Fiber loss testing for data centers and GPU clusters

This article covers why, what those tests actually are, and the practical workflow from field to acceptance report. If you have been working outside plant for years and just landed on your first data center or GPU cluster project, this is the shift in mindset you are walking into.

The one thing I keep running into personally: I cannot open OLTS or OPM files on my phone. On the desktop I am either using the vendor's proprietary software or a cross-vendor tool. That is a real friction point in the workflow, and it is not going to get better on its own.

OLTS vs OPM. What’s the difference

Both use a power meter. They are commonly confused. They solve different problems.

OPM (Optical Power Meter) is a receiver. It measures light power at a point. Standalone instrument. Used for live circuit checks, transceiver output verification, Tx power confirmation, and reference readings. In DC and hyperscaler MF work, the OPMs in play are typically the power meter half of a paired multi-fiber set, purpose-built for the connectors on the job, not the handheld single-fiber units most field techs are used to.

OLTS (Optical Loss Test Set) is a matched pair. A calibrated light source at one end, a power meter at the other. Measures end-to-end insertion loss. Modern OLTS kits often add polarity check, length measurement, and bidirectional averaging.

Quick way to keep them straight in the field:

OTDR is a different tool for a different job. It locates events along the fiber and characterizes what those events are. It also has a dead zone that makes it unsuited to short data center links. OTDR gets its own full treatment in the Fiber Trace Analyzer guide. For DC and GPU work, OLTS and OPM are the tests that get run, saved, and delivered.

Why data centers and GPU clusters use OLTS and OPM

Four reasons, all rooted in how modern high-density networks are built.

Physical. OTDRs have a dead zone of typically 1 to 3 meters. A large share of DC and cluster connectivity is short-run patching where dead zone eats a meaningful fraction of the trace. Campus builds can still run several kilometers where OTDR is entirely appropriate, but even on those longer runs, DC and cluster plant has far fewer splices, no aerial or buried unknowns, and a much simpler event picture than OSP long-haul. The acceptance report needs a total loss number anyway, and OLTS gives you that directly.

Practical. Data center acceptance is pass or fail against a link loss budget. You do not need event location. You need to know if the link meets spec. OLTS gives you exactly that answer, per fiber, in seconds.

Speed. Modern multi-fiber (MF) OLTS testers scan every fiber in an MPO or MMC connector simultaneously. A 24-fiber MMC trunk that used to be 24 duplex measurements now returns one pass or fail in under 10 seconds. On a project with thousands of trunks, the speed multiplier is the reason MF testers exist.

Bit rate. 400G, 800G, and 1.6T links have very tight loss budgets. 800G runs on 8-fiber MPO with 4 transmit and 4 receive at 100 Gb/s per lane. 1.6T uses 16-fiber MPO or MMC-16. A single bad fiber in a 12-fiber trunk reduces effective bandwidth by roughly 8 percent plus retransmission overhead. In AI training environments, marginal links are the leading cause of training instability. Nobody wants to explain to the engineering lead why the cluster hangs on epoch 47.

Connector reality in modern DC and GPU builds:

• Duplex LC still common for lower-speed and switch uplinks
• MPO-8, MPO-12, MPO-16, MPO-24 for 40G through 1.6T
• MMC (Very Small Form Factor) in 12, 16, and 24 fiber counts for the highest-density AI cluster designs

Where OLTS and OPM dominate:

• Hyperscaler data centers
• GPU clusters running NVIDIA H100, H200, GB200, and GB300 with InfiniBand NDR and XDR fabrics
• Enterprise structured cabling
• Colo interconnect and cross-connect fields
• 5G fronthaul

Test equipment. Who makes what

OLTS units come as matched pairs. A light source unit and a power meter unit. Each vendor names theirs differently. Understanding the pairing matters because it dictates which files you will have when you are done, and which report software will read them.

Approved-vendor reality. Some acceptance specs restrict which OLTS vendors are permitted. Owners sometimes limit approved units to specific vendors or product lines, and buying an off-list unit means re-buying an approved one. Always check the spec before ordering equipment. This has cost more than one contractor real money.

File format reality. Every unit produces a proprietary file. Each vendor's format is different. Common practice on hyperscaler jobs is to deliver native raw files, a PDF summary for human review, and where the tester supports it, a structured export for ingestion into the owner's systems. Native files matter because some vendor formats are tamper-evident, which owners value for authenticity.

Software reality. Every vendor's report software is Windows-only, requires download and installation, and reads only that vendor's files. Some require a license key to unlock features that matter for real work. Opening more than one OPM file in a single session can be gated in some packages. And the download-and-install requirement itself is often a roadblock for techs on locked-down enterprise laptops who cannot install software without IT approval.

How to actually run the test

The how-to spine. This section assumes single-mode single-fiber and multi-fiber testing. Multimode uses the same equipment class with different wavelengths.

Test equipment prep

• Calibration must be current. Confirm against manufacturer guidance before use.
• Warm up units for the manufacturer's recommended minimum before referencing. A cold unit gives drifting readings. This is not optional.
• Re-reference after significant temperature changes.

Test Reference Cords (TRCs)

• Reference-grade cords with the connector loss and length the spec calls for.
• Replace TRCs after their rated mating cycles or sooner if damage is observed.
• Log usage. This matters when a job gets audited.

Reference method: one Test Cord Reference, side-by-side

Hyperscaler acceptance specs commonly require the one-jumper reference in side-by-side configuration for panel-to-panel testing. Loopback reference is an exception, not standard. After setting reference, verification with a known-good trunk cord is required before you are allowed to test the Fiber Under Test. This catches reference errors before they cost you a day.

Wavelengths

Multimode is typically 850 and 1300 nm. Single-mode is typically 1310 and 1550 nm for G.652D and G.657A cable, or 1550 and 1625 nm for G.654E and G.655. Follow the spec for what applies to your cable type.

Bidirectional averaging

Best practice where the spec allows it. Some specs require unidirectional. Check before testing.

Duplex LC testing

Two-fiber test using paired jumpers. Standard for switch uplinks and older topologies. Still the majority of enterprise links even where MPO dominates the trunk plant.

MF (multi-fiber) testing

The section that matters most for DC and GPU work.

• MPO polarity check (Type A, B, C, or Universal)
• Per-fiber pass or fail vs whole-connector pass or fail
• Unpinned vs pinned MPO and MMC connectors
• MF testers scan all fibers simultaneously and return per-fiber pass or fail in seconds
• Y-cable and reference cord selection depends on pinned or unpinned gender in the test unit

Pre-terminated cable testing

A distinct scenario in hyperscaler DC builds. Common configurations:

• MMC to MMC
• MPO to MPO of a different fiber count (breakout)
• MPO to MPO of the same fiber count

Each configuration has its own test cord requirements with defined polarities, pinning, and lengths. The spec dictates the reference procedure, verification threshold, and acceptance criteria. Always follow the owner's spec exactly, because pre-terminated cable acceptance is a common source of rework when it is wrong.

Standards you will see referenced in specs

• ANSI/TIA-526-7-A (OFSTP-7) or IEC 61280-4-2, for single-mode installed cable plant loss measurement
• TIA-568.3-D, for optical fiber cabling and components
• ISO/IEC 14763-3, for testing of optical fibre cabling
• IEC 61300-3-35, for fiber optic connector endface inspection
• IEC TR 62627-01, for connector cleaning methods

Link loss budget

Every OLTS measurement is compared against a link loss budget, the total dB the link is allowed to lose across cable, splices, and connectors. Insertion loss limits, splice loss limits, and reflectance thresholds are set by the owner in the acceptance spec, and they vary by owner, network type (non-DWDM, DWDM, DWDM with RAMAN), connector type, and generation of the spec. Test to whatever your spec calls for. Do not assume industry-wide numbers.

For a full walkthrough of how link loss budgets are built and used, see the link loss budget guide. To calculate a budget for your own span, use the Link Loss Calculator. Punch in cable length, connector count, splice count, and per-element loss values, and it produces the budget you should be testing against.

Common mistakes

Dirty ferrules. Wrong reference method for the spec. Missing polarity check. Testing before warm-up. Not verifying reference before testing. Every one of these has cost a project a day or more of rework.

Inspection and cleaning. Non-negotiable

Every loss reading is only as good as the connectors used to take it. This is standalone because it is the single most common source of bad test results.

Inspection standard: IEC 61300-3-35. Defines pass or fail zones on the ferrule endface. Every modern inspection scope automates against this standard.

Direct-view scopes are prohibited by most hyperscaler specs. Digital inspection probes only. Auto-focus and auto-centralization strongly preferred. Field of view suitable for the connector type.

Approved inspection tools:

• EXFO FIP-500
• Fluke FI-3000 FiberInspector Ultra
• Viavi INX 700 Series
• AFL FOCIS Lightning 2

What to inspect and when

• Every test reference cord endface, before every reference
• Every tester port before every session
• Every fiber under test at both ends before every measurement
• Every mating adapter that gets connected
• Both sides of every connector, the plug endface AND the bulkhead port

The protocol: inspect, clean if necessary, replace or connect. Initial inspection is required before any cleaning. Unnecessary cleaning damages the ferrule. This sequence maintains endface integrity.

Cleaning

Follow IEC TR 62627-01.

• LC uses click-cleaner or reel-type cleaner
• MPO, MTP, and MMC use dedicated multi-fiber cleaners. Fluke Quick Clean, EXFO Cletop, and similar work here. LC cleaners do not work on multi-fiber ferrules.
• Clean, inspect, mate. If it fails inspection, clean and re-inspect. If it still fails, stop and investigate.

Environmental discipline. In a data hall under active install, dust can contaminate a ferrule between the moment you clean it and the moment you plug it in. Cover ports the moment you disconnect. Never leave a connector open in an active install area.

Reference verification. After setting reference, verify with a known-good verification cord. If the verification loss exceeds the spec, stop, inspect, clean, and re-reference before proceeding.

Getting the files off the tester

USB export. Every modern OLTS supports USB export in the vendor's proprietary format. Some testers also offer CSV export, but with less data than the native format. Data you will wish you had when the acceptance package gets kicked back.

Cloud sync.

• Fluke: LinkWare Live
• Viavi: StrataSync
• EXFO: EXFO Exchange

The reality in hyperscaler environments. Cloud sync is often not permitted for security reasons. Files come off on USB, get delivered to project managers, sit in shared drives. Owners typically require:

• Native files for authenticity, since some vendor formats are tamper-evident
• PDF for human-readable acceptance review
• A structured export (where the tester supports it) for machine ingestion into owner systems

Reading and reporting the files

Each vendor's software reads only its own files. FastReporter reads EXFO. LinkWare reads Fluke. FiberChekPRO reads Viavi. All are Windows-only, all require download and installation, and some require a license key to unlock features you actually need. Opening multiple OPM files in one session can be gated in some packages.

For techs on locked-down laptops, or for techs who do not want to install four different vendor packages just to read four different file types, the install requirement itself is a real roadblock. Some customers pay their equipment vendor to build a custom report template for their acceptance deliverable format. That is a real consulting engagement.

If you need to read files across vendors on the same job, options include exporting to CSV and rebuilding in Excel, installing every vendor's software, or using a cross-vendor tool. The JTS Fiber Analysis tool is one option that reads across vendors, runs in the browser (no install), and produces a report from mixed OLTS, OPM, and OTDR files. It reads these directly:

Report content. What owners expect

Hyperscaler acceptance packages have specific content requirements. Building your report to match saves rework.

Header content typically required:

• Location A and Location Z with site and panel identifiers
• Cable identifier (labeling identifier)
• Fiber identifier
• Timestamp of measurement
• Personnel who conducted the measurement

Test equipment information:

• Model, software release, serial number
• Last calibration date
• Reference and verification values, which prove the tester was set up correctly

Delivery format matrix:

• Native raw file, per fiber, per wavelength, per direction
• PDF for the human-readable summary report
• Structured export (JSON or similar) when the tester supports it, for machine ingestion

Naming conventions vary by owner. Every hyperscaler and every large contractor has its own. Ask for the naming convention document as part of scope kickoff, or you will rework the deliverable at the end.

Practical field-to-report workflow

Step by step.

1. Warm up test units per manufacturer guidance
2. Verify calibration date on all equipment
3. Inspect and clean all TRCs and unit ports
4. Set reference per the spec (one-jumper side-by-side is standard for hyperscalers)
5. Verify reference with a known-good verification cord
6. Inspect and clean each Fiber Under Test at both ends
7. Test per the spec (bidirectional if allowed, unidirectional if required)
8. Save native files. Do not rely on CSV export
9. Get files to a laptop (USB, or cloud sync where allowed)
10. Process through vendor software or cross-vendor tool
11. Build report with required header content and the owner's naming convention
12. Deliver native + PDF (+ structured export where required)

Related reading

• Link loss budget, the concept walkthrough
• Link Loss Calculator, the working tool
• Fiber Analysis tool announcement, for how batch analysis fits alongside OLTS work
• CD and PMD testing overview, for the dispersion tests that also apply to long-haul spans
• Fiber Trace Analyzer, for OTDR-side testing
• Field guides, for the full JTS reference