Cabling_
OM4 vs OM5 vs OS2: Choosing Fiber for AI Cluster Reaches
A field engineer's guide from Leviathan Systems to selecting OM4, OM5, or OS2 fiber for AI cluster interconnects, based on reach, optics cost, and future-proofing for GPU fabrics like InfiniBand and Ethernet.
Key facts
- OM4 supports 100GBASE-SR4 to 400 meters and 400GBASE-SR8 to 100 meters under TIA-568.3-D; OM5 extends 400GBASE-SR8 to 150 meters only when used with SWDM4 optics (wideband multimode fiber, TIA-492AAAD).
- OS2 single-mode fiber supports 400GBASE-DR4 to 500 meters and 400GBASE-LR4 to 10 km; coherent optics (e.g., 400G-ZR) can reach 40+ km for campus links.
- In AI clusters, the dominant link is 400 Gbps (4x100G PAM4) or 800 Gbps (8x100G PAM4); OM5 benefits only SWDM4 optics, which are rarely used in GPU fabrics—most links use parallel optics that do not utilize the wideband capability.
- Optics cost: VCSEL-based multimode transceivers (SR4/SR8) are typically 30–50% cheaper per port than single-mode EML or silicon-photonics transceivers at 400G, but the gap narrows at 800G and beyond.
- MPO trunk cables for AI clusters are factory-terminated (never field-crimped) with UPC (multimode) or APC (single-mode) polish; field work is patching, routing, cleaning, inspection, and testing per IEC 61300-3-35.
- For GPU fabrics like InfiniBand NDR400 or Ethernet 400G, switch-to-rack reach is typically under 100 meters, making OM4 or OM5 viable; OS2 is required for inter-row, inter-aisle, or inter-building links beyond 150 meters.
Reach and Bandwidth: Where Each Fiber Type Excels
OM4 is the workhorse for AI cluster leaf-spine topologies where switch-to-rack distances are under 100 meters. It supports 100GBASE-SR4 to 400 meters and 400GBASE-SR8 to 100 meters per TIA-568.3-D, covering most intra-row and intra-aisle runs in a typical data center. OM5, defined by TIA-492AAAD, adds wideband multimode fiber (WBMMF) to support four wavelengths in the 850–950 nm range, extending 400GBASE-SR8 to 150 meters. However, OM5 only benefits SWDM4 optics (e.g., 100G-SWDM4), which are rare in AI clusters—most GPU fabrics use parallel optics (e.g., 400G-SR8 with 8 fibers per direction) that do not leverage the wideband capability. The practical effect: for standard parallel optics, OM4 and OM5 have identical reach limits.
OS2 single-mode fiber is the choice for reaches beyond 150 meters, such as inter-row, inter-aisle, or inter-building links. It supports 400GBASE-DR4 to 500 meters and 400GBASE-LR4 to 10 km using standard EML or silicon-photonics transceivers. For longer distances, coherent optics (e.g., 400G-ZR) can push to 40+ km. In large AI clusters spanning multiple data halls or campuses, OS2 is mandatory. The trade-off is higher optics cost: single-mode transceivers are typically 30–50% more expensive per port than multimode at 400G, but the fiber itself is cheaper per meter. The decision hinges on whether the link distance exceeds the multimode reach limit—if it does, OS2 is the only option.
Optics Cost and Power: Why Multimode Still Dominates the Rack
Multimode optics use VCSELs (vertical-cavity surface-emitting lasers) that are cheaper to manufacture and consume less power than single-mode EML or silicon-photonics transceivers. At 400G, a multimode SR8 transceiver (e.g., QSFP-DD) costs roughly 30–50% less per port than a single-mode DR4 or FR4 transceiver. For a cluster with thousands of switch-to-rack links, this cost difference can be significant—potentially hundreds of thousands of dollars. Additionally, multimode transceivers typically draw 8–12 W per port versus 10–15 W for single-mode, which reduces power and cooling overhead.
However, the cost advantage narrows as speeds increase. At 800G, single-mode silicon-photonics transceivers are becoming cost-competitive with VCSEL-based multimode, especially for reaches above 50 meters. The industry trend is toward single-mode for next-generation speeds (1.6T and beyond) because VCSEL bandwidth scaling faces physical limits. For current AI clusters (H100/B200 with NDR400 or 400G Ethernet), multimode remains the most cost-effective choice for intra-rack and leaf-spine links under 100 meters. For future-proofing, some operators run OS2 dark fiber alongside OM4 to allow upgrades without re-cabling.
Field Deployment: Cleaning, Inspection, and Testing
Fiber performance in AI clusters is critically dependent on connector cleanliness and end-face quality. A single dust particle on an MPO connector can cause bit errors, retransmissions, and GPU collective operation timeouts. The standard practice is to inspect every connector end-face with a handheld microscope (200x or 400x magnification) before mating, and clean with a dry clicker or wet-dry method per IEC 61300-3-35. For single-mode, the pass/fail criteria are stricter: no scratches or defects in the core zone (0–25 µm radius). For multimode, the core zone is larger (0–50 µm), but contamination is equally problematic.
Testing should include an optical loss test set (OLTS) with a light source and power meter at the operating wavelength (850 nm for OM4/OM5, 1310/1550 nm for OS2). For MPO trunks, use a calibrated MPO continuity tester or an OTDR to verify loss and reflectance. The TIA-568.3-D standard specifies maximum insertion loss per connector pair (e.g., 0.75 dB for multimode, 0.5 dB for single-mode). In AI clusters, we often target tighter loss budgets (e.g., 0.3 dB per mated pair) to ensure margin for high-speed PAM4 signals. Always test end-to-end after installation and document results for troubleshooting. Leviathan Systems recommends maintaining a test log with fiber IDs, measured loss, and inspection images for every link.
Future-Proofing: OS2 for Longevity, OM4/OM5 for Density
Single-mode fiber (OS2) has virtually unlimited bandwidth potential—it can support 400G, 800G, 1.6T, and beyond using wavelength-division multiplexing (WDM) and coherent optics. For AI clusters expected to last 5–10 years, running OS2 trunk cables between rows and to the MDF provides a future-proof backbone that can be upgraded by swapping optics only. The fiber itself is cheaper per meter than OM4/OM5, but the termination and patch panels are more expensive due to tighter tolerances. Many hyperscalers now deploy OS2 as the default for all structured cabling, reserving multimode for short patch cords within the rack.
OM5 was marketed as future-proof because it supports multiple wavelengths, but in practice, the industry has not adopted SWDM4 widely. Most AI clusters use parallel optics (e.g., 8 fibers per direction for 400G-SR8), which do not benefit from OM5's wideband capability. OM4 remains the standard for multimode because it is cheaper and widely available. For new builds, the recommendation is to use OM4 for intra-rack and leaf-spine links under 100 meters, and OS2 for all longer runs. Avoid OM5 unless you have a specific need for SWDM4 optics, which is rare in GPU fabrics.
Common Failure Modes in the Field
The most frequent failure is contamination on MPO connectors causing high insertion loss or back-reflection. This is often missed during installation because technicians skip inspection or use improper cleaning methods. Always inspect every connector end-face per IEC 61300-3-35 before mating; a single smudge can cause link errors in PAM4 signaling. A second common failure is mismatched polish types: UPC (ultra-physical contact) on multimode and APC (angled physical contact) on single-mode. Mating a UPC connector to an APC adapter causes high loss and potential damage. Verify polish type on both ends before mating.
A third failure mode is exceeding the bend radius during cable routing. Multimode fiber has a minimum bend radius of 10× the cable diameter under load, and 7.5× under no load (per TIA-568.3-D). Exceeding this causes micro-bends that increase loss and can cause intermittent errors. In GPU clusters with dense cabling, cables are often pulled too tight around corners or through cable trays. Use a bend-radius limiter or cable management fingers to enforce the spec. Finally, polarity errors in MPO trunks (e.g., Type A vs. Type B) can cause link failures. Always document polarity and test with a continuity tester before connecting optics. Catching these failures early—during installation testing—prevents costly downtime later.
Decision Criteria: When to Choose OM4 vs OM5 vs OS2
For a typical AI cluster with 100-meter maximum reach between switches and racks, OM4 is the best choice. It is cost-effective, widely supported by optics vendors, and meets the reach requirements for 400G-SR8 and 800G-SR8. If the reach exceeds 100 meters but is under 150 meters, OM5 is an option only if you plan to use SWDM4 optics—otherwise, stick with OM4 or move to OS2. For any link over 150 meters, OS2 is mandatory. For inter-building or campus links, OS2 with coherent optics is the only viable solution.
Future-proofing considerations: If the cluster is expected to upgrade to 1.6T or 3.2T within 5 years, running OS2 now avoids re-cabling later. However, if the cluster is built for a specific GPU generation (e.g., H100 or B200) with a 3–4 year lifespan, OM4 is sufficient and saves capital. The decision also depends on optics availability: at 400G, multimode optics are mature and readily available; at 800G, single-mode optics are more common. For new builds, a hybrid approach works: OM4 for intra-rack and leaf-spine, OS2 for all backbone and inter-row links. One of the largest hyperscalers in Texas uses this approach for their latest AI clusters.
Standards referenced: TIA-568.3-D (Optical Fiber Cabling Components Standard) · IEC 60793-2-10 (Multimode Fibers, Categories A1a, A1b, A1d) · ISO/IEC 11801 (Information Technology – Generic Cabling for Customer Premises) · IEC 61300-3-35 (Connector End-Face Inspection Criteria) · TIA-492AAAD (OM5 Wideband Multimode Fiber Specification)
Frequently asked_
Can I use OM5 for 400G-SR8 links to get extra reach?
No. 400G-SR8 uses 8 parallel fibers at a single wavelength (850 nm) and does not benefit from OM5's wideband capability. OM5 only extends reach for SWDM4 optics that use multiple wavelengths. For 400G-SR8, OM4 and OM5 have the same reach limit (100 meters per TIA-568.3-D). If you need more than 100 meters, use OS2 with 400G-DR4 optics.
What is the cost difference between OM4 and OS2 cabling for a 100-meter run?
The fiber itself is cheaper for OS2 (single-mode) than OM4 (multimode) per meter, but the connectors, patch panels, and termination labor are more expensive for OS2 due to tighter tolerances. Overall, a 100-meter OS2 trunk cable with MPO connectors is typically 10–20% more expensive than an equivalent OM4 trunk. However, the optics cost difference (multimode vs. single-mode transceivers) is much larger and dominates the total cost of ownership. Leviathan Systems recommends a total-cost analysis including optics for your specific link count.
Should I use APC or UPC connectors for AI cluster fiber?
For multimode (OM4/OM5), use UPC connectors. For single-mode (OS2), use APC connectors to minimize back-reflection, which is critical for high-speed PAM4 and coherent optics. Never mix APC and UPC—they are mechanically incompatible and will cause high loss. Always verify polish type on both ends before mating.
How do I test fiber links in an AI cluster to ensure they meet performance requirements?
Use an optical loss test set (OLTS) with a light source and power meter at the operating wavelength. For multimode, test at 850 nm; for single-mode, test at 1310 nm and optionally 1550 nm. Measure insertion loss and compare to the TIA-568.3-D limits (e.g., 0.75 dB max per mated pair for multimode). Also use an OTDR to locate faults and verify reflectance. For MPO trunks, use a calibrated MPO continuity tester to check polarity and continuity on all fibers. Document all results for future troubleshooting.
What is the typical reach for GPU-to-GPU NVLink in an NVL72 rack, and does it use fiber?
In an NVL72 rack, GPU-to-GPU NVLink runs over the copper NVLink spine/backplane inside the rack—not over fiber or MPO. The reach is less than 1 meter. Fiber (MPO) is used for the scale-out compute network (InfiniBand or Ethernet) between switches and racks. Never conflate NVLink with fiber—they are separate domains. nvidia-smi 'NVLink up' status depends on the copper backplane, not fiber links.