Optics and transceivers: form factors and speeds

Almost every network port Jamal turns up holds an optic: a small module, a transceiver, that slides into a cage on the switch and turns the switch's electrical signals into light on a fiber and back again. Optics are the part of the network a technician handles most, the part that fails most often, and one of the largest costs in an AI fabric. A single 800 Gb/s optic costs as much as a good laptop, and Corvid's pod in Hall D holds thousands of them. DCT-210 covered the fibers, connectors and loss budgets the optics plug into. This lesson covers the modules themselves: their shapes, their names and how to tell, before plugging one in, whether it is the right one.

Lanes and speeds

An optic's speed is built from lanes. Each lane is one electrical channel from the switch chip and, usually, one optical channel on its own fiber or wavelength. A 100 Gb/s optic of a few years ago used four lanes of 25 Gb/s; a 400 Gb/s optic uses eight lanes of 50 Gb/s or four of 100 Gb/s; an 800 Gb/s optic uses eight lanes of 100 Gb/s. The next generation runs 200 Gb/s per lane: the IEEE 802.3dj project defines 200, 400 and 800 Gb/s and 1.6 Tb/s Ethernet on 200 Gb/s lanes, and it was in its final drafting stages during 2026, with 1.6 Tb/s optics already shipping in early deployments. Lane counts matter on the floor because they decide how a port can be broken out and how many fibers a link uses.

Faster lanes need stronger error correction. Links at 50 Gb/s per lane and above run forward error correction, FEC, which adds check data to the signal so the receiver can repair a certain number of bit errors without asking for anything to be sent again. Both ends of a link must use the same FEC mode, a detail that Lesson 11 returns to, because a mismatch keeps a link down even when everything physical is perfect.

Form factors

A form factor is the module's shape and its electrical connection to the switch. The switch's cages accept only their own form factor, so the form factor is the first thing to check.

QSFP-DD, the double-density QSFP, keeps the width of the older QSFP but adds a second row of electrical contacts for eight lanes, and a QSFP-DD cage will usually accept older QSFP modules. OSFP is slightly wider and longer, has its own heat sink on top and is designed for higher power; an OSFP cage does not accept QSFP modules at all. Corvid's GPU servers and switches use OSFP, including twin-port modules that carry two independent 400 Gb/s links in one 800 Gb/s module, each with its own connector on the front.

Reading the name

An optic's name tells most of what a technician needs. The standard Ethernet names follow a pattern: the speed, the letters GBASE, and a suffix for the reach and the medium. The suffix is the part to read.

So 400GBASE-DR4 means 400 Gb/s on four parallel single-mode fibers in each direction, an MPO connector and about 1,640 ft of reach, and 2x400G-FR4 in an OSFP means two separate 400 Gb/s links, each on its own duplex connector, reaching about 1.2 mi. The connector type matters as much as the speed: a DR4 optic will not mate with a duplex LC cord, and an FR4 optic will not mate with an MPO trunk. DCT-210 Lesson 5 covered the fiber types and the polarity each kind of link needs.

# Choose an optic type from the link's speed and route length (typical values, meters)
options = [  # (name, speed Gb/s, medium, max reach m)
    ("400G-SR4", 400, "multimode MPO", 100), ("400G-DR4", 400, "single-mode MPO", 500),
    ("400G-FR4", 400, "single-mode duplex", 2000), ("800G-SR8", 800, "multimode MPO", 100),
    ("800G-DR8", 800, "single-mode MPO", 500), ("2x400G-FR4", 800, "single-mode duplex x2", 2000),
    ("800G AEC", 800, "copper cable", 7),
]
TO_UNIT, UNIT = 3.281, "ft"
def choose(speed, route_m):
    fits = [o for o in options if o[1] == speed and o[3] >= route_m]
    return min(fits, key=lambda o: o[3]) if fits else None
for speed, route in [(800, 4), (800, 38), (800, 310), (400, 85), (800, 1300)]:
    o = choose(speed, route)
    r = route * TO_UNIT
    print(f"{speed}G over {r:,.0f} {UNIT}: " + (f"{o[0]} ({o[2]}, up to {o[3] * TO_UNIT:,.0f} {UNIT})" if o else "no standard option"))

800G over 13 ft: 800G AEC (copper cable, up to 23 ft)
800G over 125 ft: 800G-SR8 (multimode MPO, up to 328 ft)
800G over 1,017 ft: 800G-DR8 (single-mode MPO, up to 1,640 ft)
400G over 279 ft: 400G-SR4 (multimode MPO, up to 328 ft)
800G over 4,265 ft: 2x400G-FR4 (single-mode duplex x2, up to 6,562 ft)

The rule in the code is the rule engineers use: the shortest-reach option that covers the route, because shorter-reach optics are cheaper and use less power. The technician's part is to check that what arrives in the box matches what the design chose, and that the cabling at the port is the kind the optic needs.

Breakouts

A port with several lanes can sometimes be split into several slower ports, a breakout. A 400G-DR4 optic's four lanes can serve four 100 Gb/s servers through an MPO-to-duplex breakout cable or cassette; an 800G port can run as two 400G or eight 100G ports. The switch must be configured for the breakout, and the far ends must be the right optics, so breakouts appear on tickets as a set: the optic, the breakout cable or cassette, and the port names the switch gives to each lane.

# What a port can break out into, and the port names the switch will show
ports = {"Ethernet1/9": (800, 8), "Ethernet1/10": (400, 4)}   # speed Gb/s, lanes
for port, (speed, lanes) in ports.items():
    per_lane = speed // lanes
    print(f"{port}: {speed}G on {lanes} lanes of {per_lane}G")
    for n in (1, 2, 4, 8):
        if n <= lanes and lanes % n == 0:
            names = ", ".join(f"{port}/{k}" for k in range(1, n + 1)) if n > 1 else port
            print(f"   {n} x {speed // n}G -> {names}")

Ethernet1/9: 800G on 8 lanes of 100G
   1 x 800G -> Ethernet1/9
   2 x 400G -> Ethernet1/9/1, Ethernet1/9/2
   4 x 200G -> Ethernet1/9/1, Ethernet1/9/2, Ethernet1/9/3, Ethernet1/9/4
   8 x 100G -> Ethernet1/9/1, Ethernet1/9/2, Ethernet1/9/3, Ethernet1/9/4, Ethernet1/9/5, Ethernet1/9/6, Ethernet1/9/7, Ethernet1/9/8
Ethernet1/10: 400G on 4 lanes of 100G
   1 x 400G -> Ethernet1/10
   2 x 200G -> Ethernet1/10/1, Ethernet1/10/2
   4 x 100G -> Ethernet1/10/1, Ethernet1/10/2, Ethernet1/10/3, Ethernet1/10/4

Coding and compatibility

Every optic stores information about itself in a small memory chip: its type, its maker, its part and serial numbers, its wavelength and reach, and its diagnostic thresholds. The switch reads this when the optic is inserted. Some switch makers check the coded vendor and refuse, or only grudgingly accept, optics they did not approve; many owners therefore buy optics coded for their switch brand, even from third-party makers. Corvid keeps an approved list, and an optic not on it may come up but be unsupported if anything goes wrong.

Newer optics use a management interface called CMIS, the Common Management Interface Specification, which also lets the switch tell the module how to configure its lanes for a breakout. CMIS optics take a few seconds to initialize after insertion, so a link that does not appear the instant the optic clicks in is not necessarily faulty. Older QSFP modules use the SFF-8636 interface and SFP modules SFF-8472. All of them report the diagnostics that Lesson 8 explains.

Checking before inserting

Jamal's check before any optic goes into a port takes half a minute. Read the part number on the module and compare it with the ticket. Check the form factor against the cage. Check the connector type against the cord or trunk that will plug in. For a breakout, check the far-end optics too. And check that the module is new or comes from the approved spares stock with a test record, not from a drawer of loose optics removed from other jobs, which may be faulty, the wrong type or coded for another switch.

Copper and active cables in the same cages

Not every module in an optic cage is an optic. Direct attach copper cables, DACs, have a module shape fixed at each end of a short twinax cable, and plug into the same cages; active electrical cables, AECs, add a chip in each end that re-times the signal to reach further. Active optical cables, AOCs, have optics built permanently into each end of a fiber cable. DCT-210 Lesson 4 covered their reaches and where each is allowed. In the cages they look like optics from the front, and the switch reads their memory in the same way, so the port shows the cable type, its length and its maker. The practical differences for a technician are that these cables cannot be cleaned or inspected, since there is no connector to open, and that a faulty one is replaced whole; and that a DAC's weight and stiffness strain the cage if it is not supported in the cable manager.

Wavelengths and colors

Multimode optics such as SR4 and SR8 use light at about 850 nm, on several parallel fibers. Single-mode optics use longer wavelengths: DR4 and DR8 use 1,310 nm on parallel fibers, and FR4 and LR4 combine four wavelengths near 1,310 nm on a single pair, separated inside the optic. Some optics carry a colored pull tab or label that hints at their type, but the colors are not consistent across makers, and Stonebridge's rule is to read the part number, not the color. Bidirectional optics, which send and receive on different wavelengths over a single fiber, must be paired with their partner type at the far end: one end's transmit wavelength is the other end's receive wavelength. A ticket that pairs two of the same kind of bidirectional optic will never link, however clean the fiber.

Optics and power in the hall

Optics are now a significant part of a data center's power. At roughly 15 to 18 watts per 800G optic, a pod of a few thousand backend links spends tens of kilowatts on optics alone, all of it heat for the cooling of DCT-230 to remove. This is one reason the industry is working on linear-drive optics, which drop the signal-processing chip inside the module and rely on the switch to do that work, and on co-packaged optics, which put the optical engines on the switch package itself, with fibers connected directly to the switch. Some co-packaged switches were already in early deployment in 2026. For technicians, co-packaged optics will change the work: there will be fewer pluggable modules to swap, more fiber connectors on the front of the switch to inspect and clean, and a failed optical engine may mean replacing a larger part or the whole switch.