d617524f10
Topics: CWDM4/PSM4, MSA compliance, DAC/AOC TCO, grey vs DWDM, ESD damage, tunable DWDM, FEC deep-dive, CPO hype cycle, CMIS 4.0, vendor evaluation. Ø 1,180 words each.
25 lines
7.0 KiB
Markdown
25 lines
7.0 KiB
Markdown
---
|
|
title: "CWDM4 vs PSM4 for 100G: Why the Four-Wavelength Decision Matters More Than You Think"
|
|
type: comparison
|
|
target_audience: technical
|
|
score: 9/10
|
|
---
|
|
|
|
The 100G QSFP28 market bifurcated cleanly along two lines when IEEE 802.3bm ratified in 2015: CWDM4 and PSM4. Both deliver 4x25G lanes over SMF to 500m, both land at roughly the same optical reach, and at a glance both seem interchangeable for the same cabling run. They are not. The decision between them compounds across thousands of ports in a real data center build, and getting it wrong means either pulling fiber or throwing away optics, neither of which is cheap.
|
|
|
|
PSM4 — Parallel Single Mode 4-lane — is conceptually the simplest architecture imaginable. Four 25G lanes each travel over a separate single-mode fiber at approximately 1310nm (the individual lane wavelengths are not tightly controlled since they don't need to be wavelength-division multiplexed), with all four using NRZ modulation at 25.78125 Gbps per lane. The connector is an MPO-12 (the outer 4 fibers on each side unused), which means every PSM4 link consumes eight fiber strands. This is the critical arithmetic: a 48-port leaf switch with PSM4 uplinks requires 192 individual fibers just for the uplinks. In a spine-leaf fabric with 10,000 server-facing 25G ports and 400 100G uplinks, PSM4 alone demands 3,200 strands of single-mode fiber between the layers. The Senko MPO connectors on production PSM4 modules — as used in Innolight TR-FC13J-NCD or Flexoptix P.10741 — have a mechanical life of roughly 500 insertion cycles before ferrule wear degrades the contact geometry enough to affect loss budget.
|
|
|
|
CWDM4 takes those same four 25G lanes and wavelength-division multiplexes them onto two fibers using four distinct center wavelengths: 1271nm, 1291nm, 1311nm, and 1331nm, with 20nm channel spacing. The two fibers are LC-duplex, which is the same connector your existing 10G and 40G plant almost certainly uses. The mux/demux is done with thin-film filter arrays inside the module itself. Each lane has its own CDR (Clock and Data Recovery) circuit, which is why CWDM4 modules burn approximately 3.5W versus PSM4's 2.5W — an additional 1W per module that, across a 10,000-port fabric, adds up to 10kW of additional cooling load. Flexoptix P.10733 and the Finisar FTLC1152RDNM are representative production examples. The CDR also introduces approximately 100ps of additional lane-to-lane deskew processing, though this is irrelevant for Ethernet since 802.3bm Clause 87 allows up to 120ns of skew between lanes.
|
|
|
|
The cost differential has narrowed considerably from 2017 highs when CWDM4 modules cost nearly three times PSM4, but a material gap remains. In volume pricing as of early 2026, compatible CWDM4 QSFP28 modules from a quality vendor like Flexoptix or ProLabs land at approximately €180-220 per unit, while PSM4 equivalents are €120-150. On a 400-port spine layer that is a €24,000 to €28,000 difference just in optics. That number must be weighed against fiber plant cost: an 8-fiber MPO trunk cable costs roughly 40% more than a 2-fiber LC-duplex equivalent for the same run length, and MPO cassettes for breakout add another €15-25 per port of termination cost. The crossover point where PSM4's cheaper optics are eaten by higher fiber plant costs typically occurs around the 200-300 port threshold for new greenfield builds where fiber is being installed anyway.
|
|
|
|
For brownfield environments, CWDM4 almost always wins on economics even at its optics premium. Any data center built after 2010 has LC-duplex SMF infrastructure to every cabinet. Pulling new 8-fiber MPO trunks to replace 2-fiber LC runs costs €8-15 per meter in installation labor plus materials, so a 50-meter average run to 400 switch ports is €160,000-300,000 in fiber plant costs before a single PSM4 module is purchased. The CWDM4 optics premium of €70 per module times 400 modules is €28,000 — a trivial fraction.
|
|
|
|
The interoperability risk that gets overlooked in vendor comparisons is connector polarity. PSM4 uses Type B MPO polarity (per TIA-568-C.3), meaning the fiber labeled 1 at one end connects to fiber 1 at the other. A Type A MPO cassette — the most commonly pre-installed type in legacy data centers — crosses the fibers, which will work fine for 40G QSFP+ where both ends use MPO, but PSM4 QSFP28 requires methodical polarity management. Plugging a PSM4 module into an incorrectly polarized MPO plant is a non-obvious failure: the module will power on, DOM will show nominal TX power on all four lanes at the transmitting end, but the far end will show either zero RX power or a scrambled fiber-to-lane mapping that produces persistent bit errors. Field engineers unfamiliar with PSM4 will spend 45 minutes inspecting the optics before realizing the MPO cassette orientation is wrong.
|
|
|
|
Platform support nuances also favor CWDM4 in heterogeneous environments. Cisco Nexus 9332C and 93180YC-FX both support CWDM4 and PSM4, but the 9200 series requires a firmware upgrade to enable PSM4 auto-negotiation correctly, and Juniper QFX5120-48Y had a known bug in Junos 20.2R1 where PSM4 modules would intermittently fail to come up after a port flap until the bug was addressed in 20.2R3. CWDM4 with its LC-duplex interface is electrically and mechanically simpler from the platform's perspective — the transceiver looks and behaves more like a conventional duplex interface, which means fewer edge cases in NOS port drivers.
|
|
|
|
The decision framework is straightforward once you quantify the numbers. For new hyperscale builds where leaf-to-spine cabling is being installed from scratch, PSM4 saves real money at scale when the fabric exceeds roughly 500 ports per tier. For enterprise data centers operating on existing LC-duplex SMF plant, any calculation that ends with pulling and replacing fiber plant for PSM4 should be rejected — CWDM4 at its optics premium is the rational choice. For inter-building runs where the fiber plant is OS2 single-mode but the connectors are already MPO for 40G migration, PSM4 is worth evaluating only if you have verified Type B polarity throughout. Mixed environments — where some switches use CWDM4 and some PSM4 — require optical-to-electrical breakout panels at the connection point, since you cannot directly couple a CWDM4 module to a PSM4 module regardless of the fiber plant. These modules are not optically compatible, full stop.
|
|
|
|
One final consideration: CWDM4 gives you a more credible upgrade path to 400G CWDM4 (100G per lane, 4 lanes on the same 1271/1291/1311/1331nm wavelength plan per IEEE 802.3bs Clause 87), meaning your fiber plant investment carries forward. PSM4 fiber infrastructure does the same job for 400G-DR4 (IEEE 802.3bs Clause 124), but DR4 requires OS2 with 0.2dB/km loss specification and highly polished MPO connectors, not the generic OM3/OM4 that 40G PSM4 sometimes ran on with margin to spare. If your 10-year fiber plant investment needs to justify both present-day 100G and future 400G density, the wavelength route with LC-duplex is the lower-risk architectural bet.
|