Modern data center networks are defined by a relentless drive toward higher bandwidth, lower latency, and better power efficiency. At the heart of this evolution sits the humble optical transceiver—a small module that converts electrical signals from switches and adapters into light, and back again. Choosing the right transceiver form factor can determine whether an AI cluster, cloud fabric, or enterprise core network scales smoothly or stalls at 100G.

Among the most common form factors deployed today are SFP28, QSFP28, QSFP-DD, and OSFP. Each serves a different bandwidth tier, port density target, and migration path. For buyers of enterprise and data center hardware, understanding these differences is not just an engineering exercise—it directly affects procurement, interoperability, and future-proofing.

At Zhongxin Data, we supply NVIDIA Mellanox InfiniBand and Ethernet adapters, Spectrum-X and SN-series switches, and the optical infrastructure that ties them together. This guide compares the leading transceiver form factors and explains when to choose each one.

Why Form Factor Matters

Optical transceivers are standardized by size, electrical interface, and thermal envelope. A switch port that accepts QSFP28 cannot accept QSFP-DD without a mechanical adapter, and OSFP ports are physically incompatible with QSFP-DD modules. Beyond physical fit, the form factor determines:

  • Lane count and lane speed — 25G NRZ, 50G PAM4, or 100G PAM4 lanes
  • Aggregate bandwidth — 25G, 100G, 400G, or 800G per port
  • Power consumption and heat dissipation — critical at 400G and above
  • Cable and reach options — DAC, AOC, SR, LR, ER, ZR, ZR+
  • Backward compatibility — whether 100G modules can run in a 400G port

Making the wrong choice means buying modules that do not fit, cannot negotiate speed, or create thermal hotspots in dense switch chassis. The sections below break down each form factor and show how to map them to real workloads.

Form Factor Comparison: SFP28, QSFP28, QSFP-DD, and OSFP

Form Factor Max Speed Lanes × Lane Rate Typical Use Case Backward Compatibility
SFP28 25G / 28G 1 × 25G NRZ Server NIC access, leaf-to-server links SFP+ (10G) modules usually work
QSFP28 100G / 128G 4 × 25G NRZ 100G leaf-spine, storage backends QSFP+ (40G) modules sometimes work with speed negotiation
QSFP-DD 400G / 800G 8 × 50G PAM4 (400G) or 8 × 100G PAM4 (800G) 400G/800G spine, AI cluster backends QSFP28 (100G) modules fit mechanically and can operate at 100G
OSFP 400G / 800G 8 × 50G PAM4 (400G) or 8 × 100G PAM4 (800G) 400G/800G high-density AI fabric Not compatible with QSFP28/QSFP-DD without adapter

QSFP-DD is the most flexible of the high-density form factors. Its “double density” design keeps the same footprint as QSFP28 but adds a second row of electrical contacts, doubling the lane count from four to eight. This is why a QSFP-DD port can accept a QSFP28 module and run it at 100G, making it a safe migration choice for data centers moving from 100G to 400G.

OSFP, by contrast, is a slightly larger module designed for superior thermal performance. It is favored by NVIDIA Spectrum-X and many next-generation AI switches because 400G and 800G optics generate significantly more heat than 100G modules. OSFP does not accept QSFP28 directly, so organizations deploying OSFP switches need to plan for all-OSFP optics or use active adapters.

Reach and Cable Type Selection

Beyond form factor, data center engineers must select the right reach class. Transceiver labels such as SR, LR, ER, and ZR describe distance, fiber type, and optical budget. The table below maps common reach options to typical data center distances.

Reach Class Typical Distance Fiber Type Best For
DAC 1–3 meters Copper twinax Very short rack or adjacent switch links
AOC 3–30 meters Multimode fiber Intra-row links, top-of-rack to spine
SR (SR4/SR8) 50–100 meters Multimode fiber Same-room or same-floor data center links
LR (LR4/LR8) Up to 10 km Single-mode fiber Building-to-building or campus interconnect
ER / ZR Up to 80 km Single-mode fiber Metro links, disaster-recovery sites

For most data center fabrics, AOC or SR optics handle the majority of leaf-to-spine and server-to-leaf links. Only when connecting data centers across a campus or metro area do LR, ER, or ZR optics become necessary. A common mistake is over-specifying reach: buying LR optics for intra-rack connections increases cost and power with no benefit.

Power and Thermal Considerations

As lane rates climb, transceiver power climbs with them. A 10G SFP+ module might draw under 1 watt, while an 800G OSFP module can draw 15–20 watts. In a 32-port or 64-port switch, this translates to hundreds of watts of heat concentrated in the front panel. Switches designed for OSFP modules therefore include more aggressive thermal management, larger heat sinks, and higher airflow fans.

When designing a high-density AI cluster, power and cooling must be budgeted for the entire link—not just the switch. Two OSFP transceivers plus cable insertion loss can add 30–40 watts per 800G connection. Across a 400-node cluster with multiple links per node, that becomes a meaningful data center power and cooling load.

This is one reason used enterprise switches remain attractive: they often include proven thermal designs and front-to-rear airflow patterns that match standard hot-aisle/cold-aisle layouts. At Zhongxin Data, we test and certify used switches so they can be deployed without surprises.

“The cheapest transceiver is not the one with the lowest unit price. It is the one that fits your switch port, runs at the right speed, reaches the right distance, and does not push your rack over its thermal budget.”

Interoperability: OEM vs Third-Party Optics

Enterprise switch vendors typically sell their own branded optics, often at a significant premium. Third-party optics that meet MSA (Multi-Source Agreement) standards are widely used and can reduce cost by 50–80%. However, interoperability depends on:

  • Firmware compatibility — some switches reject non-OEM modules or limit diagnostics
  • Encoding of vendor information in EEPROM — third-party suppliers can encode optics to emulate OEM IDs
  • Regulatory compliance — laser safety (Class 1/Class 1M), RoHS, and country-specific certifications
  • Quality consistency — optical power, extinction ratio, and temperature stability vary by supplier

For mission-critical AI fabrics, many operators run a mix: OEM optics for the most heavily used spine links and qualified third-party optics for less demanding server access links. Zhongxin Data can supply both OEM and compatible optics, helping customers balance budget and risk.

Digital Diagnostic Monitoring and Telemetry

Modern optical transceivers expose real-time telemetry through the Digital Diagnostic Monitoring (DDM) interface, also known as Digital Optical Monitoring (DOM). This data is essential for proactive network operations because it reveals the health of each link before bit errors or drops become visible to applications. The parameters reported by DDM include transmitted optical power, received optical power, laser bias current, temperature, and supply voltage.

For 400G and 800G links, DDM becomes even more important. Small degradations in receive power at these lane rates can quickly produce uncorrectable errors. Operations teams should set thresholds for each parameter and integrate them into their monitoring stack. NVIDIA Mellanox switches and adapters expose this information through standard interfaces, making it straightforward to alert on out-of-range conditions.

When buying optics in bulk, ask the supplier for DDM consistency across the lot. Inexpensive modules may report approximate values or omit certain parameters, which limits visibility. Used enterprise optics from Zhongxin Data are tested for DDM accuracy and optical power margins before shipment, reducing the chance of field failures.

Cost and Procurement Strategy

Transceiver spending often represents a significant share of a data center network budget—sometimes more than the switches themselves. A 32-port 400G switch with OSFP modules can cost tens of thousands of dollars in optics alone. Procurement strategy therefore matters as much as technical selection.

A practical approach is to segment links by criticality. Mission-critical spine-to-spine or AI backend links that carry bursty, high-value traffic should use the highest-quality optics, whether OEM or premium-compatible, with full telemetry and warranty. Access-layer links that can tolerate a short failover window are good candidates for cost-optimized compatible modules or used original modules.

Another consideration is lead time. New-generation 800G optics can have long lead times during supply shortages. Buying used or carry-over 400G optics while waiting for 800G availability can keep a project on schedule. Standardizing on QSFP-DD where possible also preserves investment because the same switch ports can accept 100G, 400G, and eventually 800G modules.

Common Mistakes When Choosing Transceivers

Even experienced buyers make predictable mistakes when selecting optics. Avoiding these pitfalls can prevent expensive returns and deployment delays.

  • Confusing reach classes — ordering SR optics for a 2 km campus link, or LR optics for intra-rack connections, wastes money and may not work.
  • Ignoring switch port compatibility — not all 400G ports accept all module types; verify whether the switch expects QSFP-DD, OSFP, or QSFP112.
  • Overlooking fiber polarity — multimode MPO cables have multiple polarity methods; mismatched polarity produces no link.
  • Mixing coding types without checking — some 100G links use CWDM4, others use PSM4 or SR4; the receiver must match the transmitter technology.
  • Underestimating power and cooling — dense 400G/800G switch deployments can exceed rack thermal budgets if optics are not accounted for.

Taking the time to audit existing fiber, switch firmware, and port types before placing an order will almost always save time and money compared to troubleshooting after the hardware arrives.

Matching Transceivers to Network Tiers

A three-tier data center fabric typically uses different form factors at each layer. The following table maps common tiers to recommended optics.

Network Tier Typical Bandwidth Recommended Form Factor Typical Reach
Server access 10G / 25G SFP28 / SFP+ DAC or AOC
Leaf-to-server 25G / 100G SFP28 / QSFP28 breakout DAC / AOC / SR
Leaf-to-spine 100G / 400G QSFP28 / QSFP-DD SR / LR
Spine-to-core 400G / 800G QSFP-DD / OSFP LR / ER / ZR
AI cluster backend 400G / 800G NDR OSFP / QSFP-DD112 AOC / SR / DR

Migration from 100G to 400G and 800G

Most enterprises are not building greenfield 800G fabrics. They are upgrading from 100G leaf-spine networks to 400G, and in some cases planning 800G for AI training clusters. The migration strategy depends on whether the existing switch hardware supports QSFP-DD or OSFP.

If the current leaf switches use QSFP28 ports, the simplest path is to replace them with QSFP-DD capable switches that can still accept the existing QSFP28 optics. This protects the installed base while enabling 400G uplinks. For AI workloads, OSFP-based switches such as the NVIDIA Spectrum-X SN5600 provide 400G/800G ports with the thermal headroom needed for sustained training traffic.

Cabling also needs a migration plan. Multimode SR optics at 100G may not scale cleanly to 400G over the same fiber plant because 400G SR8 requires eight lanes rather than four. A fiber inventory audit is a sensible first step before ordering new optics.

Frequently Asked Questions

Can I plug a QSFP28 module into a QSFP-DD port?

Yes. QSFP-DD ports are mechanically backward compatible with QSFP28 modules. The module will operate at its native 100G speed using four lanes, leaving the extra QSFP-DD electrical contacts unused. This is one of the main reasons QSFP-DD is popular for 100G-to-400G migration.

Can I plug a QSFP28 or QSFP-DD module into an OSFP port?

No, not directly. OSFP uses a different mechanical size and pinout. Active adapters or replacement modules are required to bridge OSFP and QSFP-DD/QSFP28 ports. If your switch has OSFP ports, plan for OSFP optics.

What is the difference between QSFP-DD and QSFP-DD800?

QSFP-DD typically refers to modules using eight 50G PAM4 lanes for 400G total. QSFP-DD800 (also called QSFP-DD112) uses eight 100G PAM4 lanes for 800G. The physical form factor is the same, but the electrical signaling and host interface requirements are more demanding for 800G.

When should I choose DAC over optical transceivers?

Use Direct Attach Copper (DAC) cables for very short links, typically within the same rack or between adjacent switches, up to about 3 meters. DAC is lower power, lower latency, and cheaper than optics. For longer distances or flexible routing, use AOC or optical transceivers with fiber.

Are third-party MSA optics reliable in NVIDIA Mellanox switches?

MSA-compatible optics from qualified suppliers work reliably in Mellanox SN-series switches, but firmware may report non-OEM modules. For InfiniBand NDR and high-speed Ethernet fabrics, many operators prefer OEM optics or vendor-qualified compatible modules to ensure full telemetry and support coverage.

Conclusion

Choosing between QSFP28, QSFP-DD, and OSFP is not just about bandwidth. It is about matching form factor, lane rate, reach, power, and migration path to the actual network tier. QSFP-DD offers the smoothest upgrade from 100G, while OSFP provides the thermal and electrical headroom needed for 800G AI fabrics. SFP28 remains the right choice for server access and low-density links.

At Zhongxin Data, we help data center and AI infrastructure buyers source the right optics, switches, and adapters together. Explore our product ranges:

Whether you are refreshing a 100G leaf-spine fabric or building a new 800G AI cluster, selecting the correct transceiver form factor early will save cost, reduce deployment risk, and keep your network ready for the next generation of data center workloads.