100G vs 400G Optical Transceiver

The 100G and 400G generations sit at opposite ends of the current data center transition. 100G QSFP28 is the most mature high-speed interface in the industry—supported by every switch vendor, available from dozens of module suppliers, and deployed in hundreds of millions of ports worldwide. 400G QSFP-DD and OSFP represent the current state of the art, driving the spine layer of AI clusters, the fabric of hyperscale data centers, and the interconnection of GPU nodes that demand multiple terabits per second of east-west bandwidth.

The two generations differ in almost every dimension. They use different form factors, different lane architectures, different modulation formats, and different physical media. A 100G QSFP28 module carries four 25G NRZ lanes over eight multimode fibers or two single-mode fibers. A 400G QSFP-DD module carries eight 50G PAM4 lanes over sixteen multimode fibers, eight single-mode fibers, or four wavelengths on a duplex fiber pair. The 400G module is twice the width of the 100G module, consumes roughly twice the power, and costs roughly three to four times as much per port—but delivers four times the bandwidth.

The choice between them is not about which is better. It is about which layer of the network each belongs on, and whether the bandwidth demand justifies the cost, power, and complexity of the higher-speed interface. For most data centers, 100G remains the aggregation-layer workhorse, while 400G serves the spine and core layers, the AI fabric, and the highest-bandwidth DCI links. For AI clusters, 400G has become the default interconnect for GPU nodes, and the question is not whether to deploy it but how quickly the rest of the network must be upgraded to keep pace.

This guide examines both transceiver families in depth: their lane architecture, modulation, form factors, reach options, power and cost characteristics, breakout capability, and the architectural patterns that combine them in modern data centers and AI clusters.

1. Different Form Factors, Different Lane Architecture

The most visible difference between the two generations is the form factor. QSFP28 and QSFP-DD share the same basic design lineage—the "Quad Small Form-factor Pluggable" family—but QSFP-DD is physically wider to accommodate twice the number of electrical lanes.

QSFP28 has a 38-pin edge connector and supports four electrical lanes. QSFP-DD (Double Density) has a 76-pin edge connector and supports eight electrical lanes, arranged in two rows. The QSFP-DD module is backward compatible with QSFP28 cages: a QSFP28 module can be plugged into a QSFP-DD port, but a QSFP-DD module cannot fit into a QSFP28-only port because the wider connector and additional pin rows require the larger cage.

OSFP (Octal Small Form-factor Pluggable) is a competing form factor for 400G and 800G, with a different mechanical design that includes an integrated heat sink for higher power dissipation. OSFP modules are not backward compatible with QSFP cages and require dedicated OSFP ports.

The QSFP-DD form factor's backward compatibility is a significant advantage for migration. An operator can deploy a 400G-capable switch and populate it with 100G QSFP28 modules initially, then upgrade to 400G QSFP-DD modules as bandwidth demand grows. This allows the switch hardware to be deployed with a clear upgrade path, avoiding the cost of replacing modules during the switch upgrade.

2. Lane Architecture: 4×25G NRZ vs 8×50G PAM4

The lane architecture difference is more consequential than the form factor difference. Both generations use four lanes at 25G for 100G, but 400G uses eight lanes at 50G. This means 400G modules use twice the lane count and twice the lane rate.

2.1 100G: Four 25G NRZ Lanes

A 100G QSFP28 module operates four SerDes lanes at approximately 25.78 Gbps each, using NRZ modulation with 64b/66b encoding. The total data rate is 100 Gbps. The four lanes are independent—each carries one quarter of the aggregate bandwidth—and the module aggregates them into a single 100G Ethernet stream.

The NRZ modulation format is simple: two signal levels, one bit per symbol, no mandatory FEC for short-reach links. The eye diagram shows a single opening per lane, and the receiver makes a single decision per symbol period. The noise margin is large, and the signal integrity requirements are moderate.

2.2 400G: Eight 50G PAM4 Lanes

A 400G QSFP-DD module operates eight SerDes lanes at approximately 53.125 Gbps each, using PAM4 modulation with 256b/257b transcoding and Reed-Solomon forward error correction. The total data rate is 400 Gbps. The eight lanes are independent—each carries one eighth of the aggregate bandwidth—and the module aggregates them into a single 400G Ethernet stream.

The PAM4 modulation format is more complex: four signal levels, two bits per symbol, mandatory FEC. The eye diagram shows three openings per lane, and the receiver must distinguish four levels rather than two. The noise margin is reduced by approximately 9.5 dB compared to NRZ, which is why FEC is required.

The combination of doubled lane count and doubled lane rate is what gives 400G its fourfold bandwidth advantage over 100G. It also makes 400G more sensitive to signal integrity issues: the higher lane rate requires tighter impedance control, lower insertion loss, and better crosstalk isolation, while the PAM4 modulation requires lower noise and better linearity.

3. Reach Options

The reach options for the two generations differ in fiber count, wavelength plan, and maximum distance. The 400G generation offers more variety, with variants optimized for different reaches and fiber counts.

3.1 Multimode Variants

The multimode variants of both generations use parallel optics over multiple fibers. A 100GBASE-SR4 module uses four 25G NRZ lanes over eight multimode fibers terminated in an MPO-12 connector. A 400GBASE-SR8 module uses eight 50G PAM4 lanes over sixteen multimode fibers terminated in an MPO-24 connector.

The multimode reach is nearly identical between the two generations. Both 100GBASE-SR4 and 400GBASE-SR8 reach 70 meters over OM3 and 100 meters over OM4, because the higher PAM4 noise penalty at 400G is offset by the fact that the per-lane baud rate is similar. The 400GBASE-SR4.2 variant is a bidirectional design that uses eight fibers instead of sixteen, with each fiber carrying 50G in both directions using wavelength division multiplexing.

3.2 Single-Mode Variants

The single-mode variants differ more substantially, both in reach and in wavelength plan. The 400G generation offers several distinct approaches to single-mode transmission.

The 400G single-mode variants offer more reach options than 100G. The 400GBASE-DR4 variant reaches 500 meters over eight parallel single-mode fibers, filling the gap between multimode SR8 (100 m) and single-mode FR4 (2 km). The 400GBASE-FR4 variant reaches 2 kilometers over a duplex LC interface using four CWDM wavelengths, matching the reach of 100G CWDM4 but with four times the bandwidth. The 400GBASE-LR4 variant reaches 10 kilometers over the same duplex LC interface, and the ER4 variant reaches 30 to 40 kilometers.

The CWDM wavelength grid used by 400GBASE-FR4, LR4, and ER4 is the same grid used by 100GBASE-CWDM4 (1271, 1291, 1311, 1331 nm). This means that a fiber plant built for 100G CWDM4 can be upgraded to 400G FR4, LR4, or ER4 with only module changes—the CWDM multiplexers and demultiplexers remain compatible. This is one of the strongest arguments for deploying CWDM-based single-mode infrastructure in new builds, because it provides a clear upgrade path from 100G to 400G without fiber plant changes.

4. Power Consumption

Power consumption is where the 100G and 400G generations differ most visibly. The 400G module consumes roughly twice the power of a 100G module in absolute terms—the PAM4 DSP is more complex, the higher lane rate requires more powerful drivers and amplifiers, and the module contains more optical components. But the power per gigabit is comparable or slightly better.

The comparison is nuanced. The 100G SR4 module is the most power-efficient per port, consuming only 1.5 to 2.5 watts. But it delivers only 100 Gbps. The 400G SR8 module consumes 7.0 to 10.0 watts—roughly four times the power—while delivering four times the bandwidth. The power per gigabit is therefore comparable, at approximately 18 to 25 mW per Gbps for both.

The single-mode comparison is also comparable. A 100G CWDM4 module consuming 4.0 watts delivers 40 mW per gigabit. A 400G FR4 module consuming 10.0 watts delivers 25 mW per gigabit. The 400G module is more power-efficient per gigabit despite consuming more than twice the absolute power.

For a switch with 32 QSFP-DD ports, the difference in total optical power is significant. A 32-port 400G switch dissipating 32 × 10.0 W = 320 W from FR4 optics delivers 12.8 Tbps. The equivalent bandwidth in 100G ports would require 128 ports at 4.0 W each, consuming 512 W. The 400G switch uses 37 percent less optical power for the same bandwidth—a substantial improvement in power efficiency.

The thermal implications of 400G are more demanding than 100G. A 32-port 400G switch dissipating 320 W from optics alone requires significant airflow and thermal headroom, which is why 400G deployments often use liquid cooling or enhanced airflow designs. The higher per-port power draw of 400G is one reason why the transition from 100G to 400G is accompanied by changes in rack cooling infrastructure.

5. Cost Structure

Cost follows a pattern familiar from every generation transition: the higher-speed module costs more per port but less per gigabit. The 400G QSFP-DD module is roughly three to four times the price of a 100G QSFP28 module, but delivers four times the bandwidth.

The cost per gigabit is comparable between the two generations for the multimode variant, and better for 400G in the single-mode variant. The 400G module's higher absolute price is offset by its higher bandwidth, making the cost per delivered gigabit competitive.

The switch port cost difference is more significant. A 32-port 400G switch delivers 12.8 Tbps in a single rack unit, while the equivalent bandwidth in 100G ports would require 128 ports across four switch chassis. The switching ASIC, power supply, chassis, and management hardware are shared across more bandwidth in the 400G switch, reducing the cost per gigabit of switching capacity.

One important consideration is the NIC cost. A 400G QSFP-DD NIC costs significantly more than a 100G QSFP28 NIC, typically three to five times more depending on the chipset and port count. For a server upgrade, the NIC premium may dominate the total cost, and the decision to upgrade to 400G must account for this. In AI clusters, where GPU nodes require multiple 400G links, the NIC cost is a significant fraction of the total node cost.

6. Breakout and Aggregation

Breakout is where the architectural relationship between the two generations becomes clear. A 400G QSFP-DD port can break out into multiple lower-speed links, allowing a single high-speed port to serve the 100G or 50G access layer.

The 400G-to-4×100G breakout is the most common configuration in 400G deployments. A single 400G QSFP-DD port on a spine switch connects via a breakout cable to four 100G QSFP28 ports on leaf switches. This allows the spine switch to aggregate leaf uplinks with fewer high-speed ports while presenting 100G granularity at the aggregation layer.

The 400G-to-8×50G breakout is also supported, allowing a single 400G port to serve eight 50G access links. This is useful in deployments where the access layer is at 50G and the spine layer is at 400G. The breakout cable splits the 400G port into eight independent 50G links, each carrying the full 50G data rate.

The breakout capability of the 400G port is more flexible than that of the 100G port. A 100G QSFP28 port can break out into four 25G links or two 50G links, but it cannot serve 100G access links directly because the port itself is 100G. A 400G QSFP-DD port can serve 100G, 50G, or 200G access links through breakout, making it a more versatile aggregation interface for mixed-speed deployments.

7. Architectural Roles in the Data Center

Both generations occupy the aggregation and spine layers, but they serve different generations of the access layer and different fabric architectures.

7.1 100G in the 25G Access Generation

100G QSFP28 is the standard aggregation interface for the 25G access generation. A leaf switch with 48 × 25G access ports uses 4 to 8 × 100G uplinks to connect to the spine layer. The 100G uplinks match the 25G access lanes in a 4:1 ratio—each 100G port aggregates four 25G access links.

The 100G QSFP28 ecosystem is mature and widely supported across all major switch platforms. The modules are available in a full range of reach options, from 70-meter multimode SR4 to 40-kilometer ER4, and the cost per port is the lowest of any high-speed interface. For networks where 100 Gbps per uplink is sufficient, QSFP28 remains a practical and cost-effective choice.

7.2 400G in the Spine and AI Fabric

400G QSFP-DD is the standard interface for the spine layer in hyperscale data centers and AI clusters. A spine switch with 32 to 64 ports of 400G connects to leaf switches using 400G uplinks or 4×100G breakouts. The 400G ports provide the aggregate bandwidth needed to serve hundreds of 25G or 50G access ports without oversubscription.

In AI clusters, 400G has become the default interconnect for GPU nodes. Each GPU node connects to the fabric using one or more 400G links, and the spine layer aggregates these links into a non-blocking or low-oversubscription fabric. The 400G links provide the bandwidth needed for gradient synchronization, parameter exchange, and collective communication operations that dominate AI training traffic.

The pattern is clear: each generation of access speed pairs with a spine speed that is four times the lane rate. 25G access pairs with 100G aggregation. 50G access pairs with 200G aggregation. 100G access pairs with 400G aggregation. The 100G-to-400G transition is the current step in this progression, accompanying the 25G-to-100G and 50G-to-100G access upgrades.

8. 100G and 400G in AI Data Centers

AI workloads have driven adoption of both interfaces, but in different roles and at different layers of the network.

8.1 GPU Cluster Interconnects

In AI clusters, GPU nodes use multiple high-speed links for inter-node communication. The data plane typically uses 400G or 800G links, while the management, storage, and control-plane traffic uses 100G or 200G links. The 400G QSFP-DD interface provides the bandwidth needed for high-performance GPU interconnect, and it can break out to 4×100G for the aggregation layer where leaf switches connect to the spine.

NVIDIA's DGX SuperPOD and similar AI cluster architectures use 400G QSFP-DD or OSFP for GPU interconnect, with each GPU connecting to the fabric through one or more 400G links. The 400G links provide the bandwidth needed for collective communication operations, and the spine layer aggregates them into a non-blocking fabric.

8.2 Storage Fabrics

NVMe over Fabrics and distributed storage systems increasingly use 400G interfaces for storage traffic, replacing the 100G links that were common in earlier generations. A single 400G link can serve multiple storage devices through a breakout, or connect a storage node directly to the fabric. The 100G interface remains relevant for lower-tier storage and for replication traffic that does not require the full bandwidth of 400G.

8.3 Rail-Optimized Topologies

In rail-optimized AI topologies, each GPU in a server connects to a dedicated leaf switch—one "rail" per GPU rank. The rail switches use 400G or 800G uplinks to connect to the spine layer, while the GPU connections may use 400G or 800G depending on the generation. The 100G interface remains relevant for the management network that runs alongside the data plane, connecting baseboard management controllers, storage, and orchestration services.

9. Comparison Summary

10. Selection Framework

11. Migration and Coexistence

The 100G and 400G generations are designed to coexist and to support phased migration.

11.1 Upgrading from 100G to 400G

The migration from 100G to 400G is not simply a module swap. The host port must support 50G PAM4 signaling, appropriate FEC, and the correct port mode. If the host port does not support 400G operation, the QSFP-DD module may be detected incorrectly or fail to establish link.

For switches that support both rates, the migration can be phased. Deploy 400G-capable switches with 100G QSFP28 modules initially, then upgrade the modules to 400G QSFP-DD as bandwidth demand grows. This allows the switch hardware to be deployed with a clear upgrade path, and it avoids the cost of replacing modules during the switch upgrade.

11.2 FEC Configuration

FEC configuration is a common source of deployment issues when migrating from 100G to 400G. The 400G QSFP-DD module requires RS-FEC (Reed-Solomon forward error correction) to achieve acceptable bit error rates at the PAM4 noise margin. If the FEC is configured incorrectly—or if the switch port does not support FEC—the link may come up but produce corrupt traffic, with the LED showing green while the link is effectively unusable.

The FEC configuration must match at both ends of the link. If one end applies RS-FEC and the other does not, the link will fail or operate with high error rates. This is different from 100G NRZ links, where FEC is often optional and can be negotiated. For 400G links, FEC is mandatory and must be configured correctly on both ends.

11.3 Fiber Plant Considerations

The fiber plant migration depends on the variant. A 100G SR4 deployment over multimode fiber can typically be upgraded to 400G SR8 or SR4.2 without fiber changes, because the reach specifications are nearly identical. A 100G CWDM4 deployment over single-mode fiber can be upgraded to 400G FR4 with only module changes, because both use the same CWDM wavelength grid and duplex LC interface.

A 100G LR4 deployment, however, cannot be upgraded to 400G LR4 without changes to the multiplexing infrastructure. The 100G LR4 uses the LAN-WDM grid, while 400G LR4 uses the CWDM grid. The CWDM multiplexers and demultiplexers in the fiber plant are not compatible with LAN-WDM, so a different passive infrastructure is required.

12. Market Context and Outlook

The 100G QSFP28 market remains the largest segment of the high-speed optical transceiver market by unit volume, supported by the broad deployment of 25G access / 100G aggregation leaf-spine fabrics. The ecosystem is mature, multi-vendor, and well-supported, with modules available from dozens of suppliers and compatible with all major switch platforms.

The 400G QSFP-DD market is growing rapidly, driven by AI cluster deployments and the upgrade of spine layers that can no longer be served by 100G. The 400G module is the current-generation workhorse for AI fabric interconnect, and it is being deployed in large volumes by hyperscalers, cloud providers, and AI infrastructure builders.

The broader market context is one of accelerating transition. AI data centers are shifting from 100G-per-lane to 200G-per-lane electrical and optical interfaces, driven by the bandwidth demands of dense GPU clusters. The 400G QSFP-DD module is the current-generation workhorse for this transition, providing four times the bandwidth of 100G without the cost and complexity of 800G. As AI clusters scale toward 1.6T and 3.2T fabrics, 400G will serve as the critical bridge between the mature 100G ecosystem and the emerging 800G generation.

13.Conclusion

100G QSFP28 and 400G QSFP-DD serve different layers of the network and different generations of the access layer. The 100G module uses four 25G NRZ lanes and serves the aggregation layer for 25G access. The 400G module uses eight 50G PAM4 lanes and serves the spine layer and AI fabric for 100G access and GPU interconnect.

The 400G module delivers four times the bandwidth, comparable power per gigabit, and comparable or better cost per gigabit compared to its 100G counterpart. Its multimode reach is nearly identical to 100G, and its single-mode CWDM4/FR4 variants share the same wavelength grid and duplex LC interface. The 400G QSFP-DD port can break out into four 100G links, eight 50G links, or two 200G links, making it a more versatile aggregation interface than 100G QSFP28.

The decision framework is straightforward. If the access layer is at 25G and 100 Gbps per uplink is sufficient, 100G QSFP28 remains the most cost-effective and lowest-risk choice. If the access layer is at 100G, or if the spine layer is pushing past what 100G can carry, 400G QSFP-DD provides a direct upgrade path that reuses existing QSFP cages and cabling infrastructure. The form factor is backward compatible, the fiber plant is often compatible, and the migration can be phased by swapping modules as bandwidth demand grows.

As AI clusters scale and data center traffic increases, both interfaces will continue to be deployed in large volumes. The 100G/400G generation is the current transition in data center networking, and its physical compatibility ensures that 100G and 400G will coexist in the same fabric for years to come.

14.Q&A