Start with the rack and the route.
At AI scale, an awkward cable choice repeats across every tray, leaf and spine. We design the wiring alongside the fabric: where each connector lands, how the assembly reaches it, which bundles share a pathway and how a technician replaces one connection without disturbing its neighbors.
WBN supplies custom trunks and breakouts with main lengths and individual leg lengths matched to the installation. Staggered legs can follow different rack positions instead of leaving identical loops at every port. The drawing records the measurement reference, breakout location, service allowance, bend requirements and destination labels. Fiber type, jacket, termination and optical loss belong in that same definition.
Put connector cost and service access in the same decision.
Our cable program spans economical MPO assemblies for volume deployment and premium connector constructions for dense, frequently serviced interfaces. The useful comparison includes the connector envelope and the technician's access to its release mechanism, alongside the assembly price.
SENKO's MPO EZ-Way illustrates the distinction: a push-pull boot, reduced footprint and soft-spring mechanism address access around closely packed dual-MPO transceivers. We can engineer assemblies around this kind of serviceable connector where the port layout calls for it. A premium release mechanism solves a physical handling problem; it does not replace the required optical specification. SENKO connector design ↗
We specify polish, pin gender, key orientation and populated positions explicitly. The connector shell alone cannot identify a usable connection.
Build the plane map into the assembly.
Custom passive shuffle boxes turn a fabric's port schedule into an organized fiber permutation. Input bundles are regrouped into destination bundles, with plane identity and strand positions preserved through the box and its labels. A passive shuffle neither switches packets nor splits optical power.
We define the mapping from endpoint port to connector position, including Tx-to-Rx correspondence, polarity and each intermediate mating pair. Separate planes remain separate in the map; a shared enclosure also remains a shared physical dependency. Box placement, access and connector density are engineering choices.
Corning's shuffle study describes both harness and enclosure approaches; a box adds a connector pair compared with the harness arrangement discussed. We account for that serviceability tradeoff in the complete channel-loss budget. Passive shuffle architectures ↗
Make short copper links count.
We engineer thin, cost-effective 800G and 1.6T copper connections for dense Clos fabrics. On suitable short spine-to-leaf routes, a copper assembly replaces the optical modules and fiber that would otherwise form that link. The economic opportunity grows with the number of repeated connections, so equipment placement and actual routed length matter.
Passive DAC, linear-active ACC and retimed AEC address different electrical budgets. We select the construction, length and conductor size together with the host channel, power requirements and routing constraints. Thin cable is not automatically passive cable, and the shortest physical separation is not always the installed route.
TE documents 1.6T passive and active OSFP copper options, while NVIDIA's current XDR copper portfolio is active-only. Industry feasibility and a particular platform's supported channel are different inputs to the design. TE OSFP portfolio · NVIDIA interconnects
Resolve the lanes before releasing the cable.
An 800G label can describe eight nominal 100G lanes or four nominal 200G lanes; eight 200G lanes form a 1.6T interface. We match the exact host application, logical port breakout and connector population at both ends. The nominal aggregate rate does not settle electrical compatibility.
Optical breakouts require the same discipline. NVIDIA's documented twin-port 100G-PAM4 OSFP designs contain two four-channel optical engines, while wavelength-multiplexed FR4 connections cannot be fanned out like parallel DR4. A passive assembly implements a supported lane arrangement; it cannot convert one optical application into another. Lane-rate families · Optical lane arrangements
Turn the topology into a buildable cable schedule.
Bring the rack elevations, endpoint inventory and fabric port map into one engineering brief. We can turn them into assembly drawings, main and leg lengths, termination details, plane and strand maps, destination labels and a counted bill of materials. Optical circuits, physical modules and cable assemblies stay separate in the count.
The verification plan follows the same records: continuity and polarity, end-to-end optical loss or the supported electrical channel, then link behavior at the specified lane mode. Use the strand-map guide and fiber-path worksheet to capture the inputs. The result is wiring that can be built, installed and traced as part of the fabric.
Sources and applicability
WBN custom cabling and connection engineering. Construction, lane mapping and channel requirements are defined for the selected endpoints and installation.
-
SENKO MPO EZ-Way handout
Push-pull access, connector footprint and soft-spring mechanism; connector-specific figures are not assembly specifications.
-
Corning: Fiber shuffling
Passive fiber regrouping, custom assemblies and the connector-loss tradeoff between shuffle harnesses and boxes.
-
TE Connectivity OSFP copper assemblies
1.6T passive and active copper portfolio; reach and conductor choices are construction-specific.
-
NVIDIA LinkX interconnect portfolio
100G- and 200G-per-lane families and the active-only XDR copper portfolio.
-
NVIDIA transceivers and fiber details: 100G-PAM4
Model-specific twin-port optics, parallel breakouts and wavelength-multiplexed interfaces.
