Engineers love a clean upgrade path, but the transition from QSFP28 to QSFP56 is a classic case of 'looks compatible, acts incompatible.' While both form factors share the identical 38-pin physical cage, four-lane electrical architecture, and I2C management interface, they are not electrically interchangeable in the way most buyers assume. The core difference lies in modulation: QSFP28 delivers 100 Gb/s using four lanes of 25 Gb/s NRZ, whereas QSFP56 doubles throughput to 200 Gb/s by switching to four lanes of 50 Gb/s PAM4. This single architectural shift dictates everything from cable reach to switch density and total cost of ownership.

The One-Way Street of Backward Compatibility

A common misconception is that because the cage is identical, the modules are swap-able. In reality, upgrades go one way only: from NRZ to PAM4, never back. A QSFP56 port can typically host a QSFP28 module and run it at 100G because the port's retimer or gearbox can clock down to 25G NRZ per lane. However, a QSFP28 port cannot make a QSFP56 module run at 200G because the host silicon lacks the PAM4 encoder and the switch ASIC was never specified for 50G per lane signaling. If you buy QSFP56 modules for a QSFP28 switch, you are buying expensive doorstops.

Cable Reach Collapses Under PAM4

This is the difference that bites procurement teams hardest. Engineers often assume existing cabling plans carry over unchanged, but they do not. In the NVIDIA LinkX family, 100G QSFP28 passive copper is specified up to 5 meters, while 200G QSFP56 passive copper is limited to 2.5 meters. PAM4 signaling burns the signal-integrity headroom that twinax relies on, effectively halving the passive DAC ceiling. If your racks are wired with 3-meter passive DACs at 100Gβ€”a standard in-rack practiceβ€”those exact cables will not be qualified for 200G. You must budget for active copper or active optical assemblies, or accept a reduced reach, which fundamentally changes rack layout strategies.

Density Gains and Hidden Costs

The primary driver for QSFP56 is density. A 32-port QSFP56 switch delivers 6.4 Tb/s in the same 1U chassis as a 32-port QSFP28 switch's 3.2 Tb/s. This allows spine roles to handle twice the bandwidth without doubling port counts or rack units. However, the cost model is counter-intuitive. While per-port costs are higher for QSFP56, the cost per gigabit of switching capacity is usually lower. The hidden costs appear in the host infrastructure: QSFP56 requires PCIe 4.0 x16 slots to avoid bottlenecks. Running a 200G switch with PCIe 3.0 servers means paying for silicon you cannot fully utilize. Furthermore, PAM4 links require Forward Error Correction (FEC) at the physical layer, and mismatched FEC settings are a top cause of 'light but down' ports.

Key Takeaways

  • Modulation is King: QSFP28 uses NRZ; QSFP56 uses PAM4. This dictates compatibility, not just speed.
  • Cable Reach Halves: Passive DACs drop from 5m (100G) to 2.5m (200G). Plan for active optics or shorter runs.
  • PCIe Dependency: 200G adapters need PCIe 4.0 x16. PCIe 3.0 slots will bottleneck the link.
  • FEC is Mandatory: PAM4 requires Reed-Solomon FEC. Both ends must agree on settings, or the link stays down.
  • No 25G Breakout: QSFP56 does not support a clean 200G-to-25G breakout, forcing mixed-layer designs for 25G-dense racks.

The Bottom Line

Do not upgrade to QSFP56 unless you have PCIe 4.0 hosts and a fiber plant that can survive PAM4's tighter margins. For most estates, the smart move is a two-tier design: keep QSFP28 for 25G-dense leaf layers and reserve QSFP56 for the spine where density actually pays.