Optical Power in dBm Is Not Fiber Loss in dB: What a GPU Link Report Tells You

Start with the quantity behind the unit

Start by separating a power level from a power ratio. A milliwatt (mW) is a linear unit of power. The dBm scale expresses that power logarithmically against the fixed reference of 1 mW. Fluke Networks defines 0 dBm as 1 mW; the Fiber Optic Association gives the relationship as power level in dBm = 10 log10(P / 1 mW).

The logarithm compares powers by their ratio. A level of -10 dBm therefore represents 0.1 mW: 10 log10(0.1) = -10. The minus sign says the power is below the reference. It does not describe negative light or, by itself, a failed receiver.

A dB value instead expresses a ratio between two powers. For insertion loss using the positive attenuation convention, loss in dB = 10 log10(Pin / Pout). Pin and Pout must use the same linear power unit and represent the intended input and output boundaries. Fluke's November 15, 2017 connector-loss article gives this input/output relationship. Other optical quantities also use dB, so retain the field name insertion loss when that is what the result reports.

References: [F1][A1][F2]

Why -10 dBm and 10 dB are different answers

Consider a mathematical example with 10 dB of attenuation. The power ratio Pin / Pout is 10, so Pout / Pin is 0.1. That identifies the fraction transmitted, not the amount of light arriving.

With an input of 1 mW, the output is 0.1 mW: 0 dBm at the input and -10 dBm at the output. With an input of 0.1 mW, the same attenuation gives an output of 0.01 mW: -10 dBm at the input and -20 dBm at the output. Both cases have 10 dB loss even though their output powers differ.

The reverse ambiguity matters when reading a GPU link summary. Suppose two fictional paths each end at -10 dBm. If one starts at -4 dBm, its input/output difference is 6 dB. If the other starts at -8 dBm, the difference is 2 dB. These stipulated values demonstrate the calculation; they are not source recommendations, measurements or acceptance limits. The same received power can accompany different losses because the starting powers differ.

This is why a low-power observation cannot, on its own, diagnose excessive fiber attenuation. The starting power is part of the relationship.

References: [A1][F2]

Read transceiver power as endpoint monitoring

NVIDIA's Onyx Ethernet Interface Commands page at documentation path 3102002 states that its transceiver diagnostics report Tx and Rx power in mW and dBm. The page is marked last updated January 19, 2021. This is a specific example of a product reporting power in two representations, rather than reporting two separate quantities called power and loss.

For a fictional Rx entry of -10 dBm, conversion to 0.1 mW changes the representation of the same reported power. It adds no information about the far-end transmitter, intervening connectors or passive-path loss. The cited paragraph establishes the output units; it does not establish accuracy, calibration or applicability for every transceiver or current GPU networking platform.

Whether an actual receive level satisfies the endpoint requirement depends on the exact equipment and the applicable specification's field definitions and conditions. More negative dBm means less power on this scale, but neither a minus sign nor a comparison with another port supplies a universal pass criterion. A power reading also does not identify the location or cause of a physical problem.

References: [N1][A1]

A valid subtraction still needs a valid comparison

Subtracting an output power level in dBm from an input power level in dBm cancels the shared 1 mW reference: 10 log10(Pin / 1 mW) minus 10 log10(Pout / 1 mW) equals 10 log10(Pin / Pout). The resulting unit is dB. The mathematical identity does not prove that the chosen records represent those two powers across the intended passive path.

For a bidirectional connection between endpoints A and B, the A-to-B comparison concerns A's transmitted power and B's received power. A's own receive reading concerns the reverse direction. The comparison must also refer to the matching optical channel and wavelength and the same defined power quantity. Two fields labelled Tx and Rx in one local export are not enough to establish that pairing.

Timing and measurement location matter too. Values from different times do not establish a common operating state. The physical locations represented by the reported powers, often called reference planes, must be understood before assigning their difference to a particular cable boundary. The exact monitoring specifications and limitations also matter; sharing a unit does not establish equivalent measurement performance.

These are consequences of the input/output relationship, not an operating instruction for Onyx. A matched monitoring comparison may inform an investigation, with its assumptions stated. It does not become a certified passive-channel result merely because a spreadsheet labels the difference loss. A performed, referenced method and its supporting evidence remain necessary for that claim.

References: [A1][F2][N1][F3]

Read a passive loss result with its reference

An optical loss test set, or OLTS, uses a light source and power meter to determine loss through the tested path. Fluke's OLTS and OTDR explanation describes this source/meter relationship and shows that different reference arrangements include or exclude different end connections. The reference defines which losses remain in the reported comparison; it is not simply another name for the transmitter's nominal output.

This distinction lets a buyer ask a more precise question than whether the fiber tested well. Does the returned loss cover the equipment-to-equipment connection being purchased, or a smaller segment? At what recorded wavelength, in which direction and under which performed reference method was it obtained? Which criterion applies to that boundary? Those fields determine what can be compared with the requirement.

A qualified passive result can answer its attenuation question without supplying live endpoint power or configured-network performance. Conversely, endpoint monitoring cannot fill an absent referenced loss measurement. The linked fiber-report guide covers the wider acceptance package; the OTDR guide covers event loss and reflection quantities.

References: [F3]

Use two fictional records to choose the next evidence

Suppose a buyer wants to know whether the complete equipment-to-equipment optical path has acceptable insertion loss. An invented handoff contains an endpoint screenshot reporting Rx = -10 dBm and a separate passive report stating 2.0 dB for a panel-to-panel segment. In this example the passive report explicitly excludes the equipment patch cords. No actual test or project is represented.

The screenshot supplies an endpoint power observation, equivalent to 0.1 mW. Without a corresponding input-power basis it cannot establish the complete path's attenuation. The passive report supplies a loss value for a narrower boundary; the omitted patch cords cannot be treated as measured simply because the segment belongs to the same connection.

The two numbers are not contradictory: one is absolute power and the other is a ratio for a segment. Neither alone answers the buyer's complete-path question. The useful next step is to establish the required complete boundary and criterion, then determine what applicable evidence already exists and what additional qualified measurement is needed. A separate endpoint-specification question may also require the network or equipment owner. Replacing an optic, cleaning a connector or accepting the channel does not follow from these two entries.

References: [A1][F3]

Commission evidence for the question you need answered

For an endpoint-power question, request the network or equipment owner's monitoring interpretation for the identified endpoint, optical channel and observation time against the applicable equipment specification. For passive attenuation, define the required fiber path and included connections, the qualified measurement method and reference, and the governing criterion. For a combined investigation, keep the monitoring observations and passive results connected by identity and context while preserving what each actually establishes.

Leviathan's physical-layer verification service is the relevant place to discuss accessible cable identification, labeling, physical-connectivity checks, defect records and agreed documentation. Formal OLTS or other instrument measurements require confirmation of suitable equipment, qualified personnel, method, coverage, criteria and report requirements for the engagement. The project inquiry below can start with a nonconfidential outline of the path and the physical question still open. Endpoint monitoring interpretation and network configuration remain with the responsible equipment and network parties.