Networking_
400G vs 800G vs 1.6T Optics: Selecting Transceivers for AI Fabric
This article gives deployment engineers explicit criteria for matching 400G/800G/1.6T transceivers to switch ports, fiber reach, and MPO infrastructure in AI scale-out fabrics while keeping NVLink copper domains separate.
Key facts
- IEEE 802.3bs defines 400 Gb/s Ethernet using four 100 Gb/s lanes.
- IEEE 802.3ck extends 100 Gb/s electrical signaling per lane for 800 Gb/s and 1.6 Tb/s modules.
- OSFP modules allow higher power dissipation than QSFP-DD due to larger surface area and cage design.
- MPO-12 or MPO-16 trunks for parallel optics are factory-terminated; field work consists of routing, cleaning, inspection, and testing only.
- Scale-out InfiniBand or Ethernet links use fiber between racks; intra-rack GPU-to-GPU NVLink remains on copper backplanes in NVL72-class systems.
- A calibrated MPO continuity tester or OTDR is required to verify insertion loss before bringing 800G or 1.6T links online.
- Forward error correction settings must match between transceiver and switch ASIC or the link will not train.
Determining required port speed from switch ASIC and fabric bandwidth needs
Start with the switch ASIC port map and target bisection bandwidth of the AI fabric. 400G ports suffice when job scale fits within existing leaf-spine capacity; move to 800G only when aggregate east-west traffic exceeds what 400G lanes sustain without oversubscription.
Confirm the ASIC supports the lane rate before ordering optics. Some 400G switches accept 800G modules in breakout mode while others require a full-speed 800G ASIC to enable the higher rate. Document the exact port configuration and FEC type before procurement.
Choosing between OSFP and QSFP-DD based on switch faceplate and power budget
OSFP provides better thermal headroom for 800G and 1.6T optics because of its larger surface area and higher allowed power envelope. QSFP-DD fits existing high-density 400G deployments but may throttle when populated with 800G modules.
Inspect the switch faceplate and cage type first. Mixing form factors on the same switch is possible only when the hardware supports both; otherwise standardize on one cage type across the fabric to simplify spares and replacement procedures.
Leviathan Systems crews verify cage compatibility and power rail capacity on every rack before transceiver installation begins.
Assessing reach requirements against existing single-mode or multi-mode fiber infrastructure
Measure the actual fiber distance and loss budget between leaf and spine switches. 400G SR4 or DR4 modules work on short-reach multi-mode or single-mode plants; 800G and 1.6T equivalents often require DR8 or FR4 optics once distance exceeds a few hundred meters.
Compare measured loss against the transceiver specification. If the plant was installed for 400G, re-test every segment with an OTDR before installing higher-speed optics because margin shrinks rapidly at 100 Gb/s and 200 Gb/s per lane.
Routing and managing MPO trunk cables in high-density rack environments
Route MPO trunks with the minimum number of bends and maintain the manufacturer-specified minimum bend radius at every turn. Label both ends of every trunk with port and lane mapping before installation so that any later fault can be isolated without disturbing live links.
Install trunks in the vertical managers first, then patch to the switch ports. This sequence prevents accidental tension on the switch cage when adding or removing modules during commissioning.
Verifying transceiver compatibility with host FEC and signal integrity parameters
Load the intended FEC mode on the switch port before inserting the transceiver. Mismatched FEC prevents link training even when optical power is within specification.
Run the switch port diagnostics after the link comes up. Check pre-FEC bit error rate, lane skew, and module temperature against the values listed in the OEM port guide. Any parameter outside the documented range requires replacement of the transceiver or re-termination of the MPO connector.
Common failure modes encountered during 800G and 1.6T deployments
Dirty or damaged MPO end faces remain the leading cause of link flaps at 800G and above. Even microscopic contamination produces enough loss to push the link outside the FEC correction window. Always inspect and clean both the transceiver and trunk ferrules with a calibrated scope before mating.
Another frequent failure is mismatched break-out mapping when 800G modules are used in 2x400G or 4x200G configurations. The lane ordering on the switch must exactly match the fiber mapping in the trunk; a single swapped pair prevents training. Field crews catch this by verifying continuity on every fiber with an MPO tester before powering the ports.
Over-temperature shutdowns occur when higher-power OSFP modules are placed in cages designed for lower power. Measure airflow and ambient temperature at the switch intake before population and derate port count if the rack cooling cannot maintain the required inlet temperature.
Sequencing upgrades from 400G to higher speeds without fabric disruption
Upgrade one spine or leaf pair at a time while the fabric runs on the remaining capacity. Bring the new 800G or 1.6T ports online in a maintenance window, then migrate traffic using the routing protocol metrics rather than hard cutovers.
Retain the 400G transceivers as spares until the entire fabric has been validated at the higher speed. This approach limits exposure if an unexpected compatibility issue appears during the first production jobs.
Standards referenced: IEEE 802.3bs · IEEE 802.3ck · TIA-568.3-D
Frequently asked_
Can I plug an 800G transceiver into a 400G switch port?
Only when the switch port and ASIC explicitly support the higher rate or a valid breakout configuration. Inserting an unsupported module usually results in the port remaining down or training at the lower speed with reduced lanes. Check the switch release notes for the exact part number before attempting insertion.
What test equipment is required before lighting an 800G MPO link?
Use a calibrated MPO continuity tester or OTDR to confirm polarity and insertion loss on every fiber. Follow with an end-face inspection scope on both the transceiver and trunk connectors. Skip any of these steps and the link will fail to train or flap under load.
How does NVLink affect transceiver selection for the scale-out network?
NVLink traffic stays on the copper backplane inside the rack. Fiber and transceivers carry only the InfiniBand or Ethernet scale-out fabric between racks. Do not attempt to run NVLink over MPO trunks; the domains are electrically and topologically separate.
What is the first step when an 800G link fails to come up?
Inspect and clean the MPO end faces on both sides. Next verify FEC settings and lane mapping match between transceiver and switch. Only after those checks pass should you replace the transceiver or retest the fiber plant with an OTDR.
When should Leviathan Systems be engaged for a speed upgrade project?
Engage before any fiber work begins so that port maps, loss budgets, and cooling capacity can be validated against the target optics. Early involvement prevents rework when the new transceivers arrive and the existing plant proves marginal.