100G QSFP28 Selection: SR4, LR4, CWDM4, PSM4

SR4, PSM4, CWDM4, LR4, ER4 - five variants of the same QSFP28 module, five different optical engines, and a procurement decision that gets more people in trouble than it should. The module housing is identical across all of them. The electrical interface (CAUI-4, formalized in IEEE 802.3bm-2015) is identical. What differs is the laser, the wavelength, the connector, and the fiber it needs. Get that part wrong and the link either won't come up or - worse - it comes up with errors you won't trace back to the optics for weeks.

 

 

QSFP28 Basics

Four electrical lanes, each running at roughly 25.78 Gbps, in a housing mechanically identical to 40G QSFP+. Power consumption sits below 3.5 W per module. A 1U switch can pack 36 or more QSFP28 ports, which is why the form factor killed off CFP and CFP2 for most 100GbE applications - those older packages drew 6–24 W each and took up far more faceplate space.

 

 

Variant Comparison at a Glance

Parameter	SR4	PSM4	CWDM4	LR4	ER4
IEEE / MSA Standard	IEEE 802.3bm 100GBASE-SR4	100G PSM4 MSA	100G CWDM4 MSA	IEEE 802.3ba 100GBASE-LR4	IEEE 802.3ba 100GBASE-ER4
Wavelength	850 nm	1310 nm	1271 / 1291 / 1311 / 1331 nm	1295.56 / 1300.05 / 1304.58 / 1309.14 nm	~1295–1310 nm (LAN-WDM)
Laser Type	VCSEL	DML	DML (DFB)	EML	EML + APD Rx
Fiber Type	OM3 / OM4 MMF	OS2 SMF	OS2 SMF	G.652 SMF	G.652 SMF
Connector	MTP/MPO-12	MTP/MPO-12	Duplex LC	Duplex LC	Duplex LC
Fiber Count per Link	8 (4 Tx + 4 Rx)	8 (4 Tx + 4 Rx)	2 (1 Tx + 1 Rx)	2 (1 Tx + 1 Rx)	2 (1 Tx + 1 Rx)
Max Reach	70 m (OM3) / 100 m (OM4)	500 m	2 km	10 km	40 km
Typical Power	~2.0 W	~2.5 W	~2.5 W	~3.5 W	~4.5 W
Relative Module Cost	Lowest	Low–Medium	Medium	Medium–High	High
Best Fit	Intra-rack, server-to-leaf	Cross-building (existing 8-fiber SMF trunks)	Inter-building, spine-to-spine ≤ 2 km	Campus / metro aggregation	Metro backbone, DR links

 

 

Topology: Where Each Variant Lands

Core / WAN (Metro backbone) ER4 - 40 km duplex LC, SMF LR4 - 10 km duplex LC, SMF Spine Switches (campus / multi-building fabric) CWDM4 - 2 km duplex LC, SMF CWDM4 / PSM4 500 m – 2 km, SMF Leaf Switch (Bldg A) Leaf Switch (Bldg B) SR4 - 100 m MTP/MPO, OM4 SR4 - 100 m MTP/MPO, OM4 Servers / Storage Servers / Storage

 

SR4 lives at the bottom - server-to-leaf, inside a single hall. CWDM4 or PSM4 handles the leaf-to-spine segment across buildings. LR4 covers spine-to-core at campus scale. ER4 for anything past 10 km.

 

SR4: Short-Reach Multimode

SR4 runs four parallel 850 nm VCSEL channels over OM3 or OM4 multimode fiber at 25.78125 GBd per lane (IEEE 802.3bm). MTP/MPO-12 connector, eight active fibers, 70 m reach on OM3 and 100 m on OM4. VCSELs are the cheapest laser in the QSFP28 family, multimode termination costs less than single-mode, and the Cisco QSFP-100G-SR4-S draws under 2.5 W. Not much to deliberate here - if your link is under 100 m and you have OM4 in the ground, SR4 is the obvious pick.

 

The PSM4 vs. CWDM4 Decision

This is where the actual procurement debate lives. Both PSM4 and CWDM4 target the 100 m–2 km range over single-mode fiber, and both exist because IEEE's original 100G standards left a gap - SR4 topped out at 100 m on multimode, and LR4 at 10 km on single-mode cost too much for a 300 m cross-building run. The PSM4 and CWDM4 MSAs were written specifically to fill that space, but they filled it in very different ways.

 

PSM4 is the parallel approach: four independent 1310 nm DML channels, each on its own fiber, through an MTP/MPO-12 connector. Eight fibers per link, 500 m maximum reach. CWDM4 is the wavelength-multiplexed approach: four 25 Gbps channels packed onto four coarse wavelengths (1271, 1291, 1311, 1331 nm per ITU-T G.694.2) sent through a single duplex LC connector. Two fibers per link, 2 km maximum reach, approximately 5.0 dB link budget per the CWDM4 MSA.

 

The module price on PSM4 is usually lower. But every PSM4 link eats eight fibers, and that changes the math fast. In a brownfield campus that already has 12- or 24-fiber SMF trunks terminated with MTP connectors, PSM4 is a clean upgrade from 40G QSFP+ - same cables, same patch panels, just swap the optics. That is a real advantage. But in a greenfield build, or anywhere running lean two-fiber LC patch panels between buildings, provisioning new eight-fiber MTP trunks adds hundreds of dollars per link that never show up on the transceiver line item. A 100G QSFP28 that costs $30 less per module but requires $400 more in cabling per link is not a savings.

 

CWDM4 avoids the whole fiber-count problem. Duplex LC patch cords are cheap. Most facilities already have two-fiber SMF runs left over from 1G or 10G deployments. And duplex LC on single-mode happens to be the same physical interface that 400G FR4 and DR4 optics use, so the fiber you light today for a 100G CWDM4 link carries 400G traffic in the next upgrade cycle without recabling. For any link between 100 m and 2 km where you do not already have MTP trunks in place, CWDM4 is almost always the lower total-cost option.

 

LR4: 10 km Campus and Metro

LR4 multiplexes four LAN-WDM channels (1295.56, 1300.05, 1304.58, 1309.14 nm per IEEE 802.3ba) onto a duplex LC single-mode connection. The tighter channel spacing requires EML transmitters - better extinction ratio, better chromatic dispersion tolerance than the DMLs in PSM4 and CWDM4. Cisco's QSFP-100G-LR4-S data sheet confirms this PHY works without FEC, no error correction overhead, clean multi-vendor interop. The premium over CWDM4 is significant, so LR4 only makes sense when your actual measured path exceeds 2 km: campus aggregation across a hospital complex or university, carrier hand-off at a colo, that sort of span. If the path is 1.8 km, buy CWDM4 and spend the difference on an OTDR characterization.

 

 

ER4 and ZR4

ER4 reaches 40 km with higher-power EML transmitters and APD receivers. 100GBASE ZR4 extends that to 80 km. Both use duplex LC on G.652 single-mode in the standard QSFP28 housing. Power runs 4–6 W. Metro backbone modules, not general-purpose data center optics.

 

Laser Types and Link Budget Math

The laser is what actually drives the cost and reach spread across the QSFP28 family, and understanding the differences changes how you evaluate data sheets.

 

SR4 uses VCSEL arrays. Low cost, low power, good coupling into multimode fiber, limited to 850 nm and short distances. PSM4 and CWDM4 use DML transmitters at 1310 nm - the injection current modulates the light directly, which introduces chirp (wavelength drift under modulation), but that stays tolerable across 500 m to 2 km of single-mode fiber. LR4 and ER4 step up to EML transmitters. An EML separates the laser from the modulator - an external electro-absorption layer modulates the output independently of the laser cavity, producing a cleaner optical eye with lower residual dispersion. That cleaner signal is what buys you 10–40 km of reach without relying on FEC.

 

Data sheets list a "maximum distance" number, but that number assumes a factory-clean fiber plant. Real plants have splice losses, connector insertion losses, patch panel attenuation, tight bends in cable trays. The numbers that actually tell you whether a link will work are transmitter output power, receiver OMA (Optical Modulation Amplitude) sensitivity, and the difference between them - the power budget. If the budget exceeds the measured loss of your fiber path, the link works. If it is marginal, you get elevated bit error rates that may or may not be correctable by FEC depending on how far over the threshold you are. Running an OTDR trace on each path before installing optics takes about an hour. That hour eliminates most Day 1 troubleshooting - the kind where you swap three transceivers before someone finally measures the fiber and finds a 1.5 dB splice loss nobody documented.

 

 

MTP/MPO Polarity - Worth Its Own Section

SR4 and PSM4 both use MTP/MPO-12 connectors, and polarity mistakes are responsible for a disproportionate number of 100G link failures that get misdiagnosed as bad optics.

 

The standard configuration for parallel-optics data center builds is Type-B (straight-through) polarity. Male connector on the transceiver side, female on the trunk. Elite-grade low-loss adapters on longer trunk runs - not standard-loss parts, which eat into the already-tight SR4 link budget. Getting any of this wrong does not necessarily kill the link. What happens more often is that traffic passes at low utilization but CRC errors spike under load. The symptoms mimic a marginal transceiver or a dirty connector, so the usual troubleshooting sequence is: clean the connector, no change; swap the optic, no change; swap the other end's optic, no change; finally someone pulls a visual fault locator, traces the fibers, and realizes the polarity is crossed. That sequence can burn an entire maintenance window. Verifying polarity during installation, not during troubleshooting, avoids the problem entirely.

 

For high-density deployments, MTP-based trunk and cassette systems help standardize polarity management across large numbers of connections. CWDM4, LR4, and ER4 avoid the whole issue - they use duplex LC connectors with UPC polish on OS2 single-mode fiber. APC ferrules are not compatible and will cause high return loss.

 

 

Breakout and Backward Compatibility

SR4, PSM4, and CWDM4 support 4×25G breakout mode - one 100G port split into four independent 25G connections through a breakout cable or MTP cassette. Useful for connecting 25G server NICs to a 100G leaf switch. Not every NOS release activates breakout on every port ASIC, so verify with the switch vendor's compatibility matrix. QSFP28 fits mechanically into QSFP+ ports but does not negotiate 100G against a QSFP+ module at the other end.

 

 

Future-Proofing

100G QSFP28 is not going anywhere soon. The optical transceiver market is pushing 400G and 800G at the spine, but 100G at the leaf and access tier still has years of runway in enterprise campuses, colo cross-connects, and mid-size data centers. The most useful thing you can do today is bias new fiber runs toward single-mode OS2 with duplex LC. That plant carries CWDM4 now, FR4 at 400G next cycle, and probably 800G after that. Multimode distance ceilings shrink with each speed generation. Anyone planning a 400G QSFP-DD migration in the next three years should be running single-mode everywhere it is practical to do so.