Cabling_
MPO Trunk Planning & Fiber-Count Math for GPU Halls
A field-proven method for calculating MPO trunk counts and fiber counts from GPU cluster port maps, covering MPO-8/12/16 selection, cable routing constraints, and common ordering errors that cause mid-build shortfalls in AI data centers.
Key facts
- Each H100 GPU requires one scale-out network port (400 Gb/s), typically served by a single MPO-12 trunk with 8 active fibers (for 400G-SR8) or MPO-8 with 4 active fibers (for 400G-SR4.2).
- A 64-GPU rack requires 64 MPO-12 trunks for the scale-out network (one per GPU), plus additional trunks for management and storage—total fiber count per rack: 768 fibers if using MPO-12 (64 trunks × 12 fibers).
- MPO trunk cables are factory-terminated; field operations are limited to patching, routing, cleaning, inspection, and testing—never field-crimp MPO connectors.
- Typical best practice is to provision 20–30% spare fibers per trunk (e.g., 4 spare fibers in a 12-fiber trunk used for 8 active fibers) to allow for future upgrades or single fiber break.
- In NVL72-class racks, NVLink copper backplane runs inside the rack; MPO/fiber carries only the scale-out network (InfiniBand or Ethernet) between racks and leaf/spine switches—never conflate the two domains.
- Order trunks with 10% overcount for spares and plan for 8–10 week lead time for factory-terminated custom-length trunks.
Mapping Port Counts to MPO Trunk Types
Start with the GPU port map. Each H100 or B200 GPU has one scale-out network port (typically 400 Gb/s) that connects to a leaf switch via an MPO trunk. The transceiver type determines the MPO count: 400G-SR8 uses MPO-12 (8 active fibers, 4 spare); 400G-SR4.2 uses MPO-8 (4 active fibers, 4 spare). For 800 Gb/s ports (e.g., B200 with dual 400G), you need two MPO-12 trunks per GPU. Always confirm the transceiver part number with the OEM—the port shape alone does not guarantee compatibility.
For a 64-GPU rack, the trunk count equals the number of GPU ports: 64 trunks. If using MPO-12 with 8 active fibers, each trunk carries one 400 Gb/s link. The leaf switch must have matching transceivers. Calculate total fiber count per rack by multiplying the number of trunks by the fiber count per trunk (12 for MPO-12, 8 for MPO-8). This gives 768 fibers for MPO-12 or 512 for MPO-8 in a 64-GPU rack. Do not assume multiple ports per trunk unless using breakouts for lower speeds—verify the transceiver's fiber requirement per link.
Fiber-Count Math with Spare Capacity
Never order trunks with zero spare fibers. Industry best practice (per TIA-568.3-D) is to provision 20–30% spare fibers per trunk. For an MPO-12 trunk carrying 8 active fibers, that means 4 spares (33% spare). For an MPO-8 trunk carrying 4 active fibers, 4 spares (50% spare). This spare capacity allows for future transceiver upgrades (e.g., from 400G to 800G using more fibers) or replacing a damaged fiber without pulling a new trunk.
Calculate total fiber count: multiply the number of trunks by the fiber count per trunk. For 64 MPO-12 trunks (8 active + 4 spare each), total fibers = 64 × 12 = 768. The active fiber count is only 64 × 8 = 512. The remaining 256 fibers are spare. If you later upgrade to 800G transceivers that use all 12 fibers, you have the capacity without re-cabling. Always add 10% spare trunks for breakage or routing errors—order 70 trunks for a 64-trunk requirement.
Routing Constraints and Bend-Radius Planning
MPO trunk cables have a minimum bend radius under load (typically 10× the cable diameter per manufacturer spec) and unloaded (5× diameter). For a standard 3 mm diameter trunk, that’s 30 mm loaded, 15 mm unloaded. Exceeding this causes micro-bends that increase insertion loss beyond the -0.35 dB per mated pair limit (per TIA-568.3-D). In GPU halls, cable trays must have 90-degree sweeps or larger, not sharp corners. Use cable managers with a minimum radius of at least 2 inches (50 mm) or per manufacturer guidance.
Route trunks from the GPU rack to the leaf switch row using overhead trays or underfloor pathways. Keep trunk lengths under 100 meters for 400 Gb/s SR8 optics (per IEEE 802.3db). For longer runs, switch to single-mode fiber (OS2) with MPO-12 APC connectors and LR4 optics. Label both ends with a unique ID (e.g., R01-L01-T01) using a durable labeler—never write on the boot. Leave service loops (at least 1 meter) at both ends for re-termination or rerouting.
Testing and Validation Before Commissioning
After routing and patching, test every fiber in every trunk with an OTDR and a calibrated MPO continuity tester. The OTDR measures insertion loss and reflectance at each connector and splice. Acceptable loss per mated pair is typically ≤0.35 dB for multimode (OM4/OM5) and ≤0.5 dB for single-mode (OS2), per TIA-568.3-D. Reflectance should be ≤ -26 dB for multimode, ≤ -40 dB for single-mode. Any fiber exceeding these limits must be cleaned and retested; if still failing, replace the trunk.
Use a microscope (200× or 400×) to inspect every ferrule endface before mating. Look for scratches, pits, or contamination per IEC 61300-3-35. Grade 1 (no defects) is required for all new installations. Never mate a dirty connector—cleaning with a dry click-cleaner or isopropyl alcohol and lint-free wipes is mandatory. Document all test results in a spreadsheet with trunk ID, fiber number, loss, reflectance, and pass/fail. This baseline is important for warranty claims and future troubleshooting.
Common Failure Modes and How to Catch Them
The most frequent field failure is ordering the wrong trunk type: MPO-12 when the transceiver requires MPO-8, or vice versa. This causes a mid-build shortfall because the trunk won’t physically mate with the transceiver. Always verify the transceiver spec (e.g., 400G-SR8 uses MPO-12 APC, 400G-SR4.2 uses MPO-8 UPC) before ordering. Second failure: insufficient spare fibers. A 64-GPU rack with 8 active fibers per trunk and no spares leaves zero room for a single fiber break—one bad fiber means replacing the entire trunk. Third failure: exceeding bend radius during installation, causing latent loss that only shows up during OTDR testing. Catch this by inspecting cable pathways before pulling and using a bend-radius gauge.
Another common issue is mismatched polish types: UPC (ultra-physical contact) vs APC (angled physical contact). UPC connectors have a flat endface; APC has an 8-degree angle. Mixing them causes high reflectance and loss. All MPO trunks in a GPU hall should use the same polish type—typically APC for single-mode, UPC for multimode. Label the patch panels clearly. Finally, dirty connectors from handling during installation are the #1 cause of high loss. Implement a clean-before-every-mate policy: inspect and clean both ends before plugging. Use a video microscope to verify cleanliness.
Ordering Process and Lead-Time Management
Order trunks at least 8–10 weeks before installation day. Factory-terminated MPO trunks are custom-length (e.g., 10 m, 25 m, 50 m) and require lead time for polishing and testing. Specify the exact length from the GPU rack to the leaf switch, plus 2–3 meters for service loops. Use a cable length calculator or measure the actual pathway with a laser distance meter—never guess. Order 10% extra trunks for spares (e.g., 70 trunks for a 64-trunk requirement).
Include a test report from the manufacturer for every trunk, showing insertion loss and reflectance for each fiber. Reject any trunk that exceeds the -0.35 dB per mated pair limit. Store trunks in a clean, dry environment (temperature 15–30°C, humidity 30–70%) before installation. Never coil trunks tighter than 20× the cable diameter. At Leviathan Systems, we pre-stage all trunks in a staging area, test them with an OTDR, and label them before moving to the rack row—this catches factory defects before they delay the build.
Scaling from One Rack to a GPU Hall
For a GPU hall with 100 racks (6,400 GPUs), each GPU requires one MPO-12 trunk to the leaf tier, totaling 6,400 trunks. The leaf-to-spine network adds additional trunks depending on the oversubscription ratio. For a common 1:1 design, each leaf switch needs as many uplink ports as downlink ports, roughly doubling the trunk count to ~12,800 total for the hall. This is a starting estimate; adjust based on your specific topology.
Plan cable tray capacity: a standard 12-inch (300 mm) tray holds about 200 MPO-12 trunks (2,400 fibers) if laid flat with 1-inch spacing. For 12,800 trunks, you need approximately 64 trays per row, or multiple rows of trays. Use vertical cable managers every 6 racks to prevent congestion. Label every trunk at both ends with a barcode that links to a database of source rack, destination rack, and fiber count. This is essential for troubleshooting and future adds. A consistent color-coding scheme—blue for GPU-to-leaf, yellow for leaf-to-spine, green for management—reduces mispatches.
Standards referenced: TIA-568.3-D (Optical Fiber Cabling and Component Standard) · IEEE 802.3db (400 Gb/s and 800 Gb/s Ethernet over multimode fiber) · IEC 61300-3-35 (Fiber optic connector endface inspection criteria)
Frequently asked_
How do I know whether to use MPO-8 or MPO-12 trunks for my H100 cluster?
Check the transceiver spec for your GPU and leaf switch. 400G-SR8 transceivers use MPO-12 (8 active fibers, 4 spare). 400G-SR4.2 transceivers use MPO-8 (4 active fibers, 4 spare). If you’re unsure, ask the OEM for the exact part number. Never assume based on the port shape—some ports accept both but require different breakout cables. For future-proofing, MPO-12 is more common in AI clusters because it supports 800G upgrades using all 12 fibers.
What is the maximum distance for MPO trunks in a GPU hall?
For 400 Gb/s SR8 optics over OM4 multimode fiber, the maximum distance is 100 meters per IEEE 802.3db. For longer runs (up to 2 km), use single-mode OS2 fiber with LR4 optics and MPO-12 APC connectors. Always measure the actual pathway length with a laser distance meter and add 2–3 meters for service loops. Exceeding the distance limit causes bit errors and link drops.
Why do my MPO trunks fail OTDR testing even though they are new?
The most common causes are dirty connectors, damaged ferrules from shipping, or exceeding the bend radius during installation. Inspect every ferrule with a 200× microscope before testing—even new trunks can have contamination from the factory. Clean with a dry click-cleaner or isopropyl alcohol. If the loss is still high, check for kinks or sharp bends in the cable path. Replace any trunk with a ferrule scratch deeper than 2 µm per IEC 61300-3-35.
How many spare trunks should I order for a 64-rack GPU hall?
Order 10% extra trunks for spares—that’s about 7 spare trunks per 64-trunk requirement. Additionally, keep 5–10% spare fibers per trunk (e.g., 4 spare fibers in an MPO-12 trunk). This covers fiber breaks, future upgrades, and routing errors. At Leviathan Systems, we also stock 5% spare trunks in a central storage area for emergency replacements during commissioning.
Can I field-terminate MPO connectors on trunks?
No. MPO connectors are factory-polished and require precision alignment of 12 or 16 fibers. Field termination kits exist but rarely achieve the -0.35 dB loss per mated pair required for 400 Gb/s links. Always use factory-terminated trunks. If you need a custom length, order it from the manufacturer with a test report. Field work is limited to patching, routing, cleaning, inspection, and testing.