete network path before production deployment.
- Define AI workload and GPU communication requirements.
- Define leaf, spine, and oversubscription targets.
- Confirm 800G port count and breakout requirements.
- Confirm switch firmware and NOS support.
- Review NIC type, firmware, and SerDes capability.
- Confirm the target Axiom switch + LPO configuration.
- Define fiber reach, connector type, and physical path.
- Confirm FEC mode and negotiation behavior.
- Define PFC and ECN configuration strategy.
- Review MMU, queue, and shared-buffer behavior.
- Define telemetry and historical trending requirements.
- Model optical, rack, and facility power.
- Validate thermal behavior with populated ports.
- Test traffic and congestion at representative utilization.
- Run representative AI collective workloads.
- Define failure recovery and acceptance criteria.
An Integrated Axiom 800G Networking Solution
The primary advantage of the Axiom architecture is that the network switch and Linear Pluggable Optics (LPO) are offered as an integrated solution rather than unrelated components.
Axiom Switch → Axiom LPO → Fiber → NIC → GPU Server
The Axiom 800G LPO is equalized for the Axiom Broadcom Tomahawk 6-based switch platform, providing a system-level approach to 800G connectivity focused on power efficiency, latency, supply-chain flexibility, and network economics.
Related Technical Resources
AI Networking Infrastructure Guide
Plan connectivity across 400G, 800G, 1.6T, switches, adapters, optics, cables, power, thermals, and validation.
800G Optics Guide
Review 800G optical architectures, deployment considerations, form factors, power, high-density use cases, and migration planning.
800G Transceiver Validation Guide
Review optical power, FEC behavior, thermals, traffic stability, platform support, and production-readiness checks.
Related Product Categories
Networking
Optical Connectivity
Network Adapters
Networking Cables
Axiom 800G Switch + LPO FAQs
The solution combines an Axiom Broadcom Tomahawk 6-based 102.4 Tbps network switch with 64 x 800G OSFP interfaces and Axiom 800G Linear Pluggable Optics (LPO). The Axiom LPO is equalized for the Axiom switch platform so the switch and optics can be deployed as an integrated 800G link solution.
Linear Pluggable Optics rely more heavily on the host SerDes and equalization capabilities than conventional DSP-retimed optics. Axiom's 800G LPO is equalized for the Axiom Broadcom Tomahawk 6-based switch platform, helping align the host electrical path and optical module as one solution.
Axiom 800G Linear Pluggable Optics (LPO) remove the retiming DSP used in conventional DSP-based optical modules. In Axiom's modeled example, an 18W DSP-based optic is compared with a 9W Axiom LPO, representing a 50 percent reduction in optical module power for that scenario.
LPO removes retiming DSP processing stages from the optical data path. This reduces optical signal-processing delay compared with architectures that retime signals inside the optical module.
Axiom 800G LPO does not use the retiming DSP found in traditional DSP-based optical modules. As a result, the LPO module is not dependent on availability of that retiming DSP component.
The architecture is designed to lower optical acquisition and maintenance costs while also reducing power and cooling requirements. Actual savings depend on port count, hardware configuration, electricity cost, PUE, utilization, and deployment scale.
They describe different platform characteristics. The Broadcom Tomahawk 6-based Axiom platform provides 102.4 Tbps of switching capacity, while its 64 x 800G OSFP front-panel interfaces represent 51.2 Tbps of front-panel bandwidth.
Yes. The architecture is designed for high-density 800G AI and HPC Ethernet fabrics, including distributed AI training, GPU-to-GPU communication, RoCEv2 environments, scientific computing, and other bandwidth- and power-intensive east-west networks.
Planning an 800G AI or HPC Network?
Talk with Axiom about the integrated Axiom 800G network switch and Linear Pluggable Optics (LPO) solution for high-density AI and HPC fabrics. Axiom can help evaluate port requirements, optical power, latency, RoCEv2, compatibility, and deployment requirements.
