Skip to main content
Ctrl+K
NVIDIA DCGM Documentation - Home NVIDIA DCGM Documentation - Home

NVIDIA DCGM Documentation

NVIDIA DCGM Documentation - Home NVIDIA DCGM Documentation - Home

NVIDIA DCGM Documentation

Table of Contents

  • About
  • Installation
    • DCGM
    • DCGM Exporter
  • Learn
    • Getting Started for System Administrators
      • Customizing the systemd Service
      • Configure Prometheus for DCGM Exporter
    • Getting Started for Application Developers
      • Build Your First DCGM API Clients
    • Core Services
      • Entities and Groups
      • Fields and Watches
      • Topology and NVLink
      • Process and Job Statistics
    • Modules
      • NVSwitch and ConnectX
      • vGPU
      • Introspection
      • Health Monitoring
      • Policy
      • Configuration
      • Diagnostics
      • Profiling
      • CPU Monitoring
      • Multi-Node Diagnostics
    • Detached GPUs
    • Error Injection
    • Debugging and Troubleshooting
  • Reference
    • Command Line
      • nv-hostengine
      • dcgmi
        • dcgmi discovery
        • dcgmi group
        • dcgmi fieldgroup
        • dcgmi config
        • dcgmi policy
        • dcgmi health
        • dcgmi set
        • dcgmi diag
        • dcgmi mndiag
        • dcgmi stats
        • dcgmi dmon
        • dcgmi topo
        • dcgmi nvlink
        • dcgmi profile
        • dcgmi introspect
        • dcgmi modules
        • dcgmi test
      • dcgmproftester
      • dcgm-exporter
    • Field Identifiers
    • Diagnostics
      • Diagnostic Configuration File
      • Diagnostic Errors
      • Diagnostic Plugins
        • Software Deployment Checks
        • Context Create Plugin
        • GPU Memory Plugin
        • PCIe - GPU Bandwidth Plugin
        • Diagnostic Plugin
        • Memory Bandwidth Plugin
        • Targeted Stress Plugin
        • Targeted Power Plugin
        • NVBandwidth Plugin
        • Memtest Diagnostic
        • Pulse Test Diagnostic
    • Multi-Node Diagnostics
      • mnubergemm
    • C/C++ API
      • Administrative
      • System
      • Configuration
      • Field APIs
      • Process Statistics
      • Job Statistics
      • Health Monitor
      • Policies
      • Topology
      • Metadata
      • Modules
      • Profiling
      • Enums and Macros
      • Structure Definitions
      • Field Types
      • Field Scope
      • Field Entity
      • Execution Control
      • Data Structures
      • Field Constants
    • DCGM Exporter Metrics
    • Go SDK
  • Release Notes
    • DCGM
    • DCGM Exporter
  • Reference
  • Command Line
  • dcgmi
  • dcgmi nvlink

dcgmi nvlink#

NAME#

dcgmi-nvlink - display NVLink status and GPU error counters

SYNOPSIS#

dcgmi nvlink --errors --gpuid <gpu-id> [--json] [--host <address>]
dcgmi nvlink --link-status [--show-entity-ids] [--host <address>]

DESCRIPTION#

dcgmi nvlink reports either system-wide link state or generation-specific error fields for one GPU. Link-state output covers GPUs and NVSwitches known to the host engine. The error names and whether they are per-link or GPU-level depend on the GPU generation.

OPTIONS#

-e, --errors

Display NVLink error counters for the GPU selected by --gpuid.

-g <gpu-id>, --gpuid <gpu-id>

Select the GPU for --errors. This option is required in error mode.

-s, --link-status

Display every GPU and NVSwitch link and its current state.

--show-entity-ids

With --link-status, include the dcgm_link_t entity ID assigned to each link.

-j, --json

Use JSON for --errors output. Link-status output is text only. See JSON output.

--host <address>

Connect to the specified host engine.

-h, --help

Print generated usage information and exit.

LINK STATES AND ENTITY IDS#

--link-status prints one character per physical port:

Link-state key#

Character

State

Meaning

U

Up

The link is active

D

Down

The port is supported but the link is down

X

Disabled

The supported link has been disabled

_

Not supported

There is no supported link at that port index

With --show-entity-ids, each supported port is shown as <port>:<packed-id>. Pass that packed ID as link:<packed-id> to a command such as dcgmi dmon, or use the clearer gpu_link:<gpu>:<port> and switch_link:<switch>:<port> selectors. Not-supported _ ports have no link entity ID.

ERROR COUNTER REFERENCE#

--errors chooses its field set from the selected GPU’s architecture:

Error output by GPU generation#

Architecture

Output organization

Values

Hopper and older

One Link <index> section for each supported link, up to link 17

CRC FLIT Error, CRC Data Error, Replay Error, and Recovery Error totals

Blackwell and newer

One GPU-level Error Detail section

Malformed Packet Error, Buffer Overrun Error, Rx Error, Rx Remote Error, Rx General Error, Rx Symbol Error, Effective Error, Tx Discards, Link Integrity Error, Symbol BER, and Effective BER

An unsupported legacy link is omitted. A blank Blackwell value is also omitted, so absence does not mean a zero count. The operation reads current field values and does not clear them. For definitions, field IDs, aggregate versus per-link behavior, and direct dmon queries, see Compare NVLink Error Counters, Watch Traffic and Error Fields, and Field Identifiers.

EXAMPLES#

Display all link states and entity IDs:

$ dcgmi nvlink --link-status --show-entity-ids

Display error counters for GPU 0 as JSON:

$ dcgmi nvlink --errors --gpuid 0 --json

SEE ALSO#

dcgmi, dcgmi topo, dcgmi dmon, Inspect Current Link State, Compare NVLink Error Counters, Watch Traffic and Error Fields, Field Identifiers

previous

dcgmi topo

next

dcgmi profile

On this page
  • NAME
  • SYNOPSIS
  • DESCRIPTION
  • OPTIONS
  • LINK STATES AND ENTITY IDS
  • ERROR COUNTER REFERENCE
  • EXAMPLES
  • SEE ALSO
NVIDIA NVIDIA
Privacy Policy | Your Privacy Choices | Terms of Service | Accessibility | Corporate Policies | Product Security | Contact

Copyright © 2018-2026, NVIDIA Corporation.

Last updated on 2026-07-25.