Latency Comparison of OTN Router QSFP-DD in Intelligent Computing Center

QSFP-DD SR4 modules generally offer lower latency (~85 ns) than SR8 modules (~102 ns), making SR4 preferable for latency-sensitive HPC and AI workloads, while SR8 provides higher channel density at slightly higher latency.

QSFP-DD Module Overview

QSFP-DD (Quad Small Form-factor Pluggable Double Density) modules are widely used in high-performance data centers and intelligent computing centers due to their dual-density 8-channel design, backward compatibility with QSFP28, and support for 400G and 800G Ethernet connections . They are optimized for low-latency environments, particularly when forward error correction (FEC) is reduced or bypassed, which can otherwise introduce up to 100 ns of additional latency .

Latency Comparison

• QSFP-DD SR4: Uses 4-channel 100G PAM4 modulation, requiring 8-core fibers (MPO-12). Latency is typically around 85 ns, as it avoids complex signal multiplexing and Gearbox chips, making it suitable for server-to-switch connections and latency-sensitive HPC or AI workloads .
• QSFP-DD SR8: Uses 8-channel 50G PAM4 modulation, requiring 16-core fibers (MPO-16). Latency is slightly higher, around 102 ns, due to additional DSP and Gearbox processing. SR8 is optimized for switch-to-switch interconnects where higher channel density is prioritized over minimal latency .
• QSFP112 SR4: Can achieve latency below 60 ns with linear drive architectures, making it ideal for GPU cluster communication and ultra-low-latency HPC scenarios .

Factors Affecting Latency

1. FEC Usage: Enabling FEC increases latency; low-latency QSFP-DD modules often operate with reduced or no FEC to minimize processing delays .
2. DSP and Gearbox Chips: SR8 modules rely on these for signal multiplexing, which adds latency compared to SR4 modules .
3. Fiber Type and Distance: OM3/OM4 multimode fibers support short-range low-latency connections (up to 50 m for 8x25G without FEC), while longer distances may require single-mode fibers and coherent optics, slightly increasing latency .
4. Integration with OTN Routers: Bypassing traditional OTN layers using IP-over-DWDM architectures can further reduce end-to-end latency in data center interconnects .

Practical Implications for Intelligent Computing Centers

• HPC and AI Workloads: QSFP-DD SR4 or QSFP112 SR4 modules are preferred due to lower latency and compatibility with GPU clusters.
• High-Density Switch Interconnects: QSFP-DD SR8 modules provide higher channel density, suitable for aggregating multiple 100G/400G flows, though with slightly higher latency.
• Latency Optimization: Minimizing FEC, using short OM4 fiber runs, and leveraging direct IP-over-DWDM integration can significantly reduce latency in OTN router deployments . In summary, QSFP-DD SR4 modules strike a balance between low latency and moderate channel density, making them ideal for latency-sensitive intelligent computing centers, while SR8 modules prioritize higher throughput at the cost of slightly increased latency. QSFP112 SR4 offers the lowest latency for specialized HPC and AI applications.