As large-scale language models exceed trillions of parameters and scientific simulations enter the exascale realm, data transmission within data centers faces unprecedented pressure. When traditional 100Gb/s links become a computing bottleneck, InfiniBand NDR (Next Data Rate) technology emerges as the next-generation “vascular system” for supercomputing and AI clusters. The core driving this system’s efficient operation is its revolutionary optical communication engine—the NDR optical module.

Technical Architecture: The Physical Layer Cornerstone of NDR Networks

Optical-Electrical Co-Design

Modulation: The entire system utilizes PAM4 (4-level pulse amplitude modulation), achieving 4×106.25 Gb/s transmission simultaneously on four electrical and optical ports, for a total bandwidth of 425 Gbps (including overhead).

Laser Technology:

DR4 single-mode module: 1310 nm DFB laser (distributed feedback laser) paired with a PIN detector, supporting long-haul transmission of 500 m with a bit error rate ≤ 1E-15.

SR4 multi-mode module: 850 nm VCSEL laser (vertical cavity surface emitting laser), power consumption ≤ 8.5 W, suitable for 50 m intra-cabinet interconnection (OM4 fiber).

Signal Processing: A built-in 7 nm DSP chip (e.g., Broadcom solution) supports forward error correction (FEC) and clock data recovery (CDR), significantly reducing signal distortion.

Packaging and Heat Dissipation Innovation

OSFP-XD package: Slightly larger than the traditional QSFP, it supports 16-channel expansion and reserves space for 1.6T upgrades.

Thermal Management: Utilizing COB (chip-scale packaging) and micro heat pipes, it maintains an operating temperature below 70°C at ≤10W power consumption, ensuring stability in high-temperature equipment room environments.

Interface Standardization: Standardized use of MPO-12/APC connectors (8-degree angled physical contact) reduces multi-fiber insertion loss to below 0.2dB.

Application Scenario: How the NDR Module Drives High-Performance Networks

Networking Logic for Supercomputing and AI Clusters

Breakout Link: The output of a single 800G switch (such as the NVIDIA Quantum-2) is split into two 400G signals via an MPO-24 breakout cable and connected to the OSFP port of the ConnectX-7 network interface card.

Latency-Sensitive Tasks: In all-to-all communications (such as AI parameter synchronization), the NDR module offers sub-microsecond latency (<100ns), even lower than PCIe 4.0, eliminating GPU computational waits.

Fault Tolerance and Energy Efficiency Optimization

Dynamic Routing Support: Combined with InfiniBand SHARP technology, the module can offload network-layer aggregation computations (such as AllReduce), reducing CPU load by 30%.

Power Consumption Comparison:

SR4 multi-mode: 8.5W (4 lanes fully operational) / 6.5W (2 lanes reduced speed)

DR4 single-mode: ≤10W (500m transmission)

→ 2x the energy efficiency of 100G QSFP28 modules at the same bandwidth.

Comparison of Mainstream Modules on the Market and Selection Recommendations

Selection Guide:

Distance > 50m → Requires DR4 single-mode (avoids multimode fiber modal dispersion) 3

Cost-sensitive and distance ≤ 50m → SR4 multi-mode (saves laser and single-mode fiber costs) 6

Future Compatibility → Prioritize modules that support the CMIS 4.0 protocol (software-upgradeable speed).

Deployment Considerations

Polarity Management: MPO-12 ports must be used with Type B polarity patch cables (fiber transceiver and transceiver crossover), otherwise the link will fail.

Cooling Redundancy: In dense deployment scenarios (such as a 1U switch fully configured with 64 ports), ensure an airflow velocity > 4 m/s to prevent localized overheating.

Compatibility Verification:

Third-party modules (such as FS/Jiuzhou Interconnect) must pass IBTA BER testing to ensure compatibility with the original QM9700 switch.

Flat-top modules are only used for network cards; heat sink modules are used for switches.

Summary: The Core Value of NDR Modules

As the “last centimeter” engine of InfiniBand NDR networks, the 400G OSFP module, through optoelectronic co-optimization (PAM4 + DSP + silicon photonics) and thermal/signal stability design, boosts communication efficiency to over 90% (compared to 60% in the EDR era) for tens of thousands of cards in AI training. With silicon photonics yield exceeding 85% by 2025 and liquid cooling solutions becoming more widespread, its cost is expected to drop by 40%, further promoting the affordable deployment of exascale supercomputing and models with tens of billions of parameters.