Increasing Energy Efficiency for AI Clusters and Data Centers while Maintaining Performance

Dr. Carlos Berto

Feb. 3, 2025

The network interconnection components between the elements inside or outside the racks or nodes of compute engines (GPU, CPU and storage) required for these applications consume significant electrical power from the grid. Based on the power constraints, the AI cluster or data centers are redesigning the network interconnection to reduce the power consumption of the network elements. Mixing different technologies and topologies is the key to minimizing overall power consumed by the network and having the most powerful and efficient AI and data center infrastructure.

Linear Receive Optics (LRO) and Linear Pluggable Optics (LPO) enable engineers to build AI infrastructures and reduce power consumption from network equipment.

The LRO and LPO architecture are related to Silicon Photonics technology which moved the fully retimed module (DSP) to different locations, such as a separate driver and TIA modules. The advantage of removing the DSP is reducing power consumption by 40%.

The figure below illustrates the differences between a Fully Retimed (DSP), LRO and LPO.

Linear Pluggable Optics (LPO) Technology

Linear Pluggable Optics (LPO) is an advanced technology used in data centers and high-speed networks to enhance data transmission efficiency. Unlike traditional optical modules, LPO transceivers lack full retiming (DSP) circuitry, relying on the host system for retiming and signal conditioning. This approach offers several advantages and disadvantages compared to traditional technologies.

Advantages of LPO Technology:

1. Lower Power Consumption: By eliminating the DSP, LPO significantly reduces power consumption, which is crucial for power-constrained environments like AI clusters and data centers.
2. Cost Efficiency: The absence of DSP reduces the bill of materials (BOM) cost, making LPO modules more cost-effective.
3. Low Latency: Without the DSP, LPO modules can achieve lower latency, which is beneficial for applications requiring real-time data processing.
4. Easy Maintenance: LPO modules are easier to maintain due to their simpler design.

Disadvantages of LPO Technology:

1. Signal Integrity Challenges: LPO systems need to handle signal integrity with about 16 dB of loss from the host switch to the module, which can be challenging.
2. Compatibility Issues: Ensuring robust communication between different LPO modules and switches, especially from different manufacturers, can be complex.
3. Limited Distance: LPO is generally more suitable for short-distance transmission compared to traditional fully retimed modules.

Summary

LPO technology offers significant benefits in terms of power efficiency, cost, and latency, making it suitable for modern data centers and AI clusters. However, it also presents challenges in signal integrity and compatibility. When compared to traditional optical modules and CPO, LPO stands out for its simplicity and cost-effectiveness, though it may not be the best choice for long-distance or highly complex interconnect scenarios.

-------------------------------------------------------------------------------------------------------------------------------

About the Author

Carlos Berto, PhD

Dr. Berto is the Director of Network Engineering at Axiom. With over 25 years of leadership experience in the Telecommunications industry, he has a proven track record of developing innovative new products and solutions. Dr. Berto holds a Ph.D. in Engineering and possesses in-depth knowledge of network devices.