Technical specifications of the 100G QSFP28 SR4 product category

Of the four mainstream routes to 100G, SR4 is technically the dumbest. No wavelength division, no multiplexing - just four fiber pairs stuffed side by side into one module. It is also the cheapest, the most power-efficient, and the highest-volume. That kind of "dumb" is exactly what makes a design so hard to displace.

This article discusses the generic technical specifications of the 100G QSFP28 SR4 product form factor. Figures are taken from a representative 100G QSFP28 SR4 datasheet and contain no specific model or brand information. They can serve as a technical reference for comparable products.

1. The Landscape: Four Roads to 100G

100G optics have converged on four mainstream routes. They differ entirely in which fiber they use and how they multiplex:

The distance boundaries between these four are remarkably clean. Below 100 meters, SR4 has essentially no rival.

The reason is economics, not optics. Distances inside a rack and between adjacent racks are typically 3–30 m, at most 100 m, and an 850 nm VCSEL is more than adequate over that span. Using a single-mode solution for this segment is like building a highway to reach your garage - technically valid, economically absurd.

2. Why Parallel, Not Wavelength Division

This design choice deserves unpacking.

The core diameter of multimode fiber (OM3/OM4) is 50 µm - more than six times that of single-mode. This brings two consequences: relaxed coupling tolerance, which makes connectors cheap; and modal dispersion, where different modes travel at different speeds and spread the pulse as distance grows.

Modal dispersion closes off the WDM route on multimode fiber:

• Dense WDM is not viable. Putting 4-wavelength CWDM on multimode would require a narrowband filter and a wavelength-stabilized laser per channel, while modal noise in a multimode system rapidly degrades the signal-to-noise ratio of any narrowband scheme.

• Component cost inverts. Multimode WDM components have far lower volumes than their single-mode counterparts, so unit prices are actually higher - wiping out the cost advantage of multimode fiber and connectors.

Engineering therefore took the other path: rather than packing multiple wavelengths into one fiber, run more fibers and put one wavelength on each.

4 lanes × 25.78125 Gb/s = 103.125 Gb/s, which carries 100GbE after line-coding overhead. Each lane needs only one 25G VCSEL and one 25G PIN photodiode - both of which reached high-volume production during the 40G era and have extremely thin cost bases.

The price is fiber count: a duplex 100G module needs 2 fibers, SR4 needs 8. Inside a rack that price is cheap, because pre-terminated MPO trunk cables have long been the standard cabling method in data centers - one 12-fiber bundle costs far less than twelve individual patch cords.

3. Inside the Implementation

3.1 Optical Architecture: 4 Transmit, 4 Receive, 8 of 12 Fibers Used

Internally the module contains two fully symmetric paths:

TX0~TX3 --> CDR array --> VCSEL driver array --> 850nm VCSEL array (4) --+

+--> MPO12 (12 fibers, 8 used)

RX0~RX3 <-- CDR array <-- TIA array <-- 25Gb/s PIN array (4) <--+

^

MCU (I2C / SCL / SDA)

Two points deserve attention:

• Only 8 of the 12 MPO fibers are used: 4 transmit, 4 receive, 4 idle. This is also why 100G SR4 can be fanned out into 4 × 25G through an MPO breakout - each 25G lane occupies exactly one transmit/receive pair.

• Each direction has its own 4-channel CDR bank. The CDR retimes the eye and suppresses jitter, reshaping the CAUI-4 signal from the host before it drives the VCSEL. This is also central to hitting 2.5 W: the more integrated the CDR, the fewer peripheral components are needed.

3.2 The 850 nm VCSEL: The Optimal Answer for Short Reach

A VCSEL (vertical-cavity surface-emitting laser) is structurally unlike a conventional edge-emitting DFB: its cavity is perpendicular to the chip surface, and light exits vertically from the top.

That geometry yields several properties that are extremely friendly to short-reach applications:

• Circularly symmetric beam, naturally matched to the round core of multimode fiber, giving high coupling efficiency with no complex beam shaping.

• Low threshold current, typically in the milliamp range, so power consumption stays low even when directly modulated at 25G.

• Wafer-level testability, allowing screening across the entire wafer. Unit cost is far below edge emitters, which require cleaving, coating, and packaging.

• High reliability, with a simpler failure mechanism in the GaAs material system and lifetimes typically measured in millions of hours.

The datasheet specifies an RMS spectral width limit of 0.6 nm - far wider than a DFB, yet completely harmless for multimode transmission, because the limiting factor there is modal dispersion, not chromatic dispersion.

3.3 Transmitter: Why Extinction Ratio Is Only 2 dB

This is what confuses most engineers on first reading an SR4 datasheet: LR4 requires ER ≥ 4 dB, while SR4 requires only ≥ 2.0 dB.

The reason is again the VCSEL. Its P-I curve is very steep near threshold, so achieving a high extinction ratio means pushing the "0" level close to threshold. That causes two problems: modulation speed drops (carrier recovery takes longer) and relaxation oscillation on the lower edge of the eye intensifies, which actually worsens TDEC.

802.3bm therefore defined a separate, looser extinction-ratio requirement for multimode parallel optics, and instead constrains transmit quality with TDEC (Transmitter and Dispersion Eye Closure), a composite metric. The datasheet caps TDEC at 4.3 dB and separately requires that "launch power in OMA minus TDEC" be no lower than −7.3 dBm. That latter figure is what actually determines whether a link closes.

Principal transmitter specifications:

3.4 Receiver: A 12.7 dB Window Between Sensitivity and Overload

Three numbers matter most on the receive side:

• Average receive power range: −10.3 dBm to +2.4 dBm, a dynamic range of 12.7 dB.

• Stressed receiver sensitivity (SRS, OMA basis): ≤ −5.2 dBm. This is sensitivity under stress, with vertical eye closure, jitter, and adjacent-channel crosstalk deliberately injected. It sits 3–4 dB below unstressed sensitivity and is much closer to real operating conditions.

• Damage threshold: 3.4 dBm. Exceeding this may permanently damage the photodiode. During commissioning, shining a live source directly into a module can easily cross this line.

Two more parameters are easily overlooked:

• Receiver reflectance ≤ −12 dB. If a multimode connector reflects excessively, it induces relative intensity noise in the VCSEL, which shows up as a bit error rate that changes each time a connector is mated.

• LOS thresholds: assert at −30 dBm, de-assert at −12 dBm, 0.5 dB hysteresis. The wide 18 dB hysteresis is deliberate: it prevents the LOS signal from toggling repeatedly when input power dithers near threshold, which would otherwise trigger repeated alarms upstream.

3.5 Link Budget: Spending 1.9 dB

This is SR4's most counterintuitive number.

Using averages: −8.4 dBm (minimum launch) minus −10.3 dBm (minimum receive) leaves a link budget of only 1.9 dB.

1.9 dB would be nearly unusable on a single-mode system, but it is sufficient over 100 m of multimode:

Note that the cable itself accounts for less than 20% of the budget; the rest is consumed by connectors and modal dispersion.

This also explains the most common field problem with SR4: it is rarely a failed module, it is almost always a dirty MPO endface. A single fingerprint on an MPO ferrule can easily contribute more than 1 dB of excess loss - half the entire margin.

3.6 Reach: 70 m on OM3, 100 m on OM4 - and FEC Is Mandatory

The datasheet reach is 70 m on OM3, 100 m on OM4, and FEC is required.

Why FEC? Because 25.78 Gb/s is already pushing multimode fiber to its bandwidth limit. OM3 has an effective modal bandwidth (EMB) of roughly 2000 MHz·km and OM4 about 4700 MHz·km; at the 100 m scale, the 25G eye is visibly closed. 802.3bm acknowledges this and uses RS(544,514) KP4 FEC to pull the bit error rate from the 10⁻⁵ range back to 10⁻¹².

Hence a hard rule in practice: 100G SR4 links must have FEC enabled. Links with FEC disabled generally will not run reliably. This is the root cause of many cases where swapping modules does nothing but enabling FEC makes the link come up.

3.7 Electrical Interface

The electrical interface follows the standard QSFP28 definition:

• CAUI-4: four 25.78125 GBd differential lanes, one TX and one RX pair each, CML levels.

• 38-pin card edge: four high-speed TX/RX differential pairs, I2C (SCL/SDA), ModSelL, ResetL, LPMode, ModPrsL, IntL, and Vcc1 / VccTx / VccRx 3.3 V rails.

• Supply range 3.14–3.46 V, nominal 3.3 V.

• Power ceiling 2.5 W, operating case temperature 0 to +70 °C.

The 2.5 W figure deserves a note. On a 48- or 64-port 100G switch, 2.5 W × 64 = 160 W - already close to a third of total chassis power. SR4 holds 2.5 W thanks to the VCSEL's low drive voltage and a highly integrated 4-channel CDR, which is the most substantive advance over early 100G modules.

3.8 Reliability and Compliance

ESD: 1 kV on high-speed data pins, 2 kV on all other electrical pins, tested per MIL-STD-883 Method 3015.4 / JESD22-A114-A (HBM).

Laser safety: Class 1, compliant with EN 60825-1:2014 and 21 CFR 1040.10 / 1040.11.

Environmental: RoHS compliant.

Mechanics: approximately 121.5 mm long, 18.35 mm wide, 8.50 mm high, per the QSFP28 form factor.

4. Typical Applications

4.1 In-Rack: ToR to Server

This is where SR4 ships in the greatest volume. A 32-port 100G top-of-rack switch connects down to 32 servers over MPO trunk cable, each server using 100G SR4 or an MPO breakout to 4 × 25G. Distances are typically 5–30 m, and the 1.9 dB budget is more than sufficient.

4.2 Adjacent Racks: Leaf-Spine Horizontal Links

East-west traffic between spine and leaf is the dominant flow in a data center. Within a single pod or column, leaf-spine spans are generally under 50 m, and SR4 is the standard configuration.

4.3 Fan-Out from 100G to 4 × 25G

Because only 8 of the 12 MPO fibers are used, SR4 is naturally suited to fan-out: one 12-fiber MPO trunk with an SR4 at each end, split in the middle by an MPO-to-4×LC breakout harness into four 25G links feeding 25G switches or server NICs. This makes mixed-rate "100G backbone, 25G access" networks very simple to build.

4.4 Storage and AI Training Clusters

Node-to-node interconnect in distributed storage and GPU training clusters demands short reach, high port density, and low latency. Here SR4's low power and small footprint matter far more than long-distance capability.

5. Deployment Notes

Ordered by how often they cause trouble in the field:

1. Verify MPO polarity. Type A (straight), Type B (crossed), and Type C (pairwise crossed) are not interchangeable; mixing them misaligns transmit and receive fibers. The symptom is a link that is completely dead rather than erroring. Standardize on one polarity scheme across the site and label it.

2. Endface cleanliness is the number one failure source. One contaminated MPO affects all 12 fibers at once, and the large multimode core is especially sensitive. Clean with a dedicated tool before every mating and keep endface inspection records.

3. FEC must be enabled and consistent at both ends. One end on and one end off guarantees the link will fail.

4. Bend radius. Multimode bend loss at 850 nm is far more sensitive than at 1310 nm; right-angle bends in rack cable managers are a common hidden hazard.

5. Connect DDM monitoring. Reading per-lane transmit power, receive power, temperature, and bias current over I2C is the only way to catch VCSEL aging early. A spread of more than 1.5 dB between the four lanes warrants investigation.

6. Control case temperature. 0 to +70 °C refers to case temperature, not ambient. In high-density racks module cases typically run 15–20 °C above inlet air, so design against measured case temperature.

6. Where the Boundary Lies

SR4's boundary is unambiguous: 100 meters.

Beyond that, a single-mode approach is required, and there are three options:

• PSM4: 4-lane parallel single-mode at 1310 nm, 500 m, MPO12 - the same cabling model as SR4, suited to cross-row or cross-room runs.

• CWDM4: 4-wavelength CWDM single-mode, 2 km, duplex LC - the most fiber-efficient option.

• LR4: 4-wavelength LAN-WDM single-mode, 10 km, duplex LC.

At this crossover the cost structure shifts abruptly: VCSEL becomes DFB, MPO becomes duplex LC, multimode becomes single-mode fiber. This is also why data center cabling is typically built in layers - "multimode inside the rack, single-mode in the backbone" - with each physical layer covering the segment it serves most economically.

7. Conclusion

SR4's technical merit lies not in being advanced, but in being exactly right.

It gives up wavelength division, gives up long reach, and gives up fiber efficiency, in exchange for the lowest component cost, 2.5 W of power, and full compatibility with the installed MPO cabling base. In an environment where a single rack holds thousands of 100G ports and that number doubles every three years, any one of those three advantages matters far more than "can reach 10 km."

The key to understanding SR4 is not memorizing those dBm figures, but understanding that 1.9 dB budget line. It explains why SR4 is cheap, why it stops at 100 m, and why half of all field failures are dirty endfaces. One budget line ties together the entire engineering logic of the product.