Fibre in the data centre: OM4, OM5 and OS2

Jamal already knows how optical fibre works from LVT-210: light guided in a glass core by total internal reflection, multimode fibres with large cores and many paths for the light, single-mode fibres with tiny cores and one path, attenuation in decibels per kilometre and the colours of the jackets. This lesson does not repeat that. It is about which fibres a data centre uses, why, and what that means for the links a technician installs. Nadia begins with a question: why does Hall A have orange, aqua and lime-green cords in its panels, Hall B mostly aqua and yellow, and Hall D almost nothing but yellow?

The fibre types in the halls

The table tells a story. When Hall A was built, short server links ran at 10 Gb/s over multimode fibre, because multimode transceivers were far cheaper than single-mode ones. As speeds climbed, the reach of multimode links shrank and single-mode transceivers got cheaper, until for 400 and 800 Gb/s many designers chose single-mode for everything longer than a rack or two. Hall D, designed in 2024, uses single-mode for all of its optical links. Hall A will carry its mixture of fibre types for as long as its customers keep their equipment.

Why multimode reach shrinks

A multimode fibre carries light along many paths of slightly different lengths, so a pulse of light spreads out as it travels. This modal dispersion limits how fast the pulses can follow each other over a given distance, which is described by the fibre's modal bandwidth. OM3 and OM4 have their core's refractive index carefully graded so that the paths arrive almost together, and OM4 is graded more precisely than OM3, which is why it reaches farther at the same speed. At higher speeds, each pulse is shorter, so the same spreading matters more and the reach drops.

Multimode links in data centres use 850 nm vertical-cavity lasers, VCSELs, which are cheap and efficient. Faster multimode links get their speed by running several lanes in parallel, one fibre pair per lane, rather than by running one lane faster. A 400GBASE-SR4 link, for example, uses four lanes of 100 Gb/s on four fibres in each direction. OM5 was designed to carry several wavelengths between 850 and 950 nm on each fibre, so that the lanes can share fibres instead, but it has seen limited use because single-mode took over the longer links.

# Typical maximum reach on multimode and single-mode fibre (metres), for planning only
reach = {
    "10GBASE-SR":   {"OM3": 300, "OM4": 400, "OM5": 400},
    "100GBASE-SR4": {"OM3": 70,  "OM4": 100, "OM5": 100},
    "400GBASE-SR4": {"OM3": 70,  "OM4": 100, "OM5": 100},
    "800GBASE-SR8": {"OM3": 60,  "OM4": 100, "OM5": 100},
    "800GBASE-VR8": {"OM3": 30,  "OM4": 50,  "OM5": 50},
}
single_mode = {"400GBASE-DR4": 500, "800GBASE-DR8": 500, "400GBASE-FR4": 2000, "400GBASE-LR4": 10000}
for app, row in reach.items():
    print(f"{app:14}", "  ".join(f"{k} {v:>5} m" for k, v in row.items()))
for app, v in single_mode.items():
    print(f"{app:14} OS2 {v:>5} m")

10GBASE-SR     OM3   300 m  OM4   400 m  OM5   400 m
100GBASE-SR4   OM3    70 m  OM4   100 m  OM5   100 m
400GBASE-SR4   OM3    70 m  OM4   100 m  OM5   100 m
800GBASE-SR8   OM3    60 m  OM4   100 m  OM5   100 m
800GBASE-VR8   OM3    30 m  OM4    50 m  OM5    50 m
400GBASE-DR4   OS2   500 m
800GBASE-DR8   OS2   500 m
400GBASE-FR4   OS2  2000 m
400GBASE-LR4   OS2 10000 m

The reaches are the typical figures from the IEEE Ethernet standards and the transceiver makers; the exact limit for a given link depends on its loss and the number of connections, which Lesson 11 covers. The pattern is what matters: at 10 Gb/s multimode reaches across a hall, at 400 and 800 Gb/s it reaches across a row or two, and single-mode reaches the far end of the campus.

800 Gb/s optics

The IEEE 802.3df amendment, published in 2024, defined Ethernet at 800 Gb/s over eight lanes of 100 Gb/s: 800GBASE-SR8 and VR8 on multimode, DR8 on single-mode to about 500 m (1,640 ft) and DR8-2 to about 2 km (1.2 mi), as well as copper and backplane versions. Some 800 Gb/s optics are instead built as two 400 Gb/s engines in one module, for example 2×400G-FR4, which use two duplex connectors and can talk to two separate 400 Gb/s ports. Work on 200 Gb/s per lane, which will bring 1.6 Tb/s links, continues in the IEEE 802.3dj project, and the first such optics are already appearing in the newest AI equipment.

For Jamal, the details of the electronics matter less than the connector on the front of the module, because that decides what he plugs in:

• Duplex LC: one fibre pair, used by single-wavelength 10 and 25 Gb/s links and by multiplexed links such as 400G-FR4 and LR4.
• MPO-12: twelve positions in one row, of which a 400G-SR4 or DR4 link uses eight, four to transmit and four to receive.
• MPO-16: sixteen positions in one row, the connector chosen for eight-lane 800 Gb/s links such as DR8 and SR8, which use all sixteen.
• Dual MPO-12 or dual duplex connectors on twin-port modules, which carry two independent links.

Corvid's backend in Hall D runs 800 Gb/s links on single-mode fibre with MPO-16 connectors at both ends, and its frontend on 2×400 Gb/s twin-port modules. Lesson 6 is about the multi-fibre connectors.

Single-mode in the data centre

Single-mode fibre in data centres is almost always OS2, a low-water-peak fibre with low loss across the 1260 to 1625 nm window. Data centre single-mode links use two kinds of optics. Parallel single-mode, such as DR4 and DR8, uses one wavelength near 1310 nm and one fibre per lane per direction, on MPO connectors. Wavelength-multiplexed single-mode, such as FR4 and LR4, puts four wavelengths on one fibre pair with duplex connectors. Parallel optics are cheaper per bit for short links but use more fibre; multiplexed optics save fibre at a higher price per module. Hall D's backend is parallel because its links are short and fibre inside a hall is cheap compared with the optics.

Single-mode connectors in Hall D are angled physical contact, APC: the end face of each ferrule is polished at about 8 degrees so that light reflected at the joint leaves the core instead of returning to the laser. APC connectors are usually green and must never be mated with flat-polished UPC connectors, which are usually blue on single-mode, because the mismatched faces leave an air gap and can be damaged. Many parallel single-mode transceivers require APC MPO connectors; multimode MPO connectors are normally flat-polished. Nadia makes Jamal say the polish out loud every time he picks up a cord in Hall D.

Bend-insensitive fibre and small cables

Data centre cables are packed tightly and bent around managers, so nearly all fibre installed since about 2015 is bend-insensitive: multimode fibres with a trench around the core, and single-mode fibres meeting the G.657.A1 or A2 bend classes. They lose much less light in tight bends than older fibres. That does not make bend radius irrelevant. Bend-insensitive fibre still breaks if it is kinked, crushed or pinched by a hook-and-loop strap pulled too tight, and the cable around it still has a minimum bend radius set by its maker. The LVT-210 rules for bend radius still apply; what changes is that a bend that would have cost a dB in an old cable may now cost almost nothing, so a bad bend may not show on a loss test until the fibre cracks.

Data centre fibre cables are also thinner than building cables. Patch cords use 1.6 or 2.0 mm (0.063 or 0.079 in) jackets, uniboot duplex cords put both fibres in one jacket, and high-count trunks use ribbon or rollable ribbon construction to fit hundreds of fibres into a cable the size of a finger.

Estimating the loss of a link

Before a link is installed, a technician should know roughly what its loss will be, so that a test result can be judged. The rule from LVT-210 is that the loss is the fibre's attenuation times its length, plus the loss of each mated connector pair and each splice. Data centre links are short, so the connectors dominate.

# Rough loss of three Hall links: fibre plus connections (planning values)
links = [
    ("A-31-03 to A-HDA-31, OM4 at 850 nm", 0.038, 3.5, 2, 0.5),
    ("B-12-09 to B-NET-04, OS2 at 1310 nm", 0.052, 0.5, 2, 0.35),
    ("D-3-07 to D-4-02, OS2 MPO-16", 0.061, 0.5, 2, 0.35),
]
for name, km, db_per_km, pairs, per_pair in links:
    fibre = km * db_per_km
    conn = pairs * per_pair
    total = fibre + conn
    print(f"{name}: fibre {fibre:.2f} dB + {pairs} pairs {conn:.2f} dB = {total:.2f} dB  (length {km * 1000:.0f} m)")

A-31-03 to A-HDA-31, OM4 at 850 nm: fibre 0.13 dB + 2 pairs 1.00 dB = 1.13 dB  (length 38 m)
B-12-09 to B-NET-04, OS2 at 1310 nm: fibre 0.03 dB + 2 pairs 0.70 dB = 0.73 dB  (length 52 m)
D-3-07 to D-4-02, OS2 MPO-16: fibre 0.03 dB + 2 pairs 0.70 dB = 0.73 dB  (length 61 m)

On every link, the fibre itself contributes a few hundredths of a decibel and the two mated pairs contribute nearly all the loss. That is why connector quality and cleanliness, the subjects of Lessons 6 and 8, matter so much more in a data centre than the length of the run.

Fibre counts and future speeds

Cabling is meant to outlast several generations of equipment. Hall A's 2014 multimode plant has already seen servers move from 10 to 25 Gb/s, and some of its customers would now like 100 or 400 Gb/s links that the plant cannot carry far enough. That history is why Stonebridge's newer halls are cabled with single-mode trunks of high fibre count, often more than the first tenant needs. Single-mode fibre has no modal bandwidth limit at data centre distances, so the same strands that carry 800 Gb/s today will carry the next generation of optics, and spare strands in each trunk can be lit when a tenant adds links. Nadia's rule of thumb is that the trunks are installed for ten or fifteen years, the cords for three or four and the transceivers for less, so the trunks are the place to be generous.

For technicians, the practical result is that the fibre type of a trunk is a decision made once, at design time, and recorded in the trunk schedule. When a ticket asks for a link of a given speed and length, the technician checks that the trunk serving it is of a type and count that can carry it, and if not, the request goes back to the cabling lead rather than being forced onto the wrong fibre.