| 1 | plugin_name: go.d.plugin |
| 2 | modules: |
| 3 | - meta: |
| 4 | id: collector-go.d.plugin-dcgm |
| 5 | plugin_name: go.d.plugin |
| 6 | module_name: dcgm |
| 7 | monitored_instance: |
| 8 | name: Nvidia Data Center GPU Manager (DCGM) |
| 9 | link: https://github.com/NVIDIA/dcgm-exporter |
| 10 | icon_filename: nvidia.svg |
| 11 | categories: |
| 12 | - data-collection.hardware-and-sensors |
| 13 | keywords: |
| 14 | - nvidia |
| 15 | - gpu |
| 16 | - dcgm |
| 17 | - dcgm-exporter |
| 18 | related_resources: |
| 19 | integrations: |
| 20 | list: |
| 21 | - plugin_name: go.d.plugin |
| 22 | module_name: nvidia_smi |
| 23 | info_provided_to_referring_integrations: |
| 24 | description: "" |
| 25 | overview: |
| 26 | data_collection: |
| 27 | metrics_description: | |
| 28 | This collector gathers NVIDIA GPU telemetry from a `dcgm-exporter` endpoint. |
| 29 | It supports all numeric fields exposed by the exporter and maps them into Netdata-native contexts. |
| 30 | method_description: | |
| 31 | It collects metrics by periodically scraping the exporter Prometheus endpoint over HTTP. |
| 32 | supported_platforms: |
| 33 | include: [] |
| 34 | exclude: [] |
| 35 | multi_instance: true |
| 36 | additional_permissions: |
| 37 | description: "" |
| 38 | default_behavior: |
| 39 | auto_detection: |
| 40 | description: This integration does not support auto-detection in v1. |
| 41 | limits: |
| 42 | description: | |
| 43 | The collector applies global and per-metric time series limits to prevent excessive cardinality. |
| 44 | performance_impact: |
| 45 | description: The impact depends on dcgm-exporter field selection and resulting series cardinality. |
| 46 | setup: |
| 47 | prerequisites: |
| 48 | list: |
| 49 | - title: Run dcgm-exporter |
| 50 | description: Install DCGM and run `dcgm-exporter` so that a Prometheus endpoint is available (default `:9400/metrics`). |
| 51 | - title: Configure exporter field list |
| 52 | description: | |
| 53 | The default exporter profile exposes a small subset of fields. |
| 54 | Use the Netdata recommended profile: |
| 55 | [`dcgm-exporter-netdata.csv`](https://github.com/netdata/netdata/blob/master/src/go/plugin/go.d/collector/dcgm/dcgm-exporter-netdata.csv) |
| 56 | (raw download: `https://raw.githubusercontent.com/netdata/netdata/master/src/go/plugin/go.d/collector/dcgm/dcgm-exporter-netdata.csv`). |
| 57 | |
| 58 | The Netdata profile enables 127 fields by default and documents all remaining known DCGM fields as commented entries. |
| 59 | To customize beyond the baseline, uncomment the field you need and comment one currently enabled field. |
| 60 | |
| 61 | Runtime validation artifact: |
| 62 | `src/go/plugin/go.d/collector/dcgm/runtime-validation-driver-590.48.01-dcgm-exporter-4.4.1-4.5.2.md` |
| 63 | and |
| 64 | `src/go/plugin/go.d/collector/dcgm/runtime-validation-driver-590.48.01-dcgm-exporter-4.4.1-4.5.2.json` |
| 65 | |
| 66 | Validation is primarily version-scoped (NVIDIA driver + DCGM/DCGM-exporter versions), so treat it as a strong baseline rather than universal compatibility. |
| 67 | |
| 68 | Example: |
| 69 | `dcgm-exporter -f /path/to/dcgm-exporter-netdata.csv` |
| 70 | - title: Keep collection intervals aligned |
| 71 | description: | |
| 72 | Set Netdata `update_every` to the same value as dcgm-exporter collection interval (default 30 seconds). |
| 73 | Example exporter interval: `dcgm-exporter -c 30000` and Netdata `update_every: 30`. |
| 74 | - title: Enable profiling capabilities (optional) |
| 75 | description: Profiling fields may require additional privileges/capabilities in your runtime environment. |
| 76 | configuration: |
| 77 | file: |
| 78 | name: go.d/dcgm.conf |
| 79 | options: |
| 80 | description: | |
| 81 | The following options can be defined globally: update_every, autodetection_retry. |
| 82 | folding: |
| 83 | title: Config options |
| 84 | enabled: true |
| 85 | list: |
| 86 | - name: update_every |
| 87 | description: Data collection interval (seconds). Keep this aligned with dcgm-exporter collection interval. |
| 88 | default_value: 30 |
| 89 | required: false |
| 90 | group: Collection |
| 91 | - name: autodetection_retry |
| 92 | description: Autodetection retry interval (seconds). Set 0 to disable. |
| 93 | default_value: 0 |
| 94 | required: false |
| 95 | group: Collection |
| 96 | - name: url |
| 97 | description: DCGM exporter metrics endpoint URL. |
| 98 | default_value: http://127.0.0.1:9400/metrics |
| 99 | required: true |
| 100 | group: Target |
| 101 | - name: timeout |
| 102 | description: HTTP request timeout (seconds). |
| 103 | default_value: 10 |
| 104 | required: false |
| 105 | group: Target |
| 106 | - name: max_time_series |
| 107 | description: Global time series limit. If exceeded, collection is skipped for this cycle. |
| 108 | default_value: 2000 |
| 109 | required: false |
| 110 | group: Limits |
| 111 | - name: max_time_series_per_metric |
| 112 | description: Per-metric time series limit. Metrics above this limit are skipped. |
| 113 | default_value: 200 |
| 114 | required: false |
| 115 | group: Limits |
| 116 | - name: username |
| 117 | description: Username for Basic HTTP authentication. |
| 118 | default_value: "" |
| 119 | required: false |
| 120 | group: HTTP Auth |
| 121 | - name: password |
| 122 | description: Password for Basic HTTP authentication. |
| 123 | default_value: "" |
| 124 | required: false |
| 125 | group: HTTP Auth |
| 126 | - name: bearer_token_file |
| 127 | description: Path to a file containing a bearer token. |
| 128 | default_value: "" |
| 129 | required: false |
| 130 | group: HTTP Auth |
| 131 | - name: tls_skip_verify |
| 132 | description: Skip TLS certificate and hostname verification (insecure). |
| 133 | default_value: no |
| 134 | required: false |
| 135 | group: TLS |
| 136 | - name: tls_ca |
| 137 | description: Path to CA bundle used to validate the server certificate. |
| 138 | default_value: "" |
| 139 | required: false |
| 140 | group: TLS |
| 141 | - name: tls_cert |
| 142 | description: Path to client TLS certificate (for mTLS). |
| 143 | default_value: "" |
| 144 | required: false |
| 145 | group: TLS |
| 146 | - name: tls_key |
| 147 | description: Path to client TLS private key (for mTLS). |
| 148 | default_value: "" |
| 149 | required: false |
| 150 | group: TLS |
| 151 | - name: proxy_url |
| 152 | description: HTTP proxy URL. |
| 153 | default_value: "" |
| 154 | required: false |
| 155 | group: Proxy |
| 156 | - name: proxy_username |
| 157 | description: Username for proxy authentication. |
| 158 | default_value: "" |
| 159 | required: false |
| 160 | group: Proxy |
| 161 | - name: proxy_password |
| 162 | description: Password for proxy authentication. |
| 163 | default_value: "" |
| 164 | required: false |
| 165 | group: Proxy |
| 166 | - name: headers |
| 167 | description: Additional HTTP headers to include in the request. |
| 168 | default_value: "" |
| 169 | required: false |
| 170 | group: Request |
| 171 | - name: method |
| 172 | description: HTTP method. |
| 173 | default_value: GET |
| 174 | required: false |
| 175 | group: Request |
| 176 | - name: body |
| 177 | description: HTTP request body. |
| 178 | default_value: "" |
| 179 | required: false |
| 180 | group: Request |
| 181 | - name: not_follow_redirects |
| 182 | description: Do not follow HTTP redirects. |
| 183 | default_value: no |
| 184 | required: false |
| 185 | group: Request |
| 186 | - name: force_http2 |
| 187 | description: Force HTTP/2 (including h2c over TCP). |
| 188 | default_value: no |
| 189 | required: false |
| 190 | group: Request |
| 191 | - name: vnode |
| 192 | description: Associate this job with a Virtual Node. |
| 193 | default_value: "" |
| 194 | required: false |
| 195 | group: Virtual Node |
| 196 | examples: |
| 197 | folding: |
| 198 | title: Config |
| 199 | enabled: true |
| 200 | list: |
| 201 | - name: Local exporter |
| 202 | description: Collect metrics from a local dcgm-exporter endpoint. |
| 203 | config: | |
| 204 | jobs: |
| 205 | - name: local |
| 206 | url: http://127.0.0.1:9400/metrics |
| 207 | update_every: 30 |
| 208 | - name: TLS endpoint |
| 209 | description: Collect metrics over HTTPS with custom CA certificate. |
| 210 | config: | |
| 211 | jobs: |
| 212 | - name: secure |
| 213 | url: https://dcgm-exporter.example.com:9400/metrics |
| 214 | update_every: 30 |
| 215 | tls_ca: /etc/netdata/certs/dcgm-ca.crt |
| 216 | - name: Increased cardinality limits |
| 217 | description: Increase limits when collecting large field sets and multiple entities. |
| 218 | config: | |
| 219 | jobs: |
| 220 | - name: dcgm_large |
| 221 | url: http://127.0.0.1:9400/metrics |
| 222 | update_every: 30 |
| 223 | max_time_series: 10000 |
| 224 | max_time_series_per_metric: 2000 |
| 225 | troubleshooting: |
| 226 | problems: |
| 227 | list: [] |
| 228 | alerts: |
| 229 | - name: dcgm_gpu_xid_errors |
| 230 | metric: dcgm.gpu.reliability.xid |
| 231 | info: NVIDIA driver reported GPU XID error on GPU ${label:gpu} |
| 232 | link: https://github.com/netdata/netdata/blob/master/src/health/health.d/dcgm.conf |
| 233 | - name: dcgm_gpu_row_remap_failure |
| 234 | metric: dcgm.gpu.reliability.row_remap_status |
| 235 | info: GPU row remapping failed on GPU ${label:gpu} |
| 236 | link: https://github.com/netdata/netdata/blob/master/src/health/health.d/dcgm.conf |
| 237 | - name: dcgm_gpu_uncorrectable_remapped_rows |
| 238 | metric: dcgm.gpu.reliability.row_remap_events |
| 239 | info: Uncorrectable remapped rows increased on GPU ${label:gpu} |
| 240 | link: https://github.com/netdata/netdata/blob/master/src/health/health.d/dcgm.conf |
| 241 | - name: dcgm_gpu_power_violation |
| 242 | metric: dcgm.gpu.throttle.violations |
| 243 | info: Power throttling detected on GPU ${label:gpu} |
| 244 | link: https://github.com/netdata/netdata/blob/master/src/health/health.d/dcgm.conf |
| 245 | - name: dcgm_gpu_thermal_violation |
| 246 | metric: dcgm.gpu.throttle.violations |
| 247 | info: Thermal throttling detected on GPU ${label:gpu} |
| 248 | link: https://github.com/netdata/netdata/blob/master/src/health/health.d/dcgm.conf |
| 249 | metrics: |
| 250 | folding: |
| 251 | title: Metrics |
| 252 | enabled: false |
| 253 | description: | |
| 254 | Metrics are grouped into static Netdata contexts. Contexts are created only when matching DCGM fields are present in the exporter output. |
| 255 | availability: [] |
| 256 | scopes: |
| 257 | - name: gpu |
| 258 | description: These metrics refer to GPU device instances. |
| 259 | labels: |
| 260 | - name: gpu |
| 261 | description: gpu label from exporter metrics. |
| 262 | - name: uuid |
| 263 | description: uuid label from exporter metrics. |
| 264 | metrics: |
| 265 | - name: dcgm.gpu.capability.support |
| 266 | description: GPU Capability Support metrics. |
| 267 | unit: state |
| 268 | chart_type: line |
| 269 | dimensions: |
| 270 | - name: cc_mode |
| 271 | - name: cuda_compute_capability |
| 272 | - name: gpm_support |
| 273 | - name: mig_attributes |
| 274 | - name: mig_ci_info |
| 275 | - name: mig_gi_info |
| 276 | - name: mig_max_slices |
| 277 | - name: supported_clocks |
| 278 | - name: supported_type_info |
| 279 | - name: dcgm.gpu.clock.frequency |
| 280 | description: GPU Clock Frequency metrics. |
| 281 | unit: MHz |
| 282 | chart_type: line |
| 283 | dimensions: |
| 284 | - name: app_mem_clock |
| 285 | visibility: hidden |
| 286 | - name: app_sm_clock |
| 287 | visibility: hidden |
| 288 | - name: max_mem_clock |
| 289 | visibility: hidden |
| 290 | - name: max_sm_clock |
| 291 | visibility: hidden |
| 292 | - name: max_video_clock |
| 293 | visibility: hidden |
| 294 | - name: memory |
| 295 | - name: sm |
| 296 | - name: video_clock |
| 297 | - name: dcgm.gpu.compute.activity |
| 298 | description: GPU Compute Pipeline Activity metrics. |
| 299 | unit: '%' |
| 300 | chart_type: line |
| 301 | dimensions: |
| 302 | - name: dram |
| 303 | - name: fp16 |
| 304 | - name: fp32 |
| 305 | - name: fp64 |
| 306 | - name: graphics_engine_active |
| 307 | - name: integer |
| 308 | - name: sm_active |
| 309 | - name: sm_occupancy |
| 310 | - name: tensor |
| 311 | - name: dcgm.gpu.compute.tensor.activity |
| 312 | description: GPU Tensor Core Activity by precision type. |
| 313 | unit: '%' |
| 314 | chart_type: line |
| 315 | dimensions: |
| 316 | - name: tensor_dfma |
| 317 | - name: tensor_hmma |
| 318 | - name: tensor_imma |
| 319 | - name: dcgm.gpu.compute.media.activity |
| 320 | description: GPU Media Engine Activity metrics. |
| 321 | unit: '%' |
| 322 | chart_type: line |
| 323 | dimensions: |
| 324 | - name: nvdec0_active |
| 325 | - name: nvdec1_active |
| 326 | - name: nvdec2_active |
| 327 | - name: nvdec3_active |
| 328 | - name: nvdec4_active |
| 329 | - name: nvdec5_active |
| 330 | - name: nvdec6_active |
| 331 | - name: nvdec7_active |
| 332 | - name: nvjpg0_active |
| 333 | - name: nvjpg1_active |
| 334 | - name: nvjpg2_active |
| 335 | - name: nvjpg3_active |
| 336 | - name: nvjpg4_active |
| 337 | - name: nvjpg5_active |
| 338 | - name: nvjpg6_active |
| 339 | - name: nvjpg7_active |
| 340 | - name: nvofa0_active |
| 341 | - name: nvofa1_active |
| 342 | - name: dcgm.gpu.compute.cache.activity |
| 343 | description: GPU Memory Cache Hit/Miss metrics. |
| 344 | unit: 'events/s' |
| 345 | chart_type: line |
| 346 | dimensions: |
| 347 | - name: hostmem_cache_hit |
| 348 | - name: hostmem_cache_miss |
| 349 | - name: peermem_cache_hit |
| 350 | - name: peermem_cache_miss |
| 351 | - name: dcgm.gpu.compute.utilization |
| 352 | description: GPU Compute Utilization metrics. |
| 353 | unit: '%' |
| 354 | chart_type: line |
| 355 | dimensions: |
| 356 | - name: decoder |
| 357 | - name: encoder |
| 358 | - name: gpu |
| 359 | - name: memory_copy |
| 360 | - name: dcgm.gpu.cpu.power |
| 361 | description: GPU CPU Power metrics. |
| 362 | unit: Watts |
| 363 | chart_type: line |
| 364 | dimensions: |
| 365 | - name: module_power_util_current |
| 366 | - name: sysio_power_util_current |
| 367 | - name: dcgm.gpu.cpu.info |
| 368 | description: GPU CPU Information metrics. |
| 369 | unit: value |
| 370 | chart_type: line |
| 371 | dimensions: |
| 372 | - name: cpu_model |
| 373 | - name: cpu_vendor |
| 374 | - name: dcgm.gpu.diagnostics.results |
| 375 | description: GPU Diagnostics Results metrics. |
| 376 | unit: state |
| 377 | chart_type: line |
| 378 | dimensions: |
| 379 | - name: diag_diagnostic_result |
| 380 | - name: diag_eud_result |
| 381 | - name: diag_memory_bandwidth_result |
| 382 | - name: diag_memory_result |
| 383 | - name: diag_memtest_result |
| 384 | - name: diag_nccl_tests_result |
| 385 | - name: diag_nvbandwidth_result |
| 386 | - name: diag_pulse_test_result |
| 387 | - name: diag_software_result |
| 388 | - name: diag_targeted_power_result |
| 389 | - name: diag_targeted_stress_result |
| 390 | - name: dcgm.gpu.diagnostics.status |
| 391 | description: GPU Diagnostics Status metrics. |
| 392 | unit: state |
| 393 | chart_type: line |
| 394 | dimensions: |
| 395 | - name: diag_status |
| 396 | - name: dcgm.gpu.health.status |
| 397 | description: GPU Health Status metrics. |
| 398 | unit: state |
| 399 | chart_type: line |
| 400 | dimensions: |
| 401 | - name: imex_daemon_status |
| 402 | - name: imex_domain_status |
| 403 | - name: dcgm.gpu.interconnect.connectx.error_status |
| 404 | description: GPU ConnectX Error Status metrics. |
| 405 | unit: state |
| 406 | chart_type: line |
| 407 | dimensions: |
| 408 | - name: connectx_correctable_err_mask |
| 409 | - name: connectx_correctable_err_status |
| 410 | - name: connectx_uncorrectable_err_mask |
| 411 | - name: connectx_uncorrectable_err_severity |
| 412 | - name: connectx_uncorrectable_err_status |
| 413 | - name: dcgm.gpu.interconnect.connectx.errors |
| 414 | description: GPU ConnectX Errors metrics. |
| 415 | unit: errors/s |
| 416 | chart_type: line |
| 417 | dimensions: |
| 418 | - name: connectx_correctable_err_mask |
| 419 | - name: connectx_correctable_err_status |
| 420 | - name: connectx_uncorrectable_err_mask |
| 421 | - name: connectx_uncorrectable_err_severity |
| 422 | - name: connectx_uncorrectable_err_status |
| 423 | - name: dcgm.gpu.interconnect.connectx.link |
| 424 | description: GPU ConnectX Link metrics. |
| 425 | unit: value |
| 426 | chart_type: line |
| 427 | dimensions: |
| 428 | - name: connectx_active_pcie_link_speed |
| 429 | - name: connectx_expect_pcie_link_speed |
| 430 | - name: dcgm.gpu.interconnect.connectx.status |
| 431 | description: GPU ConnectX Status metrics. |
| 432 | unit: state |
| 433 | chart_type: line |
| 434 | dimensions: |
| 435 | - name: connectx_health |
| 436 | - name: dcgm.gpu.interconnect.error_rate |
| 437 | description: GPU Interconnect Error Rate metrics. |
| 438 | unit: errors/s |
| 439 | chart_type: line |
| 440 | dimensions: |
| 441 | - name: c2c_link_error_intr |
| 442 | - name: c2c_link_error_replay |
| 443 | - name: c2c_link_error_replay_b2b |
| 444 | - name: dcgm.gpu.interconnect.fabric |
| 445 | description: GPU Fabric State metrics. |
| 446 | unit: state |
| 447 | chart_type: line |
| 448 | dimensions: |
| 449 | - name: fabric_clique_id |
| 450 | - name: fabric_cluster_uuid |
| 451 | - name: fabric_health_mask |
| 452 | - name: fabric_manager_error_code |
| 453 | - name: fabric_manager_status |
| 454 | - name: dcgm.gpu.interconnect.nvlink.error_rate |
| 455 | description: GPU NVLink Error Rate metrics. |
| 456 | unit: errors/s |
| 457 | chart_type: line |
| 458 | dimensions: |
| 459 | - name: gpu_nvlink_errors |
| 460 | - name: dcgm.gpu.interconnect.pcie.error_rate |
| 461 | description: GPU PCIe Error Rate metrics. |
| 462 | unit: errors/s |
| 463 | chart_type: line |
| 464 | dimensions: |
| 465 | - name: pcie_count_correctable_errors |
| 466 | - name: pcie_replay |
| 467 | - name: dcgm.gpu.interconnect.pcie.link.generation |
| 468 | description: GPU PCIe Link Generation metrics. |
| 469 | unit: generation |
| 470 | chart_type: line |
| 471 | dimensions: |
| 472 | - name: link_gen |
| 473 | - name: max_link_gen |
| 474 | visibility: hidden |
| 475 | - name: dcgm.gpu.interconnect.pcie.link.width |
| 476 | description: GPU PCIe Link Width metrics. |
| 477 | unit: lanes |
| 478 | chart_type: line |
| 479 | dimensions: |
| 480 | - name: connectx_active_pcie_link_width |
| 481 | - name: connectx_expect_pcie_link_width |
| 482 | - name: link_width |
| 483 | - name: max_link_width |
| 484 | visibility: hidden |
| 485 | - name: dcgm.gpu.interconnect.state |
| 486 | description: GPU Interconnect State metrics. |
| 487 | unit: state |
| 488 | chart_type: line |
| 489 | dimensions: |
| 490 | - name: c2c_link |
| 491 | - name: c2c_link_power_state |
| 492 | - name: c2c_link_status |
| 493 | - name: dcgm.gpu.interconnect.pcie.state |
| 494 | description: GPU PCIe State metrics. |
| 495 | unit: state |
| 496 | chart_type: line |
| 497 | dimensions: |
| 498 | - name: diag_pcie_result |
| 499 | - name: dcgm.gpu.interconnect.throughput |
| 500 | description: GPU Interconnect Throughput metrics. |
| 501 | unit: B/s |
| 502 | chart_type: area |
| 503 | dimensions: |
| 504 | - name: c2c_max_bandwidth |
| 505 | - name: c2c_rx_all_bytes |
| 506 | - name: c2c_rx_data_bytes |
| 507 | - name: c2c_tx_all_bytes |
| 508 | - name: c2c_tx_data_bytes |
| 509 | - name: dcgm.gpu.interconnect.pcie.throughput |
| 510 | description: GPU PCIe Throughput metrics. |
| 511 | unit: B/s |
| 512 | chart_type: area |
| 513 | dimensions: |
| 514 | - name: pcie_rx |
| 515 | - name: pcie_rx_throughput |
| 516 | - name: pcie_tx |
| 517 | - name: pcie_tx_throughput |
| 518 | - name: dcgm.gpu.interconnect.nvlink.throughput |
| 519 | description: GPU NVLink Throughput metrics. |
| 520 | unit: B/s |
| 521 | chart_type: area |
| 522 | dimensions: |
| 523 | - name: nvlink_rx |
| 524 | - name: nvlink_tx |
| 525 | - name: dcgm.gpu.interconnect.total.throughput |
| 526 | description: GPU Interconnect Total Throughput metrics. |
| 527 | unit: B/s |
| 528 | chart_type: area |
| 529 | dimensions: |
| 530 | - name: pcie |
| 531 | - name: nvlink |
| 532 | - name: dcgm.gpu.internal.boundary |
| 533 | description: GPU Internal Boundary Fields metrics. |
| 534 | unit: state |
| 535 | chart_type: line |
| 536 | dimensions: |
| 537 | - name: first_connectx_field_id |
| 538 | - name: first_vgpu_field_id |
| 539 | - name: internal_fields_0_end |
| 540 | - name: internal_fields_0_start |
| 541 | - name: last_connectx_field_id |
| 542 | - name: last_vgpu_field_id |
| 543 | - name: dcgm.gpu.inventory.identity |
| 544 | description: GPU Inventory Identity metrics. |
| 545 | unit: value |
| 546 | chart_type: line |
| 547 | dimensions: |
| 548 | - name: brand |
| 549 | - name: count |
| 550 | - name: cuda_visible_devices_str |
| 551 | - name: minor_number |
| 552 | - name: name |
| 553 | - name: nvml_index |
| 554 | - name: serial |
| 555 | - name: uuid |
| 556 | - name: dcgm.gpu.inventory.platform |
| 557 | description: GPU Platform Inventory metrics. |
| 558 | unit: value |
| 559 | chart_type: line |
| 560 | dimensions: |
| 561 | - name: platform_chassis_serial_number |
| 562 | - name: platform_chassis_slot_number |
| 563 | - name: platform_host_id |
| 564 | - name: platform_infiniband_guid |
| 565 | - name: platform_module_id |
| 566 | - name: platform_peer_type |
| 567 | - name: platform_tray_index |
| 568 | - name: dcgm.gpu.inventory.software |
| 569 | description: GPU Software and Firmware metrics. |
| 570 | unit: value |
| 571 | chart_type: line |
| 572 | dimensions: |
| 573 | - name: inforom_config_check |
| 574 | - name: inforom_config_valid |
| 575 | - name: inforom_image_ver |
| 576 | - name: oem_inforom_ver |
| 577 | - name: power_inforom_ver |
| 578 | - name: process_name |
| 579 | - name: vbios_version |
| 580 | - name: dcgm.gpu.memory.bar1_usage |
| 581 | description: GPU BAR1 Memory Usage metrics. |
| 582 | unit: B |
| 583 | chart_type: stacked |
| 584 | dimensions: |
| 585 | - name: free |
| 586 | - name: used |
| 587 | - name: dcgm.gpu.memory.bar1_capacity |
| 588 | description: GPU BAR1 Memory Capacity metrics. |
| 589 | unit: B |
| 590 | chart_type: line |
| 591 | dimensions: |
| 592 | - name: total |
| 593 | - name: dcgm.gpu.memory.ecc_error_rate |
| 594 | description: GPU ECC Error Rate metrics. |
| 595 | unit: errors/s |
| 596 | chart_type: line |
| 597 | dimensions: |
| 598 | - name: ecc_current |
| 599 | - name: ecc_dbe_agg |
| 600 | - name: ecc_dbe_agg_cbu |
| 601 | - name: ecc_dbe_agg_dev |
| 602 | - name: ecc_dbe_agg_l1 |
| 603 | - name: ecc_dbe_agg_l2 |
| 604 | - name: ecc_dbe_agg_reg |
| 605 | - name: ecc_dbe_agg_shm |
| 606 | - name: ecc_dbe_agg_srm |
| 607 | - name: ecc_dbe_agg_tex |
| 608 | - name: ecc_dbe_vol |
| 609 | - name: ecc_dbe_vol_cbu |
| 610 | - name: ecc_dbe_vol_dev |
| 611 | - name: ecc_dbe_vol_l1 |
| 612 | - name: ecc_dbe_vol_l2 |
| 613 | - name: ecc_dbe_vol_reg |
| 614 | - name: ecc_dbe_vol_shm |
| 615 | - name: ecc_dbe_vol_srm |
| 616 | - name: ecc_dbe_vol_tex |
| 617 | - name: ecc_pending |
| 618 | - name: ecc_sbe_agg |
| 619 | - name: ecc_sbe_agg_cbu |
| 620 | - name: ecc_sbe_agg_dev |
| 621 | - name: ecc_sbe_agg_l1 |
| 622 | - name: ecc_sbe_agg_l2 |
| 623 | - name: ecc_sbe_agg_reg |
| 624 | - name: ecc_sbe_agg_shm |
| 625 | - name: ecc_sbe_agg_srm |
| 626 | - name: ecc_sbe_agg_tex |
| 627 | - name: ecc_sbe_vol |
| 628 | - name: ecc_sbe_vol_cbu |
| 629 | - name: ecc_sbe_vol_dev |
| 630 | - name: ecc_sbe_vol_l1 |
| 631 | - name: ecc_sbe_vol_l2 |
| 632 | - name: ecc_sbe_vol_reg |
| 633 | - name: ecc_sbe_vol_shm |
| 634 | - name: ecc_sbe_vol_srm |
| 635 | - name: ecc_sbe_vol_tex |
| 636 | - name: dcgm.gpu.memory.ecc_errors |
| 637 | description: GPU ECC Errors metrics. |
| 638 | unit: errors |
| 639 | chart_type: line |
| 640 | dimensions: |
| 641 | - name: ecc_current |
| 642 | - name: ecc_dbe_agg_cbu |
| 643 | - name: ecc_dbe_agg_dev |
| 644 | - name: ecc_dbe_agg_l1 |
| 645 | - name: ecc_dbe_agg_l2 |
| 646 | - name: ecc_dbe_agg_reg |
| 647 | - name: ecc_dbe_agg_shm |
| 648 | - name: ecc_dbe_agg_srm |
| 649 | - name: ecc_dbe_agg_tex |
| 650 | - name: ecc_dbe_vol_cbu |
| 651 | - name: ecc_dbe_vol_dev |
| 652 | - name: ecc_dbe_vol_l1 |
| 653 | - name: ecc_dbe_vol_l2 |
| 654 | - name: ecc_dbe_vol_reg |
| 655 | - name: ecc_dbe_vol_shm |
| 656 | - name: ecc_dbe_vol_srm |
| 657 | - name: ecc_dbe_vol_tex |
| 658 | - name: ecc_inforom_ver |
| 659 | - name: ecc_pending |
| 660 | - name: ecc_sbe_agg_cbu |
| 661 | - name: ecc_sbe_agg_dev |
| 662 | - name: ecc_sbe_agg_l1 |
| 663 | - name: ecc_sbe_agg_l2 |
| 664 | - name: ecc_sbe_agg_reg |
| 665 | - name: ecc_sbe_agg_shm |
| 666 | - name: ecc_sbe_agg_srm |
| 667 | - name: ecc_sbe_agg_tex |
| 668 | - name: ecc_sbe_vol_cbu |
| 669 | - name: ecc_sbe_vol_dev |
| 670 | - name: ecc_sbe_vol_l1 |
| 671 | - name: ecc_sbe_vol_l2 |
| 672 | - name: ecc_sbe_vol_reg |
| 673 | - name: ecc_sbe_vol_shm |
| 674 | - name: ecc_sbe_vol_srm |
| 675 | - name: ecc_sbe_vol_tex |
| 676 | - name: dcgm.gpu.memory.page_retirements |
| 677 | description: GPU Retired Memory Pages metrics. |
| 678 | unit: pages/s |
| 679 | chart_type: line |
| 680 | dimensions: |
| 681 | - name: retired_dbe |
| 682 | - name: retired_pending |
| 683 | - name: retired_sbe |
| 684 | - name: dcgm.gpu.memory.usage |
| 685 | description: GPU Memory Usage metrics. |
| 686 | unit: B |
| 687 | chart_type: stacked |
| 688 | dimensions: |
| 689 | - name: free |
| 690 | - name: reserved |
| 691 | - name: used |
| 692 | - name: dcgm.gpu.memory.capacity |
| 693 | description: GPU Memory Capacity metrics. |
| 694 | unit: B |
| 695 | chart_type: line |
| 696 | dimensions: |
| 697 | - name: total |
| 698 | - name: dcgm.gpu.memory.utilization |
| 699 | description: GPU Memory Utilization metrics. |
| 700 | unit: '%' |
| 701 | chart_type: line |
| 702 | dimensions: |
| 703 | - name: used_percent |
| 704 | - name: dcgm.gpu.power.energy |
| 705 | description: GPU Energy Consumption Rate metrics. |
| 706 | unit: mJ/s |
| 707 | chart_type: line |
| 708 | dimensions: |
| 709 | - name: total |
| 710 | - name: dcgm.gpu.power.profiles |
| 711 | description: GPU Power Profiles metrics. |
| 712 | unit: state |
| 713 | chart_type: line |
| 714 | dimensions: |
| 715 | - name: enforced_power_profile_mask |
| 716 | - name: requested_power_profile_mask |
| 717 | - name: valid_power_profile_mask |
| 718 | - name: dcgm.gpu.power.smoothing |
| 719 | description: GPU Power Smoothing metrics. |
| 720 | unit: value |
| 721 | chart_type: line |
| 722 | dimensions: |
| 723 | - name: pwr_smoothing_active_preset_profile |
| 724 | - name: pwr_smoothing_admin_override_percent_tmp_floor |
| 725 | - name: pwr_smoothing_admin_override_ramp_down_hyst_val |
| 726 | - name: pwr_smoothing_admin_override_ramp_down_rate |
| 727 | - name: pwr_smoothing_admin_override_ramp_up_rate |
| 728 | - name: pwr_smoothing_applied_tmp_ceil |
| 729 | - name: pwr_smoothing_applied_tmp_floor |
| 730 | - name: pwr_smoothing_enabled |
| 731 | - name: pwr_smoothing_hw_circuitry_percent_lifetime_remaining |
| 732 | - name: pwr_smoothing_imm_ramp_down_enabled |
| 733 | - name: pwr_smoothing_max_num_preset_profiles |
| 734 | - name: pwr_smoothing_max_percent_tmp_floor_setting |
| 735 | - name: pwr_smoothing_min_percent_tmp_floor_setting |
| 736 | - name: pwr_smoothing_priv_lvl |
| 737 | - name: pwr_smoothing_profile_percent_tmp_floor |
| 738 | - name: pwr_smoothing_profile_ramp_down_hyst_val |
| 739 | - name: pwr_smoothing_profile_ramp_down_rate |
| 740 | - name: pwr_smoothing_profile_ramp_up_rate |
| 741 | - name: dcgm.gpu.power.usage |
| 742 | description: GPU Power Usage metrics. |
| 743 | unit: Watts |
| 744 | chart_type: line |
| 745 | dimensions: |
| 746 | - name: draw |
| 747 | - name: enforced_limit |
| 748 | visibility: hidden |
| 749 | - name: power_mgmt_limit |
| 750 | visibility: hidden |
| 751 | - name: power_mgmt_limit_def |
| 752 | visibility: hidden |
| 753 | - name: power_mgmt_limit_max |
| 754 | visibility: hidden |
| 755 | - name: power_mgmt_limit_min |
| 756 | visibility: hidden |
| 757 | - name: power_usage_instant |
| 758 | - name: dcgm.gpu.reliability.memory_health |
| 759 | description: GPU Memory Health metrics. |
| 760 | unit: state |
| 761 | chart_type: line |
| 762 | dimensions: |
| 763 | - name: banks_remap_rows_avail_high |
| 764 | - name: banks_remap_rows_avail_low |
| 765 | - name: banks_remap_rows_avail_max |
| 766 | - name: banks_remap_rows_avail_none |
| 767 | - name: banks_remap_rows_avail_partial |
| 768 | - name: memory_unrepairable_flag |
| 769 | - name: threshold_srm |
| 770 | - name: dcgm.gpu.reliability.recovery_action |
| 771 | description: GPU Recovery Action metrics. |
| 772 | unit: state |
| 773 | chart_type: line |
| 774 | dimensions: |
| 775 | - name: get_gpu_recovery_action |
| 776 | - name: dcgm.gpu.reliability.row_remap_events |
| 777 | description: GPU Row Remap Events metrics. |
| 778 | unit: rows/s |
| 779 | chart_type: line |
| 780 | dimensions: |
| 781 | - name: correctable_remapped_rows |
| 782 | - name: uncorrectable_remapped_rows |
| 783 | - name: dcgm.gpu.reliability.row_remap_status |
| 784 | description: GPU Row Remap Status metrics. |
| 785 | unit: state |
| 786 | chart_type: line |
| 787 | dimensions: |
| 788 | - name: row_remap_failure |
| 789 | - name: row_remap_pending |
| 790 | - name: dcgm.gpu.reliability.xid |
| 791 | description: GPU XID Errors metrics. |
| 792 | unit: code |
| 793 | chart_type: line |
| 794 | dimensions: |
| 795 | - name: xid |
| 796 | - name: dcgm.gpu.state.configuration |
| 797 | description: GPU Configuration State metrics. |
| 798 | unit: state |
| 799 | chart_type: line |
| 800 | dimensions: |
| 801 | - name: autoboost |
| 802 | - name: compute_mode |
| 803 | - name: persistence_mode |
| 804 | - name: sync_boost |
| 805 | - name: sync_boost_violation |
| 806 | - name: dcgm.gpu.state.performance |
| 807 | description: GPU Performance State metrics. |
| 808 | unit: state |
| 809 | chart_type: line |
| 810 | dimensions: |
| 811 | - name: pstate |
| 812 | - name: dcgm.gpu.state.virtualization |
| 813 | description: GPU Virtualization State metrics. |
| 814 | unit: state |
| 815 | chart_type: line |
| 816 | dimensions: |
| 817 | - name: mig_mode |
| 818 | - name: virtual_mode |
| 819 | - name: dcgm.gpu.thermal.fan_speed |
| 820 | description: GPU Fan Speed metrics. |
| 821 | unit: '%' |
| 822 | chart_type: line |
| 823 | dimensions: |
| 824 | - name: fan_speed |
| 825 | - name: dcgm.gpu.thermal.temperature |
| 826 | description: GPU Temperature metrics. |
| 827 | unit: Celsius |
| 828 | chart_type: line |
| 829 | dimensions: |
| 830 | - name: connectx_device_temperature |
| 831 | - name: gpu |
| 832 | - name: gpu_max_op_temp |
| 833 | visibility: hidden |
| 834 | - name: gpu_temp_limit |
| 835 | visibility: hidden |
| 836 | - name: mem_max_op_temp |
| 837 | visibility: hidden |
| 838 | - name: memory |
| 839 | - name: shutdown_temp |
| 840 | visibility: hidden |
| 841 | - name: slowdown_temp |
| 842 | visibility: hidden |
| 843 | - name: dcgm.gpu.throttle.reasons |
| 844 | description: GPU Throttle Reasons metrics. |
| 845 | unit: bitmask |
| 846 | chart_type: line |
| 847 | dimensions: |
| 848 | - name: clocks_event_reasons |
| 849 | - name: dcgm.gpu.throttle.violations |
| 850 | description: GPU Throttle Violation Duration metrics. |
| 851 | unit: milliseconds/s |
| 852 | chart_type: line |
| 853 | dimensions: |
| 854 | - name: board_limit_violation |
| 855 | - name: hw_power_brake_slowdown |
| 856 | - name: hw_therm_slowdown |
| 857 | - name: low_utilization_violation |
| 858 | - name: power_violation |
| 859 | - name: reliability_violation |
| 860 | - name: sw_power_cap |
| 861 | - name: sw_therm_slowdown |
| 862 | - name: sync_boost |
| 863 | - name: thermal_violation |
| 864 | - name: total_app_clocks_violation |
| 865 | - name: total_base_clocks_violation |
| 866 | - name: dcgm.gpu.topology.affinity |
| 867 | description: GPU Topology and Affinity metrics. |
| 868 | unit: value |
| 869 | chart_type: line |
| 870 | dimensions: |
| 871 | - name: cpu_affinity_0 |
| 872 | - name: cpu_affinity_1 |
| 873 | - name: cpu_affinity_2 |
| 874 | - name: cpu_affinity_3 |
| 875 | - name: gpu_topology_affinity |
| 876 | - name: gpu_topology_pci |
| 877 | - name: mem_affinity_0 |
| 878 | - name: mem_affinity_1 |
| 879 | - name: mem_affinity_2 |
| 880 | - name: mem_affinity_3 |
| 881 | - name: pci_busid |
| 882 | - name: pci_combined_id |
| 883 | - name: pci_subsys_id |
| 884 | - name: dcgm.gpu.virtualization.vgpu.frame_rate |
| 885 | description: GPU vGPU Frame Rate metrics. |
| 886 | unit: fps |
| 887 | chart_type: line |
| 888 | dimensions: |
| 889 | - name: vgpu_frame_rate_limit |
| 890 | - name: dcgm.gpu.virtualization.vgpu.instance |
| 891 | description: GPU vGPU Instance metrics. |
| 892 | unit: value |
| 893 | chart_type: line |
| 894 | dimensions: |
| 895 | - name: vgpu_instance_ids |
| 896 | - name: vgpu_pci_id |
| 897 | - name: vgpu_uuid |
| 898 | - name: dcgm.gpu.virtualization.vgpu.license |
| 899 | description: GPU vGPU License metrics. |
| 900 | unit: state |
| 901 | chart_type: line |
| 902 | dimensions: |
| 903 | - name: vgpu_instance_license_state |
| 904 | - name: vgpu_license_status |
| 905 | - name: vgpu_type_license |
| 906 | - name: dcgm.gpu.virtualization.vgpu.memory |
| 907 | description: GPU vGPU Memory metrics. |
| 908 | unit: B |
| 909 | chart_type: line |
| 910 | dimensions: |
| 911 | - name: vgpu_memory_usage |
| 912 | - name: dcgm.gpu.virtualization.vgpu.sessions |
| 913 | description: GPU vGPU Sessions metrics. |
| 914 | unit: value |
| 915 | chart_type: line |
| 916 | dimensions: |
| 917 | - name: vgpu_enc_sessions_info |
| 918 | - name: vgpu_enc_stats |
| 919 | - name: vgpu_fbc_sessions_info |
| 920 | - name: vgpu_fbc_stats |
| 921 | - name: dcgm.gpu.virtualization.vgpu.software |
| 922 | description: GPU vGPU Software metrics. |
| 923 | unit: value |
| 924 | chart_type: line |
| 925 | dimensions: |
| 926 | - name: vgpu_driver_version |
| 927 | - name: dcgm.gpu.virtualization.vgpu.type |
| 928 | description: GPU vGPU Type metrics. |
| 929 | unit: value |
| 930 | chart_type: line |
| 931 | dimensions: |
| 932 | - name: creatable_vgpu_type_ids |
| 933 | - name: supported_vgpu_type_ids |
| 934 | - name: vgpu_type |
| 935 | - name: vgpu_type_class |
| 936 | - name: vgpu_type_info |
| 937 | - name: vgpu_type_name |
| 938 | - name: dcgm.gpu.virtualization.vgpu.utilization |
| 939 | description: GPU vGPU Utilization metrics. |
| 940 | unit: '%' |
| 941 | chart_type: line |
| 942 | dimensions: |
| 943 | - name: vgpu_per_process_utilization |
| 944 | - name: dcgm.gpu.virtualization.vgpu.vm |
| 945 | description: GPU vGPU VM metrics. |
| 946 | unit: value |
| 947 | chart_type: line |
| 948 | dimensions: |
| 949 | - name: vgpu_vm_gpu_instance_id |
| 950 | - name: vgpu_vm_id |
| 951 | - name: vgpu_vm_name |
| 952 | - name: dcgm.gpu.workload.sessions |
| 953 | description: GPU Workload Sessions metrics. |
| 954 | unit: value |
| 955 | chart_type: line |
| 956 | dimensions: |
| 957 | - name: accounting_data |
| 958 | - name: enc_stats |
| 959 | - name: fbc_sessions_info |
| 960 | - name: fbc_stats |
| 961 | - name: mig |
| 962 | description: These metrics refer to MIG instances. |
| 963 | labels: |
| 964 | - name: gpu |
| 965 | description: gpu label from exporter metrics. |
| 966 | - name: gpu_i_id |
| 967 | description: gpu_i_id label from exporter metrics. |
| 968 | - name: gpu_i_profile |
| 969 | description: gpu_i_profile label from exporter metrics. |
| 970 | metrics: |
| 971 | - name: dcgm.mig.clock.frequency |
| 972 | description: MIG Clock Frequency metrics. |
| 973 | unit: MHz |
| 974 | chart_type: line |
| 975 | dimensions: |
| 976 | - name: app_mem_clock |
| 977 | visibility: hidden |
| 978 | - name: app_sm_clock |
| 979 | visibility: hidden |
| 980 | - name: max_mem_clock |
| 981 | visibility: hidden |
| 982 | - name: max_sm_clock |
| 983 | visibility: hidden |
| 984 | - name: max_video_clock |
| 985 | visibility: hidden |
| 986 | - name: memory |
| 987 | - name: sm |
| 988 | - name: video_clock |
| 989 | - name: dcgm.mig.compute.activity |
| 990 | description: MIG Compute Pipeline Activity metrics. |
| 991 | unit: '%' |
| 992 | chart_type: line |
| 993 | dimensions: |
| 994 | - name: dram |
| 995 | - name: fp16 |
| 996 | - name: fp32 |
| 997 | - name: fp64 |
| 998 | - name: graphics_engine_active |
| 999 | - name: integer |
| 1000 | - name: sm_active |
| 1001 | - name: sm_occupancy |
| 1002 | - name: tensor |
| 1003 | - name: dcgm.mig.compute.tensor.activity |
| 1004 | description: MIG Tensor Core Activity by precision type. |
| 1005 | unit: '%' |
| 1006 | chart_type: line |
| 1007 | dimensions: |
| 1008 | - name: tensor_dfma |
| 1009 | - name: tensor_hmma |
| 1010 | - name: tensor_imma |
| 1011 | - name: dcgm.mig.compute.media.activity |
| 1012 | description: MIG Media Engine Activity metrics. |
| 1013 | unit: '%' |
| 1014 | chart_type: line |
| 1015 | dimensions: |
| 1016 | - name: nvdec0_active |
| 1017 | - name: nvdec1_active |
| 1018 | - name: nvdec2_active |
| 1019 | - name: nvdec3_active |
| 1020 | - name: nvdec4_active |
| 1021 | - name: nvdec5_active |
| 1022 | - name: nvdec6_active |
| 1023 | - name: nvdec7_active |
| 1024 | - name: nvjpg0_active |
| 1025 | - name: nvjpg1_active |
| 1026 | - name: nvjpg2_active |
| 1027 | - name: nvjpg3_active |
| 1028 | - name: nvjpg4_active |
| 1029 | - name: nvjpg5_active |
| 1030 | - name: nvjpg6_active |
| 1031 | - name: nvjpg7_active |
| 1032 | - name: nvofa0_active |
| 1033 | - name: nvofa1_active |
| 1034 | - name: dcgm.mig.compute.cache.activity |
| 1035 | description: MIG Memory Cache Hit/Miss metrics. |
| 1036 | unit: 'events/s' |
| 1037 | chart_type: line |
| 1038 | dimensions: |
| 1039 | - name: hostmem_cache_hit |
| 1040 | - name: hostmem_cache_miss |
| 1041 | - name: peermem_cache_hit |
| 1042 | - name: peermem_cache_miss |
| 1043 | - name: dcgm.mig.compute.utilization |
| 1044 | description: MIG Compute Utilization metrics. |
| 1045 | unit: '%' |
| 1046 | chart_type: line |
| 1047 | dimensions: |
| 1048 | - name: decoder |
| 1049 | - name: encoder |
| 1050 | - name: gpu |
| 1051 | - name: memory_copy |
| 1052 | - name: dcgm.mig.interconnect.nvlink.ber |
| 1053 | description: MIG NVLink Bit Error Rate metrics. |
| 1054 | unit: ratio |
| 1055 | chart_type: line |
| 1056 | dimensions: |
| 1057 | - name: nvlink_count_effective_ber |
| 1058 | - name: nvlink_count_effective_ber_float |
| 1059 | - name: nvlink_count_symbol_ber |
| 1060 | - name: nvlink_count_symbol_ber_float |
| 1061 | - name: dcgm.mig.interconnect.nvlink.congestion |
| 1062 | description: MIG NVLink Congestion metrics. |
| 1063 | unit: events/s |
| 1064 | chart_type: line |
| 1065 | dimensions: |
| 1066 | - name: nvlink_ppcnt_ibpc_port_xmit_wait |
| 1067 | - name: dcgm.mig.interconnect.error_rate |
| 1068 | description: MIG Interconnect Error Rate metrics. |
| 1069 | unit: errors/s |
| 1070 | chart_type: line |
| 1071 | dimensions: |
| 1072 | - name: c2c_link_error_intr |
| 1073 | - name: c2c_link_error_replay |
| 1074 | - name: c2c_link_error_replay_b2b |
| 1075 | - name: dcgm.mig.interconnect.nvlink.error_rate |
| 1076 | description: MIG NVLink Error Rate metrics. |
| 1077 | unit: errors/s |
| 1078 | chart_type: line |
| 1079 | dimensions: |
| 1080 | - name: gpu_nvlink_errors |
| 1081 | - name: nvlink_count_effective_errors |
| 1082 | - name: nvlink_count_fec_history_0 |
| 1083 | - name: nvlink_count_fec_history_1 |
| 1084 | - name: nvlink_count_fec_history_10 |
| 1085 | - name: nvlink_count_fec_history_11 |
| 1086 | - name: nvlink_count_fec_history_12 |
| 1087 | - name: nvlink_count_fec_history_13 |
| 1088 | - name: nvlink_count_fec_history_14 |
| 1089 | - name: nvlink_count_fec_history_15 |
| 1090 | - name: nvlink_count_fec_history_2 |
| 1091 | - name: nvlink_count_fec_history_3 |
| 1092 | - name: nvlink_count_fec_history_4 |
| 1093 | - name: nvlink_count_fec_history_5 |
| 1094 | - name: nvlink_count_fec_history_6 |
| 1095 | - name: nvlink_count_fec_history_7 |
| 1096 | - name: nvlink_count_fec_history_8 |
| 1097 | - name: nvlink_count_fec_history_9 |
| 1098 | - name: nvlink_count_link_recovery_events |
| 1099 | - name: nvlink_count_link_recovery_failed_events |
| 1100 | - name: nvlink_count_link_recovery_successful_events |
| 1101 | - name: nvlink_count_local_link_integrity_errors |
| 1102 | - name: nvlink_count_rx_buffer_overrun_errors |
| 1103 | - name: nvlink_count_rx_errors |
| 1104 | - name: nvlink_count_rx_general_errors |
| 1105 | - name: nvlink_count_rx_malformed_packet_errors |
| 1106 | - name: nvlink_count_rx_remote_errors |
| 1107 | - name: nvlink_count_rx_symbol_errors |
| 1108 | - name: nvlink_count_tx_discards |
| 1109 | - name: nvlink_crc_data_error |
| 1110 | - name: nvlink_crc_data_error_count_l0 |
| 1111 | - name: nvlink_crc_data_error_count_l1 |
| 1112 | - name: nvlink_crc_data_error_count_l10 |
| 1113 | - name: nvlink_crc_data_error_count_l11 |
| 1114 | - name: nvlink_crc_data_error_count_l12 |
| 1115 | - name: nvlink_crc_data_error_count_l13 |
| 1116 | - name: nvlink_crc_data_error_count_l14 |
| 1117 | - name: nvlink_crc_data_error_count_l15 |
| 1118 | - name: nvlink_crc_data_error_count_l16 |
| 1119 | - name: nvlink_crc_data_error_count_l17 |
| 1120 | - name: nvlink_crc_data_error_count_l2 |
| 1121 | - name: nvlink_crc_data_error_count_l3 |
| 1122 | - name: nvlink_crc_data_error_count_l4 |
| 1123 | - name: nvlink_crc_data_error_count_l5 |
| 1124 | - name: nvlink_crc_data_error_count_l6 |
| 1125 | - name: nvlink_crc_data_error_count_l7 |
| 1126 | - name: nvlink_crc_data_error_count_l8 |
| 1127 | - name: nvlink_crc_data_error_count_l9 |
| 1128 | - name: nvlink_crc_flit_error |
| 1129 | - name: nvlink_crc_flit_error_count_l0 |
| 1130 | - name: nvlink_crc_flit_error_count_l1 |
| 1131 | - name: nvlink_crc_flit_error_count_l10 |
| 1132 | - name: nvlink_crc_flit_error_count_l11 |
| 1133 | - name: nvlink_crc_flit_error_count_l12 |
| 1134 | - name: nvlink_crc_flit_error_count_l13 |
| 1135 | - name: nvlink_crc_flit_error_count_l14 |
| 1136 | - name: nvlink_crc_flit_error_count_l15 |
| 1137 | - name: nvlink_crc_flit_error_count_l16 |
| 1138 | - name: nvlink_crc_flit_error_count_l17 |
| 1139 | - name: nvlink_crc_flit_error_count_l2 |
| 1140 | - name: nvlink_crc_flit_error_count_l3 |
| 1141 | - name: nvlink_crc_flit_error_count_l4 |
| 1142 | - name: nvlink_crc_flit_error_count_l5 |
| 1143 | - name: nvlink_crc_flit_error_count_l6 |
| 1144 | - name: nvlink_crc_flit_error_count_l7 |
| 1145 | - name: nvlink_crc_flit_error_count_l8 |
| 1146 | - name: nvlink_crc_flit_error_count_l9 |
| 1147 | - name: nvlink_error_dl_crc |
| 1148 | - name: nvlink_error_dl_recovery |
| 1149 | - name: nvlink_error_dl_replay |
| 1150 | - name: nvlink_ppcnt_physical_successful_recovery_events |
| 1151 | - name: nvlink_ppcnt_plr_rcv_uncorrectable_code |
| 1152 | - name: nvlink_ppcnt_recovery_time_since_last |
| 1153 | - name: nvlink_ppcnt_recovery_total_successful_events |
| 1154 | - name: nvlink_pprm_oper_recovery |
| 1155 | - name: nvlink_recovery_error |
| 1156 | - name: nvlink_recovery_error_count_l0 |
| 1157 | - name: nvlink_recovery_error_count_l1 |
| 1158 | - name: nvlink_recovery_error_count_l10 |
| 1159 | - name: nvlink_recovery_error_count_l11 |
| 1160 | - name: nvlink_recovery_error_count_l12 |
| 1161 | - name: nvlink_recovery_error_count_l13 |
| 1162 | - name: nvlink_recovery_error_count_l14 |
| 1163 | - name: nvlink_recovery_error_count_l15 |
| 1164 | - name: nvlink_recovery_error_count_l16 |
| 1165 | - name: nvlink_recovery_error_count_l17 |
| 1166 | - name: nvlink_recovery_error_count_l2 |
| 1167 | - name: nvlink_recovery_error_count_l3 |
| 1168 | - name: nvlink_recovery_error_count_l4 |
| 1169 | - name: nvlink_recovery_error_count_l5 |
| 1170 | - name: nvlink_recovery_error_count_l6 |
| 1171 | - name: nvlink_recovery_error_count_l7 |
| 1172 | - name: nvlink_recovery_error_count_l8 |
| 1173 | - name: nvlink_recovery_error_count_l9 |
| 1174 | - name: nvlink_replay_error |
| 1175 | - name: nvlink_replay_error_count_l0 |
| 1176 | - name: nvlink_replay_error_count_l1 |
| 1177 | - name: nvlink_replay_error_count_l10 |
| 1178 | - name: nvlink_replay_error_count_l11 |
| 1179 | - name: nvlink_replay_error_count_l12 |
| 1180 | - name: nvlink_replay_error_count_l13 |
| 1181 | - name: nvlink_replay_error_count_l14 |
| 1182 | - name: nvlink_replay_error_count_l15 |
| 1183 | - name: nvlink_replay_error_count_l16 |
| 1184 | - name: nvlink_replay_error_count_l17 |
| 1185 | - name: nvlink_replay_error_count_l2 |
| 1186 | - name: nvlink_replay_error_count_l3 |
| 1187 | - name: nvlink_replay_error_count_l4 |
| 1188 | - name: nvlink_replay_error_count_l5 |
| 1189 | - name: nvlink_replay_error_count_l6 |
| 1190 | - name: nvlink_replay_error_count_l7 |
| 1191 | - name: nvlink_replay_error_count_l8 |
| 1192 | - name: nvlink_replay_error_count_l9 |
| 1193 | - name: dcgm.mig.interconnect.pcie.error_rate |
| 1194 | description: MIG PCIe Error Rate metrics. |
| 1195 | unit: errors/s |
| 1196 | chart_type: line |
| 1197 | dimensions: |
| 1198 | - name: pcie_count_correctable_errors |
| 1199 | - name: pcie_replay |
| 1200 | - name: dcgm.mig.interconnect.nvlink.errors |
| 1201 | description: MIG NVLink Errors metrics. |
| 1202 | unit: errors |
| 1203 | chart_type: line |
| 1204 | dimensions: |
| 1205 | - name: nvlink_ppcnt_plr_rcv_uncorrectable_code |
| 1206 | - name: dcgm.mig.interconnect.fabric |
| 1207 | description: MIG Fabric State metrics. |
| 1208 | unit: state |
| 1209 | chart_type: line |
| 1210 | dimensions: |
| 1211 | - name: fabric_clique_id |
| 1212 | - name: fabric_cluster_uuid |
| 1213 | - name: fabric_health_mask |
| 1214 | - name: fabric_manager_error_code |
| 1215 | - name: fabric_manager_status |
| 1216 | - name: dcgm.mig.interconnect.pcie.link.generation |
| 1217 | description: MIG PCIe Link Generation metrics. |
| 1218 | unit: generation |
| 1219 | chart_type: line |
| 1220 | dimensions: |
| 1221 | - name: link_gen |
| 1222 | - name: max_link_gen |
| 1223 | visibility: hidden |
| 1224 | - name: dcgm.mig.interconnect.pcie.link.width |
| 1225 | description: MIG PCIe Link Width metrics. |
| 1226 | unit: lanes |
| 1227 | chart_type: line |
| 1228 | dimensions: |
| 1229 | - name: link_width |
| 1230 | - name: max_link_width |
| 1231 | visibility: hidden |
| 1232 | - name: dcgm.mig.interconnect.state |
| 1233 | description: MIG Interconnect State metrics. |
| 1234 | unit: state |
| 1235 | chart_type: line |
| 1236 | dimensions: |
| 1237 | - name: c2c_link |
| 1238 | - name: c2c_link_power_state |
| 1239 | - name: c2c_link_status |
| 1240 | - name: dcgm.mig.interconnect.pcie.state |
| 1241 | description: MIG PCIe State metrics. |
| 1242 | unit: state |
| 1243 | chart_type: line |
| 1244 | dimensions: |
| 1245 | - name: diag_pcie_result |
| 1246 | - name: dcgm.mig.interconnect.nvlink.state |
| 1247 | description: MIG NVLink State metrics. |
| 1248 | unit: state |
| 1249 | chart_type: line |
| 1250 | dimensions: |
| 1251 | - name: gpu_topology_nvlink |
| 1252 | - name: nvlink_get_state |
| 1253 | - name: nvlink_ppcnt_physical_link_down_counter |
| 1254 | - name: nvlink_ppcnt_plr_rcv_code_err |
| 1255 | - name: nvlink_ppcnt_plr_sync_events |
| 1256 | - name: nvlink_ppcnt_plr_xmit_retry_events |
| 1257 | - name: p2p_nvlink_status |
| 1258 | - name: dcgm.mig.interconnect.throughput |
| 1259 | description: MIG Interconnect Throughput metrics. |
| 1260 | unit: B/s |
| 1261 | chart_type: area |
| 1262 | dimensions: |
| 1263 | - name: c2c_max_bandwidth |
| 1264 | - name: c2c_rx_all_bytes |
| 1265 | - name: c2c_rx_data_bytes |
| 1266 | - name: c2c_tx_all_bytes |
| 1267 | - name: c2c_tx_data_bytes |
| 1268 | - name: dcgm.mig.interconnect.nvlink.throughput |
| 1269 | description: MIG NVLink Throughput metrics. |
| 1270 | unit: B/s |
| 1271 | chart_type: area |
| 1272 | dimensions: |
| 1273 | - name: nvlink_bandwidth_l0 |
| 1274 | - name: nvlink_bandwidth_l1 |
| 1275 | - name: nvlink_bandwidth_l10 |
| 1276 | - name: nvlink_bandwidth_l11 |
| 1277 | - name: nvlink_bandwidth_l12 |
| 1278 | - name: nvlink_bandwidth_l13 |
| 1279 | - name: nvlink_bandwidth_l14 |
| 1280 | - name: nvlink_bandwidth_l15 |
| 1281 | - name: nvlink_bandwidth_l16 |
| 1282 | - name: nvlink_bandwidth_l17 |
| 1283 | - name: nvlink_bandwidth_l2 |
| 1284 | - name: nvlink_bandwidth_l3 |
| 1285 | - name: nvlink_bandwidth_l4 |
| 1286 | - name: nvlink_bandwidth_l5 |
| 1287 | - name: nvlink_bandwidth_l6 |
| 1288 | - name: nvlink_bandwidth_l7 |
| 1289 | - name: nvlink_bandwidth_l8 |
| 1290 | - name: nvlink_bandwidth_l9 |
| 1291 | - name: nvlink_count_rx |
| 1292 | - name: nvlink_count_tx |
| 1293 | - name: nvlink_l0_rx |
| 1294 | - name: nvlink_l0_tx |
| 1295 | - name: nvlink_l10_rx |
| 1296 | - name: nvlink_l10_tx |
| 1297 | - name: nvlink_l11_rx |
| 1298 | - name: nvlink_l11_tx |
| 1299 | - name: nvlink_l12_rx |
| 1300 | - name: nvlink_l12_tx |
| 1301 | - name: nvlink_l13_rx |
| 1302 | - name: nvlink_l13_tx |
| 1303 | - name: nvlink_l14_rx |
| 1304 | - name: nvlink_l14_tx |
| 1305 | - name: nvlink_l15_rx |
| 1306 | - name: nvlink_l15_tx |
| 1307 | - name: nvlink_l16_rx |
| 1308 | - name: nvlink_l16_tx |
| 1309 | - name: nvlink_l17_rx |
| 1310 | - name: nvlink_l17_tx |
| 1311 | - name: nvlink_l1_rx |
| 1312 | - name: nvlink_l1_tx |
| 1313 | - name: nvlink_l2_rx |
| 1314 | - name: nvlink_l2_tx |
| 1315 | - name: nvlink_l3_rx |
| 1316 | - name: nvlink_l3_tx |
| 1317 | - name: nvlink_l4_rx |
| 1318 | - name: nvlink_l4_tx |
| 1319 | - name: nvlink_l5_rx |
| 1320 | - name: nvlink_l5_tx |
| 1321 | - name: nvlink_l6_rx |
| 1322 | - name: nvlink_l6_tx |
| 1323 | - name: nvlink_l7_rx |
| 1324 | - name: nvlink_l7_tx |
| 1325 | - name: nvlink_l8_rx |
| 1326 | - name: nvlink_l8_tx |
| 1327 | - name: nvlink_l9_rx |
| 1328 | - name: nvlink_l9_tx |
| 1329 | - name: nvlink_rx_bandwidth |
| 1330 | - name: nvlink_rx_bandwidth_l0 |
| 1331 | - name: nvlink_rx_bandwidth_l1 |
| 1332 | - name: nvlink_rx_bandwidth_l10 |
| 1333 | - name: nvlink_rx_bandwidth_l11 |
| 1334 | - name: nvlink_rx_bandwidth_l12 |
| 1335 | - name: nvlink_rx_bandwidth_l13 |
| 1336 | - name: nvlink_rx_bandwidth_l14 |
| 1337 | - name: nvlink_rx_bandwidth_l15 |
| 1338 | - name: nvlink_rx_bandwidth_l16 |
| 1339 | - name: nvlink_rx_bandwidth_l17 |
| 1340 | - name: nvlink_rx_bandwidth_l2 |
| 1341 | - name: nvlink_rx_bandwidth_l3 |
| 1342 | - name: nvlink_rx_bandwidth_l4 |
| 1343 | - name: nvlink_rx_bandwidth_l5 |
| 1344 | - name: nvlink_rx_bandwidth_l6 |
| 1345 | - name: nvlink_rx_bandwidth_l7 |
| 1346 | - name: nvlink_rx_bandwidth_l8 |
| 1347 | - name: nvlink_rx_bandwidth_l9 |
| 1348 | - name: nvlink_rx |
| 1349 | - name: nvlink_tx_bandwidth |
| 1350 | - name: nvlink_tx_bandwidth_l0 |
| 1351 | - name: nvlink_tx_bandwidth_l1 |
| 1352 | - name: nvlink_tx_bandwidth_l10 |
| 1353 | - name: nvlink_tx_bandwidth_l11 |
| 1354 | - name: nvlink_tx_bandwidth_l12 |
| 1355 | - name: nvlink_tx_bandwidth_l13 |
| 1356 | - name: nvlink_tx_bandwidth_l14 |
| 1357 | - name: nvlink_tx_bandwidth_l15 |
| 1358 | - name: nvlink_tx_bandwidth_l16 |
| 1359 | - name: nvlink_tx_bandwidth_l17 |
| 1360 | - name: nvlink_tx_bandwidth_l2 |
| 1361 | - name: nvlink_tx_bandwidth_l3 |
| 1362 | - name: nvlink_tx_bandwidth_l4 |
| 1363 | - name: nvlink_tx_bandwidth_l5 |
| 1364 | - name: nvlink_tx_bandwidth_l6 |
| 1365 | - name: nvlink_tx_bandwidth_l7 |
| 1366 | - name: nvlink_tx_bandwidth_l8 |
| 1367 | - name: nvlink_tx_bandwidth_l9 |
| 1368 | - name: nvlink_tx |
| 1369 | - name: dcgm.mig.interconnect.pcie.throughput |
| 1370 | description: MIG PCIe Throughput metrics. |
| 1371 | unit: B/s |
| 1372 | chart_type: area |
| 1373 | dimensions: |
| 1374 | - name: pcie_rx |
| 1375 | - name: pcie_rx_throughput |
| 1376 | - name: pcie_tx |
| 1377 | - name: pcie_tx_throughput |
| 1378 | - name: dcgm.mig.interconnect.total.throughput |
| 1379 | description: MIG Interconnect Total Throughput metrics. |
| 1380 | unit: B/s |
| 1381 | chart_type: area |
| 1382 | dimensions: |
| 1383 | - name: pcie |
| 1384 | - name: nvlink |
| 1385 | - name: dcgm.mig.interconnect.nvlink.traffic |
| 1386 | description: MIG NVLink Traffic metrics. |
| 1387 | unit: events/s |
| 1388 | chart_type: line |
| 1389 | dimensions: |
| 1390 | - name: nvlink_count_rx_packets |
| 1391 | - name: nvlink_count_tx_packets |
| 1392 | - name: nvlink_ppcnt_plr_rcv_codes |
| 1393 | - name: nvlink_ppcnt_plr_xmit_codes |
| 1394 | - name: nvlink_ppcnt_plr_xmit_retry_codes |
| 1395 | - name: dcgm.mig.memory.bar1_usage |
| 1396 | description: MIG BAR1 Memory Usage metrics. |
| 1397 | unit: B |
| 1398 | chart_type: stacked |
| 1399 | dimensions: |
| 1400 | - name: free |
| 1401 | - name: used |
| 1402 | - name: dcgm.mig.memory.bar1_capacity |
| 1403 | description: MIG BAR1 Memory Capacity metrics. |
| 1404 | unit: B |
| 1405 | chart_type: line |
| 1406 | dimensions: |
| 1407 | - name: total |
| 1408 | - name: dcgm.mig.memory.ecc_error_rate |
| 1409 | description: MIG ECC Error Rate metrics. |
| 1410 | unit: errors/s |
| 1411 | chart_type: line |
| 1412 | dimensions: |
| 1413 | - name: ecc_current |
| 1414 | - name: ecc_dbe_agg |
| 1415 | - name: ecc_dbe_agg_cbu |
| 1416 | - name: ecc_dbe_agg_dev |
| 1417 | - name: ecc_dbe_agg_l1 |
| 1418 | - name: ecc_dbe_agg_l2 |
| 1419 | - name: ecc_dbe_agg_reg |
| 1420 | - name: ecc_dbe_agg_shm |
| 1421 | - name: ecc_dbe_agg_srm |
| 1422 | - name: ecc_dbe_agg_tex |
| 1423 | - name: ecc_dbe_vol |
| 1424 | - name: ecc_dbe_vol_cbu |
| 1425 | - name: ecc_dbe_vol_dev |
| 1426 | - name: ecc_dbe_vol_l1 |
| 1427 | - name: ecc_dbe_vol_l2 |
| 1428 | - name: ecc_dbe_vol_reg |
| 1429 | - name: ecc_dbe_vol_shm |
| 1430 | - name: ecc_dbe_vol_srm |
| 1431 | - name: ecc_dbe_vol_tex |
| 1432 | - name: ecc_pending |
| 1433 | - name: ecc_sbe_agg |
| 1434 | - name: ecc_sbe_agg_cbu |
| 1435 | - name: ecc_sbe_agg_dev |
| 1436 | - name: ecc_sbe_agg_l1 |
| 1437 | - name: ecc_sbe_agg_l2 |
| 1438 | - name: ecc_sbe_agg_reg |
| 1439 | - name: ecc_sbe_agg_shm |
| 1440 | - name: ecc_sbe_agg_srm |
| 1441 | - name: ecc_sbe_agg_tex |
| 1442 | - name: ecc_sbe_vol |
| 1443 | - name: ecc_sbe_vol_cbu |
| 1444 | - name: ecc_sbe_vol_dev |
| 1445 | - name: ecc_sbe_vol_l1 |
| 1446 | - name: ecc_sbe_vol_l2 |
| 1447 | - name: ecc_sbe_vol_reg |
| 1448 | - name: ecc_sbe_vol_shm |
| 1449 | - name: ecc_sbe_vol_srm |
| 1450 | - name: ecc_sbe_vol_tex |
| 1451 | - name: nvlink_ecc_data_error |
| 1452 | - name: dcgm.mig.memory.ecc_errors |
| 1453 | description: MIG ECC Errors metrics. |
| 1454 | unit: errors |
| 1455 | chart_type: line |
| 1456 | dimensions: |
| 1457 | - name: ecc_current |
| 1458 | - name: ecc_dbe_agg_cbu |
| 1459 | - name: ecc_dbe_agg_dev |
| 1460 | - name: ecc_dbe_agg_l1 |
| 1461 | - name: ecc_dbe_agg_l2 |
| 1462 | - name: ecc_dbe_agg_reg |
| 1463 | - name: ecc_dbe_agg_shm |
| 1464 | - name: ecc_dbe_agg_srm |
| 1465 | - name: ecc_dbe_agg_tex |
| 1466 | - name: ecc_dbe_vol_cbu |
| 1467 | - name: ecc_dbe_vol_dev |
| 1468 | - name: ecc_dbe_vol_l1 |
| 1469 | - name: ecc_dbe_vol_l2 |
| 1470 | - name: ecc_dbe_vol_reg |
| 1471 | - name: ecc_dbe_vol_shm |
| 1472 | - name: ecc_dbe_vol_srm |
| 1473 | - name: ecc_dbe_vol_tex |
| 1474 | - name: ecc_inforom_ver |
| 1475 | - name: ecc_pending |
| 1476 | - name: ecc_sbe_agg_cbu |
| 1477 | - name: ecc_sbe_agg_dev |
| 1478 | - name: ecc_sbe_agg_l1 |
| 1479 | - name: ecc_sbe_agg_l2 |
| 1480 | - name: ecc_sbe_agg_reg |
| 1481 | - name: ecc_sbe_agg_shm |
| 1482 | - name: ecc_sbe_agg_srm |
| 1483 | - name: ecc_sbe_agg_tex |
| 1484 | - name: ecc_sbe_vol_cbu |
| 1485 | - name: ecc_sbe_vol_dev |
| 1486 | - name: ecc_sbe_vol_l1 |
| 1487 | - name: ecc_sbe_vol_l2 |
| 1488 | - name: ecc_sbe_vol_reg |
| 1489 | - name: ecc_sbe_vol_shm |
| 1490 | - name: ecc_sbe_vol_srm |
| 1491 | - name: ecc_sbe_vol_tex |
| 1492 | - name: dcgm.mig.memory.page_retirements |
| 1493 | description: MIG Retired Memory Pages metrics. |
| 1494 | unit: pages/s |
| 1495 | chart_type: line |
| 1496 | dimensions: |
| 1497 | - name: retired_dbe |
| 1498 | - name: retired_pending |
| 1499 | - name: retired_sbe |
| 1500 | - name: dcgm.mig.memory.usage |
| 1501 | description: MIG Memory Usage metrics. |
| 1502 | unit: B |
| 1503 | chart_type: stacked |
| 1504 | dimensions: |
| 1505 | - name: free |
| 1506 | - name: reserved |
| 1507 | - name: used |
| 1508 | - name: dcgm.mig.memory.capacity |
| 1509 | description: MIG Memory Capacity metrics. |
| 1510 | unit: B |
| 1511 | chart_type: line |
| 1512 | dimensions: |
| 1513 | - name: total |
| 1514 | - name: dcgm.mig.memory.utilization |
| 1515 | description: MIG Memory Utilization metrics. |
| 1516 | unit: '%' |
| 1517 | chart_type: line |
| 1518 | dimensions: |
| 1519 | - name: used_percent |
| 1520 | - name: dcgm.mig.power.energy |
| 1521 | description: MIG Energy Consumption Rate metrics. |
| 1522 | unit: mJ/s |
| 1523 | chart_type: line |
| 1524 | dimensions: |
| 1525 | - name: total |
| 1526 | - name: dcgm.mig.power.profiles |
| 1527 | description: MIG Power Profiles metrics. |
| 1528 | unit: state |
| 1529 | chart_type: line |
| 1530 | dimensions: |
| 1531 | - name: enforced_power_profile_mask |
| 1532 | - name: requested_power_profile_mask |
| 1533 | - name: valid_power_profile_mask |
| 1534 | - name: dcgm.mig.power.smoothing |
| 1535 | description: MIG Power Smoothing metrics. |
| 1536 | unit: value |
| 1537 | chart_type: line |
| 1538 | dimensions: |
| 1539 | - name: pwr_smoothing_active_preset_profile |
| 1540 | - name: pwr_smoothing_admin_override_percent_tmp_floor |
| 1541 | - name: pwr_smoothing_admin_override_ramp_down_hyst_val |
| 1542 | - name: pwr_smoothing_admin_override_ramp_down_rate |
| 1543 | - name: pwr_smoothing_admin_override_ramp_up_rate |
| 1544 | - name: pwr_smoothing_applied_tmp_ceil |
| 1545 | - name: pwr_smoothing_applied_tmp_floor |
| 1546 | - name: pwr_smoothing_enabled |
| 1547 | - name: pwr_smoothing_hw_circuitry_percent_lifetime_remaining |
| 1548 | - name: pwr_smoothing_imm_ramp_down_enabled |
| 1549 | - name: pwr_smoothing_max_num_preset_profiles |
| 1550 | - name: pwr_smoothing_max_percent_tmp_floor_setting |
| 1551 | - name: pwr_smoothing_min_percent_tmp_floor_setting |
| 1552 | - name: pwr_smoothing_priv_lvl |
| 1553 | - name: pwr_smoothing_profile_percent_tmp_floor |
| 1554 | - name: pwr_smoothing_profile_ramp_down_hyst_val |
| 1555 | - name: pwr_smoothing_profile_ramp_down_rate |
| 1556 | - name: pwr_smoothing_profile_ramp_up_rate |
| 1557 | - name: dcgm.mig.power.usage |
| 1558 | description: MIG Power Usage metrics. |
| 1559 | unit: Watts |
| 1560 | chart_type: line |
| 1561 | dimensions: |
| 1562 | - name: draw |
| 1563 | - name: enforced_limit |
| 1564 | visibility: hidden |
| 1565 | - name: power_mgmt_limit |
| 1566 | visibility: hidden |
| 1567 | - name: power_mgmt_limit_def |
| 1568 | visibility: hidden |
| 1569 | - name: power_mgmt_limit_max |
| 1570 | visibility: hidden |
| 1571 | - name: power_mgmt_limit_min |
| 1572 | visibility: hidden |
| 1573 | - name: power_usage_instant |
| 1574 | - name: dcgm.mig.reliability.memory_health |
| 1575 | description: MIG Memory Health metrics. |
| 1576 | unit: state |
| 1577 | chart_type: line |
| 1578 | dimensions: |
| 1579 | - name: banks_remap_rows_avail_high |
| 1580 | - name: banks_remap_rows_avail_low |
| 1581 | - name: banks_remap_rows_avail_max |
| 1582 | - name: banks_remap_rows_avail_none |
| 1583 | - name: banks_remap_rows_avail_partial |
| 1584 | - name: memory_unrepairable_flag |
| 1585 | - name: threshold_srm |
| 1586 | - name: dcgm.mig.reliability.recovery_action |
| 1587 | description: MIG Recovery Action metrics. |
| 1588 | unit: state |
| 1589 | chart_type: line |
| 1590 | dimensions: |
| 1591 | - name: get_gpu_recovery_action |
| 1592 | - name: dcgm.mig.reliability.row_remap_events |
| 1593 | description: MIG Row Remap Events metrics. |
| 1594 | unit: rows/s |
| 1595 | chart_type: line |
| 1596 | dimensions: |
| 1597 | - name: correctable_remapped_rows |
| 1598 | - name: uncorrectable_remapped_rows |
| 1599 | - name: dcgm.mig.reliability.row_remap_status |
| 1600 | description: MIG Row Remap Status metrics. |
| 1601 | unit: state |
| 1602 | chart_type: line |
| 1603 | dimensions: |
| 1604 | - name: row_remap_failure |
| 1605 | - name: row_remap_pending |
| 1606 | - name: dcgm.mig.reliability.xid |
| 1607 | description: MIG XID Errors metrics. |
| 1608 | unit: code |
| 1609 | chart_type: line |
| 1610 | dimensions: |
| 1611 | - name: xid |
| 1612 | - name: dcgm.mig.state.configuration |
| 1613 | description: MIG Configuration State metrics. |
| 1614 | unit: state |
| 1615 | chart_type: line |
| 1616 | dimensions: |
| 1617 | - name: autoboost |
| 1618 | - name: compute_mode |
| 1619 | - name: persistence_mode |
| 1620 | - name: sync_boost |
| 1621 | - name: sync_boost_violation |
| 1622 | - name: dcgm.mig.state.performance |
| 1623 | description: MIG Performance State metrics. |
| 1624 | unit: state |
| 1625 | chart_type: line |
| 1626 | dimensions: |
| 1627 | - name: pstate |
| 1628 | - name: dcgm.mig.state.virtualization |
| 1629 | description: MIG Virtualization State metrics. |
| 1630 | unit: state |
| 1631 | chart_type: line |
| 1632 | dimensions: |
| 1633 | - name: mig_mode |
| 1634 | - name: virtual_mode |
| 1635 | - name: dcgm.mig.thermal.fan_speed |
| 1636 | description: MIG Fan Speed metrics. |
| 1637 | unit: '%' |
| 1638 | chart_type: line |
| 1639 | dimensions: |
| 1640 | - name: fan_speed |
| 1641 | - name: dcgm.mig.thermal.temperature |
| 1642 | description: MIG Temperature metrics. |
| 1643 | unit: Celsius |
| 1644 | chart_type: line |
| 1645 | dimensions: |
| 1646 | - name: gpu |
| 1647 | - name: gpu_max_op_temp |
| 1648 | visibility: hidden |
| 1649 | - name: gpu_temp_limit |
| 1650 | visibility: hidden |
| 1651 | - name: mem_max_op_temp |
| 1652 | visibility: hidden |
| 1653 | - name: memory |
| 1654 | - name: shutdown_temp |
| 1655 | visibility: hidden |
| 1656 | - name: slowdown_temp |
| 1657 | visibility: hidden |
| 1658 | - name: dcgm.mig.throttle.reasons |
| 1659 | description: MIG Throttle Reasons metrics. |
| 1660 | unit: bitmask |
| 1661 | chart_type: line |
| 1662 | dimensions: |
| 1663 | - name: clocks_event_reasons |
| 1664 | - name: dcgm.mig.throttle.violations |
| 1665 | description: MIG Throttle Violation Duration metrics. |
| 1666 | unit: milliseconds/s |
| 1667 | chart_type: line |
| 1668 | dimensions: |
| 1669 | - name: board_limit_violation |
| 1670 | - name: hw_power_brake_slowdown |
| 1671 | - name: hw_therm_slowdown |
| 1672 | - name: low_utilization_violation |
| 1673 | - name: power_violation |
| 1674 | - name: reliability_violation |
| 1675 | - name: sw_power_cap |
| 1676 | - name: sw_therm_slowdown |
| 1677 | - name: sync_boost |
| 1678 | - name: thermal_violation |
| 1679 | - name: total_app_clocks_violation |
| 1680 | - name: total_base_clocks_violation |
| 1681 | - name: nvlink |
| 1682 | description: These metrics refer to NVLink link instances. |
| 1683 | labels: |
| 1684 | - name: gpu |
| 1685 | description: gpu label from exporter metrics. |
| 1686 | - name: gpu_uuid |
| 1687 | description: gpu_uuid label from exporter metrics. |
| 1688 | - name: nvlink |
| 1689 | description: nvlink label from exporter metrics. |
| 1690 | metrics: |
| 1691 | - name: dcgm.nvlink.interconnect.ber |
| 1692 | description: NVLink Interconnect Bit Error Rate metrics. |
| 1693 | unit: ratio |
| 1694 | chart_type: line |
| 1695 | dimensions: |
| 1696 | - name: nvlink_count_effective_ber |
| 1697 | - name: nvlink_count_effective_ber_float |
| 1698 | - name: nvlink_count_symbol_ber |
| 1699 | - name: nvlink_count_symbol_ber_float |
| 1700 | - name: dcgm.nvlink.interconnect.congestion |
| 1701 | description: NVLink Interconnect Congestion metrics. |
| 1702 | unit: events/s |
| 1703 | chart_type: line |
| 1704 | dimensions: |
| 1705 | - name: nvlink_ppcnt_ibpc_port_xmit_wait |
| 1706 | - name: dcgm.nvlink.interconnect.error_rate |
| 1707 | description: NVLink Interconnect Error Rate metrics. |
| 1708 | unit: errors/s |
| 1709 | chart_type: line |
| 1710 | dimensions: |
| 1711 | - name: gpu_nvlink_errors |
| 1712 | - name: nvlink_count_effective_errors |
| 1713 | - name: nvlink_count_fec_history_0 |
| 1714 | - name: nvlink_count_fec_history_1 |
| 1715 | - name: nvlink_count_fec_history_10 |
| 1716 | - name: nvlink_count_fec_history_11 |
| 1717 | - name: nvlink_count_fec_history_12 |
| 1718 | - name: nvlink_count_fec_history_13 |
| 1719 | - name: nvlink_count_fec_history_14 |
| 1720 | - name: nvlink_count_fec_history_15 |
| 1721 | - name: nvlink_count_fec_history_2 |
| 1722 | - name: nvlink_count_fec_history_3 |
| 1723 | - name: nvlink_count_fec_history_4 |
| 1724 | - name: nvlink_count_fec_history_5 |
| 1725 | - name: nvlink_count_fec_history_6 |
| 1726 | - name: nvlink_count_fec_history_7 |
| 1727 | - name: nvlink_count_fec_history_8 |
| 1728 | - name: nvlink_count_fec_history_9 |
| 1729 | - name: nvlink_count_link_recovery_events |
| 1730 | - name: nvlink_count_link_recovery_failed_events |
| 1731 | - name: nvlink_count_link_recovery_successful_events |
| 1732 | - name: nvlink_count_local_link_integrity_errors |
| 1733 | - name: nvlink_count_rx_buffer_overrun_errors |
| 1734 | - name: nvlink_count_rx_errors |
| 1735 | - name: nvlink_count_rx_general_errors |
| 1736 | - name: nvlink_count_rx_malformed_packet_errors |
| 1737 | - name: nvlink_count_rx_remote_errors |
| 1738 | - name: nvlink_count_rx_symbol_errors |
| 1739 | - name: nvlink_count_tx_discards |
| 1740 | - name: nvlink_crc_data_error |
| 1741 | - name: nvlink_crc_data_error_count_l0 |
| 1742 | - name: nvlink_crc_data_error_count_l1 |
| 1743 | - name: nvlink_crc_data_error_count_l10 |
| 1744 | - name: nvlink_crc_data_error_count_l11 |
| 1745 | - name: nvlink_crc_data_error_count_l12 |
| 1746 | - name: nvlink_crc_data_error_count_l13 |
| 1747 | - name: nvlink_crc_data_error_count_l14 |
| 1748 | - name: nvlink_crc_data_error_count_l15 |
| 1749 | - name: nvlink_crc_data_error_count_l16 |
| 1750 | - name: nvlink_crc_data_error_count_l17 |
| 1751 | - name: nvlink_crc_data_error_count_l2 |
| 1752 | - name: nvlink_crc_data_error_count_l3 |
| 1753 | - name: nvlink_crc_data_error_count_l4 |
| 1754 | - name: nvlink_crc_data_error_count_l5 |
| 1755 | - name: nvlink_crc_data_error_count_l6 |
| 1756 | - name: nvlink_crc_data_error_count_l7 |
| 1757 | - name: nvlink_crc_data_error_count_l8 |
| 1758 | - name: nvlink_crc_data_error_count_l9 |
| 1759 | - name: nvlink_crc_flit_error |
| 1760 | - name: nvlink_crc_flit_error_count_l0 |
| 1761 | - name: nvlink_crc_flit_error_count_l1 |
| 1762 | - name: nvlink_crc_flit_error_count_l10 |
| 1763 | - name: nvlink_crc_flit_error_count_l11 |
| 1764 | - name: nvlink_crc_flit_error_count_l12 |
| 1765 | - name: nvlink_crc_flit_error_count_l13 |
| 1766 | - name: nvlink_crc_flit_error_count_l14 |
| 1767 | - name: nvlink_crc_flit_error_count_l15 |
| 1768 | - name: nvlink_crc_flit_error_count_l16 |
| 1769 | - name: nvlink_crc_flit_error_count_l17 |
| 1770 | - name: nvlink_crc_flit_error_count_l2 |
| 1771 | - name: nvlink_crc_flit_error_count_l3 |
| 1772 | - name: nvlink_crc_flit_error_count_l4 |
| 1773 | - name: nvlink_crc_flit_error_count_l5 |
| 1774 | - name: nvlink_crc_flit_error_count_l6 |
| 1775 | - name: nvlink_crc_flit_error_count_l7 |
| 1776 | - name: nvlink_crc_flit_error_count_l8 |
| 1777 | - name: nvlink_crc_flit_error_count_l9 |
| 1778 | - name: nvlink_error_dl_crc |
| 1779 | - name: nvlink_error_dl_recovery |
| 1780 | - name: nvlink_error_dl_replay |
| 1781 | - name: nvlink_ppcnt_physical_successful_recovery_events |
| 1782 | - name: nvlink_ppcnt_plr_rcv_uncorrectable_code |
| 1783 | - name: nvlink_ppcnt_recovery_time_since_last |
| 1784 | - name: nvlink_ppcnt_recovery_total_successful_events |
| 1785 | - name: nvlink_pprm_oper_recovery |
| 1786 | - name: nvlink_recovery_error |
| 1787 | - name: nvlink_recovery_error_count_l0 |
| 1788 | - name: nvlink_recovery_error_count_l1 |
| 1789 | - name: nvlink_recovery_error_count_l10 |
| 1790 | - name: nvlink_recovery_error_count_l11 |
| 1791 | - name: nvlink_recovery_error_count_l12 |
| 1792 | - name: nvlink_recovery_error_count_l13 |
| 1793 | - name: nvlink_recovery_error_count_l14 |
| 1794 | - name: nvlink_recovery_error_count_l15 |
| 1795 | - name: nvlink_recovery_error_count_l16 |
| 1796 | - name: nvlink_recovery_error_count_l17 |
| 1797 | - name: nvlink_recovery_error_count_l2 |
| 1798 | - name: nvlink_recovery_error_count_l3 |
| 1799 | - name: nvlink_recovery_error_count_l4 |
| 1800 | - name: nvlink_recovery_error_count_l5 |
| 1801 | - name: nvlink_recovery_error_count_l6 |
| 1802 | - name: nvlink_recovery_error_count_l7 |
| 1803 | - name: nvlink_recovery_error_count_l8 |
| 1804 | - name: nvlink_recovery_error_count_l9 |
| 1805 | - name: nvlink_replay_error |
| 1806 | - name: nvlink_replay_error_count_l0 |
| 1807 | - name: nvlink_replay_error_count_l1 |
| 1808 | - name: nvlink_replay_error_count_l10 |
| 1809 | - name: nvlink_replay_error_count_l11 |
| 1810 | - name: nvlink_replay_error_count_l12 |
| 1811 | - name: nvlink_replay_error_count_l13 |
| 1812 | - name: nvlink_replay_error_count_l14 |
| 1813 | - name: nvlink_replay_error_count_l15 |
| 1814 | - name: nvlink_replay_error_count_l16 |
| 1815 | - name: nvlink_replay_error_count_l17 |
| 1816 | - name: nvlink_replay_error_count_l2 |
| 1817 | - name: nvlink_replay_error_count_l3 |
| 1818 | - name: nvlink_replay_error_count_l4 |
| 1819 | - name: nvlink_replay_error_count_l5 |
| 1820 | - name: nvlink_replay_error_count_l6 |
| 1821 | - name: nvlink_replay_error_count_l7 |
| 1822 | - name: nvlink_replay_error_count_l8 |
| 1823 | - name: nvlink_replay_error_count_l9 |
| 1824 | - name: dcgm.nvlink.interconnect.errors |
| 1825 | description: NVLink Interconnect Errors metrics. |
| 1826 | unit: errors |
| 1827 | chart_type: line |
| 1828 | dimensions: |
| 1829 | - name: nvlink_ppcnt_plr_rcv_uncorrectable_code |
| 1830 | - name: dcgm.nvlink.interconnect.state |
| 1831 | description: NVLink Interconnect State metrics. |
| 1832 | unit: state |
| 1833 | chart_type: line |
| 1834 | dimensions: |
| 1835 | - name: gpu_topology_nvlink |
| 1836 | - name: nvlink_get_state |
| 1837 | - name: nvlink_ppcnt_physical_link_down_counter |
| 1838 | - name: nvlink_ppcnt_plr_rcv_code_err |
| 1839 | - name: nvlink_ppcnt_plr_sync_events |
| 1840 | - name: nvlink_ppcnt_plr_xmit_retry_events |
| 1841 | - name: p2p_nvlink_status |
| 1842 | - name: dcgm.nvlink.interconnect.throughput |
| 1843 | description: NVLink Interconnect Throughput metrics. |
| 1844 | unit: B/s |
| 1845 | chart_type: area |
| 1846 | dimensions: |
| 1847 | - name: nvlink_bandwidth |
| 1848 | - name: nvlink_bandwidth_l0 |
| 1849 | - name: nvlink_bandwidth_l1 |
| 1850 | - name: nvlink_bandwidth_l10 |
| 1851 | - name: nvlink_bandwidth_l11 |
| 1852 | - name: nvlink_bandwidth_l12 |
| 1853 | - name: nvlink_bandwidth_l13 |
| 1854 | - name: nvlink_bandwidth_l14 |
| 1855 | - name: nvlink_bandwidth_l15 |
| 1856 | - name: nvlink_bandwidth_l16 |
| 1857 | - name: nvlink_bandwidth_l17 |
| 1858 | - name: nvlink_bandwidth_l2 |
| 1859 | - name: nvlink_bandwidth_l3 |
| 1860 | - name: nvlink_bandwidth_l4 |
| 1861 | - name: nvlink_bandwidth_l5 |
| 1862 | - name: nvlink_bandwidth_l6 |
| 1863 | - name: nvlink_bandwidth_l7 |
| 1864 | - name: nvlink_bandwidth_l8 |
| 1865 | - name: nvlink_bandwidth_l9 |
| 1866 | - name: nvlink_count_rx |
| 1867 | - name: nvlink_count_tx |
| 1868 | - name: nvlink_l0_rx |
| 1869 | - name: nvlink_l0_tx |
| 1870 | - name: nvlink_l10_rx |
| 1871 | - name: nvlink_l10_tx |
| 1872 | - name: nvlink_l11_rx |
| 1873 | - name: nvlink_l11_tx |
| 1874 | - name: nvlink_l12_rx |
| 1875 | - name: nvlink_l12_tx |
| 1876 | - name: nvlink_l13_rx |
| 1877 | - name: nvlink_l13_tx |
| 1878 | - name: nvlink_l14_rx |
| 1879 | - name: nvlink_l14_tx |
| 1880 | - name: nvlink_l15_rx |
| 1881 | - name: nvlink_l15_tx |
| 1882 | - name: nvlink_l16_rx |
| 1883 | - name: nvlink_l16_tx |
| 1884 | - name: nvlink_l17_rx |
| 1885 | - name: nvlink_l17_tx |
| 1886 | - name: nvlink_l1_rx |
| 1887 | - name: nvlink_l1_tx |
| 1888 | - name: nvlink_l2_rx |
| 1889 | - name: nvlink_l2_tx |
| 1890 | - name: nvlink_l3_rx |
| 1891 | - name: nvlink_l3_tx |
| 1892 | - name: nvlink_l4_rx |
| 1893 | - name: nvlink_l4_tx |
| 1894 | - name: nvlink_l5_rx |
| 1895 | - name: nvlink_l5_tx |
| 1896 | - name: nvlink_l6_rx |
| 1897 | - name: nvlink_l6_tx |
| 1898 | - name: nvlink_l7_rx |
| 1899 | - name: nvlink_l7_tx |
| 1900 | - name: nvlink_l8_rx |
| 1901 | - name: nvlink_l8_tx |
| 1902 | - name: nvlink_l9_rx |
| 1903 | - name: nvlink_l9_tx |
| 1904 | - name: nvlink_rx_bandwidth |
| 1905 | - name: nvlink_rx_bandwidth_l0 |
| 1906 | - name: nvlink_rx_bandwidth_l1 |
| 1907 | - name: nvlink_rx_bandwidth_l10 |
| 1908 | - name: nvlink_rx_bandwidth_l11 |
| 1909 | - name: nvlink_rx_bandwidth_l12 |
| 1910 | - name: nvlink_rx_bandwidth_l13 |
| 1911 | - name: nvlink_rx_bandwidth_l14 |
| 1912 | - name: nvlink_rx_bandwidth_l15 |
| 1913 | - name: nvlink_rx_bandwidth_l16 |
| 1914 | - name: nvlink_rx_bandwidth_l17 |
| 1915 | - name: nvlink_rx_bandwidth_l2 |
| 1916 | - name: nvlink_rx_bandwidth_l3 |
| 1917 | - name: nvlink_rx_bandwidth_l4 |
| 1918 | - name: nvlink_rx_bandwidth_l5 |
| 1919 | - name: nvlink_rx_bandwidth_l6 |
| 1920 | - name: nvlink_rx_bandwidth_l7 |
| 1921 | - name: nvlink_rx_bandwidth_l8 |
| 1922 | - name: nvlink_rx_bandwidth_l9 |
| 1923 | - name: nvlink_rx |
| 1924 | - name: nvlink_tx_bandwidth |
| 1925 | - name: nvlink_tx_bandwidth_l0 |
| 1926 | - name: nvlink_tx_bandwidth_l1 |
| 1927 | - name: nvlink_tx_bandwidth_l10 |
| 1928 | - name: nvlink_tx_bandwidth_l11 |
| 1929 | - name: nvlink_tx_bandwidth_l12 |
| 1930 | - name: nvlink_tx_bandwidth_l13 |
| 1931 | - name: nvlink_tx_bandwidth_l14 |
| 1932 | - name: nvlink_tx_bandwidth_l15 |
| 1933 | - name: nvlink_tx_bandwidth_l16 |
| 1934 | - name: nvlink_tx_bandwidth_l17 |
| 1935 | - name: nvlink_tx_bandwidth_l2 |
| 1936 | - name: nvlink_tx_bandwidth_l3 |
| 1937 | - name: nvlink_tx_bandwidth_l4 |
| 1938 | - name: nvlink_tx_bandwidth_l5 |
| 1939 | - name: nvlink_tx_bandwidth_l6 |
| 1940 | - name: nvlink_tx_bandwidth_l7 |
| 1941 | - name: nvlink_tx_bandwidth_l8 |
| 1942 | - name: nvlink_tx_bandwidth_l9 |
| 1943 | - name: nvlink_tx |
| 1944 | - name: dcgm.nvlink.interconnect.traffic |
| 1945 | description: NVLink Interconnect Traffic metrics. |
| 1946 | unit: events/s |
| 1947 | chart_type: line |
| 1948 | dimensions: |
| 1949 | - name: nvlink_count_rx_packets |
| 1950 | - name: nvlink_count_tx_packets |
| 1951 | - name: nvlink_ppcnt_plr_rcv_codes |
| 1952 | - name: nvlink_ppcnt_plr_xmit_codes |
| 1953 | - name: nvlink_ppcnt_plr_xmit_retry_codes |
| 1954 | - name: dcgm.nvlink.internal.boundary |
| 1955 | description: NVLink Internal Boundary Fields metrics. |
| 1956 | unit: state |
| 1957 | chart_type: line |
| 1958 | dimensions: |
| 1959 | - name: nvlink_ppcnt_recovery_time_between_last_two |
| 1960 | - name: dcgm.nvlink.memory.ecc_error_rate |
| 1961 | description: NVLink ECC Error Rate metrics. |
| 1962 | unit: errors/s |
| 1963 | chart_type: line |
| 1964 | dimensions: |
| 1965 | - name: nvlink_ecc_data_error |
| 1966 | - name: nvswitch |
| 1967 | description: These metrics refer to NVSwitch instances. |
| 1968 | labels: |
| 1969 | - name: nvswitch |
| 1970 | description: nvswitch label from exporter metrics. |
| 1971 | metrics: |
| 1972 | - name: dcgm.nvswitch.interconnect.nvswitch.current |
| 1973 | description: NVSwitch Current metrics. |
| 1974 | unit: value |
| 1975 | chart_type: line |
| 1976 | dimensions: |
| 1977 | - name: nvswitch_current_iddq |
| 1978 | - name: nvswitch_current_iddq_dvdd |
| 1979 | - name: nvswitch_current_iddq_rev |
| 1980 | - name: dcgm.nvswitch.interconnect.nvswitch.errors |
| 1981 | description: NVSwitch NVSwitch Errors metrics. |
| 1982 | unit: errors/s |
| 1983 | chart_type: line |
| 1984 | dimensions: |
| 1985 | - name: nvswitch_fatal_errors |
| 1986 | - name: nvswitch_link_crc_errors |
| 1987 | - name: nvswitch_link_crc_errors_lane0 |
| 1988 | - name: nvswitch_link_crc_errors_lane1 |
| 1989 | - name: nvswitch_link_crc_errors_lane2 |
| 1990 | - name: nvswitch_link_crc_errors_lane3 |
| 1991 | - name: nvswitch_link_crc_errors_lane4 |
| 1992 | - name: nvswitch_link_crc_errors_lane5 |
| 1993 | - name: nvswitch_link_crc_errors_lane6 |
| 1994 | - name: nvswitch_link_crc_errors_lane7 |
| 1995 | - name: nvswitch_link_fatal_errors |
| 1996 | - name: nvswitch_link_flit_errors |
| 1997 | - name: nvswitch_link_non_fatal_errors |
| 1998 | - name: nvswitch_link_recovery_errors |
| 1999 | - name: nvswitch_link_replay_errors |
| 2000 | - name: nvswitch_non_fatal_errors |
| 2001 | - name: dcgm.nvswitch.interconnect.nvswitch.latency |
| 2002 | description: NVSwitch NVSwitch Link Latency metrics. |
| 2003 | unit: events/s |
| 2004 | chart_type: line |
| 2005 | dimensions: |
| 2006 | - name: nvswitch_link_latency_count_vc0 |
| 2007 | - name: nvswitch_link_latency_count_vc1 |
| 2008 | - name: nvswitch_link_latency_count_vc2 |
| 2009 | - name: nvswitch_link_latency_count_vc3 |
| 2010 | - name: nvswitch_link_latency_high_vc0 |
| 2011 | - name: nvswitch_link_latency_high_vc1 |
| 2012 | - name: nvswitch_link_latency_high_vc2 |
| 2013 | - name: nvswitch_link_latency_high_vc3 |
| 2014 | - name: nvswitch_link_latency_low_vc0 |
| 2015 | - name: nvswitch_link_latency_low_vc1 |
| 2016 | - name: nvswitch_link_latency_low_vc2 |
| 2017 | - name: nvswitch_link_latency_low_vc3 |
| 2018 | - name: nvswitch_link_latency_medium_vc0 |
| 2019 | - name: nvswitch_link_latency_medium_vc1 |
| 2020 | - name: nvswitch_link_latency_medium_vc2 |
| 2021 | - name: nvswitch_link_latency_medium_vc3 |
| 2022 | - name: nvswitch_link_latency_panic_vc0 |
| 2023 | - name: nvswitch_link_latency_panic_vc1 |
| 2024 | - name: nvswitch_link_latency_panic_vc2 |
| 2025 | - name: nvswitch_link_latency_panic_vc3 |
| 2026 | - name: dcgm.nvswitch.interconnect.nvswitch.power |
| 2027 | description: NVSwitch NVSwitch Power metrics. |
| 2028 | unit: Watts |
| 2029 | chart_type: line |
| 2030 | dimensions: |
| 2031 | - name: nvswitch_power_dvdd |
| 2032 | - name: nvswitch_power_hvdd |
| 2033 | - name: nvswitch_power_vdd |
| 2034 | - name: dcgm.nvswitch.interconnect.nvswitch.status |
| 2035 | description: NVSwitch NVSwitch Status metrics. |
| 2036 | unit: state |
| 2037 | chart_type: line |
| 2038 | dimensions: |
| 2039 | - name: nvswitch_link_status |
| 2040 | - name: nvswitch_link_type |
| 2041 | - name: nvswitch_reset_required |
| 2042 | - name: dcgm.nvswitch.interconnect.nvswitch.throughput |
| 2043 | description: NVSwitch NVSwitch Throughput metrics. |
| 2044 | unit: B/s |
| 2045 | chart_type: area |
| 2046 | dimensions: |
| 2047 | - name: nvswitch_link_throughput_rx |
| 2048 | - name: nvswitch_link_throughput_tx |
| 2049 | - name: nvswitch_throughput_rx |
| 2050 | - name: nvswitch_throughput_tx |
| 2051 | - name: dcgm.nvswitch.interconnect.nvswitch.topology |
| 2052 | description: NVSwitch NVSwitch Topology metrics. |
| 2053 | unit: value |
| 2054 | chart_type: line |
| 2055 | dimensions: |
| 2056 | - name: nvswitch_device_uuid |
| 2057 | - name: nvswitch_link_device_link_id |
| 2058 | - name: nvswitch_link_device_link_sid |
| 2059 | - name: nvswitch_link_id |
| 2060 | - name: nvswitch_link_remote_pcie_bus |
| 2061 | - name: nvswitch_link_remote_pcie_device |
| 2062 | - name: nvswitch_link_remote_pcie_domain |
| 2063 | - name: nvswitch_link_remote_pcie_function |
| 2064 | - name: nvswitch_pcie_bus |
| 2065 | - name: nvswitch_pcie_device |
| 2066 | - name: nvswitch_pcie_domain |
| 2067 | - name: nvswitch_pcie_function |
| 2068 | - name: nvswitch_phys_id |
| 2069 | - name: dcgm.nvswitch.interconnect.nvswitch.voltage |
| 2070 | description: NVSwitch NVSwitch Voltage metrics. |
| 2071 | unit: mV |
| 2072 | chart_type: line |
| 2073 | dimensions: |
| 2074 | - name: nvswitch_voltage_mvolt |
| 2075 | - name: dcgm.nvswitch.internal.boundary |
| 2076 | description: NVSwitch Internal Boundary Fields metrics. |
| 2077 | unit: state |
| 2078 | chart_type: line |
| 2079 | dimensions: |
| 2080 | - name: first_nvswitch_field_id |
| 2081 | - name: last_nvswitch_field_id |
| 2082 | - name: dcgm.nvswitch.memory.ecc_error_rate |
| 2083 | description: NVSwitch ECC Error Rate metrics. |
| 2084 | unit: errors/s |
| 2085 | chart_type: line |
| 2086 | dimensions: |
| 2087 | - name: nvswitch_link_ecc_errors |
| 2088 | - name: nvswitch_link_ecc_errors_lane0 |
| 2089 | - name: nvswitch_link_ecc_errors_lane1 |
| 2090 | - name: nvswitch_link_ecc_errors_lane2 |
| 2091 | - name: nvswitch_link_ecc_errors_lane3 |
| 2092 | - name: nvswitch_link_ecc_errors_lane4 |
| 2093 | - name: nvswitch_link_ecc_errors_lane5 |
| 2094 | - name: nvswitch_link_ecc_errors_lane6 |
| 2095 | - name: nvswitch_link_ecc_errors_lane7 |
| 2096 | - name: dcgm.nvswitch.thermal.temperature |
| 2097 | description: NVSwitch Temperature metrics. |
| 2098 | unit: Celsius |
| 2099 | chart_type: line |
| 2100 | dimensions: |
| 2101 | - name: nvswitch_temperature_current |
| 2102 | - name: nvswitch_temperature_limit_shutdown |
| 2103 | - name: nvswitch_temperature_limit_slowdown |
| 2104 | - name: cpu |
| 2105 | description: These metrics refer to host CPU instances. |
| 2106 | labels: |
| 2107 | - name: cpu |
| 2108 | description: cpu label from exporter metrics. |
| 2109 | metrics: |
| 2110 | - name: dcgm.cpu.clock.frequency |
| 2111 | description: CPU Clock Frequency metrics. |
| 2112 | unit: MHz |
| 2113 | chart_type: line |
| 2114 | dimensions: |
| 2115 | - name: cpu_clock_current |
| 2116 | - name: dcgm.cpu.cpu.info |
| 2117 | description: CPU CPU Information metrics. |
| 2118 | unit: value |
| 2119 | chart_type: line |
| 2120 | dimensions: |
| 2121 | - name: cpu_model |
| 2122 | - name: cpu_vendor |
| 2123 | - name: dcgm.cpu.cpu.power |
| 2124 | description: CPU CPU Power metrics. |
| 2125 | unit: Watts |
| 2126 | chart_type: line |
| 2127 | dimensions: |
| 2128 | - name: cpu_power_limit |
| 2129 | - name: cpu_power_util_current |
| 2130 | - name: dcgm.cpu.cpu.temperature |
| 2131 | description: CPU CPU Temperature metrics. |
| 2132 | unit: Celsius |
| 2133 | chart_type: line |
| 2134 | dimensions: |
| 2135 | - name: cpu_temp_critical |
| 2136 | - name: cpu_temp_current |
| 2137 | - name: cpu_temp_warning |
| 2138 | - name: dcgm.cpu.cpu.utilization |
| 2139 | description: CPU CPU Utilization metrics. |
| 2140 | unit: '%' |
| 2141 | chart_type: line |
| 2142 | dimensions: |
| 2143 | - name: cpu_util |
| 2144 | - name: cpu_util_irq |
| 2145 | - name: cpu_util_nice |
| 2146 | - name: cpu_util_sys |
| 2147 | - name: cpu_util_user |
| 2148 | - name: dcgm.cpu.diagnostics.results |
| 2149 | description: CPU Diagnostics Results metrics. |
| 2150 | unit: state |
| 2151 | chart_type: line |
| 2152 | dimensions: |
| 2153 | - name: diag_cpu_eud_result |
| 2154 | - name: cpu_core |
| 2155 | description: These metrics refer to host CPU core instances. |
| 2156 | labels: |
| 2157 | - name: cpu |
| 2158 | description: cpu label from exporter metrics. |
| 2159 | - name: cpucore |
| 2160 | description: cpucore label from exporter metrics. |
| 2161 | metrics: |
| 2162 | - name: dcgm.cpu_core.clock.frequency |
| 2163 | description: CPU Core Clock Frequency metrics. |
| 2164 | unit: MHz |
| 2165 | chart_type: line |
| 2166 | dimensions: |
| 2167 | - name: cpu_clock_current |
| 2168 | - name: dcgm.cpu_core.cpu.info |
| 2169 | description: CPU Core CPU Information metrics. |
| 2170 | unit: value |
| 2171 | chart_type: line |
| 2172 | dimensions: |
| 2173 | - name: cpu_model |
| 2174 | - name: cpu_vendor |
| 2175 | - name: dcgm.cpu_core.cpu.power |
| 2176 | description: CPU Core CPU Power metrics. |
| 2177 | unit: Watts |
| 2178 | chart_type: line |
| 2179 | dimensions: |
| 2180 | - name: cpu_power_limit |
| 2181 | - name: cpu_power_util_current |
| 2182 | - name: dcgm.cpu_core.cpu.temperature |
| 2183 | description: CPU Core CPU Temperature metrics. |
| 2184 | unit: Celsius |
| 2185 | chart_type: line |
| 2186 | dimensions: |
| 2187 | - name: cpu_temp_critical |
| 2188 | - name: cpu_temp_current |
| 2189 | - name: cpu_temp_warning |
| 2190 | - name: dcgm.cpu_core.cpu.utilization |
| 2191 | description: CPU Core CPU Utilization metrics. |
| 2192 | unit: '%' |
| 2193 | chart_type: line |
| 2194 | dimensions: |
| 2195 | - name: cpu_util |
| 2196 | - name: cpu_util_irq |
| 2197 | - name: cpu_util_nice |
| 2198 | - name: cpu_util_sys |
| 2199 | - name: cpu_util_user |
| 2200 | - name: dcgm.cpu_core.diagnostics.results |
| 2201 | description: CPU Core Diagnostics Results metrics. |
| 2202 | unit: state |
| 2203 | chart_type: line |
| 2204 | dimensions: |
| 2205 | - name: diag_cpu_eud_result |
| 2206 | - name: exporter |
| 2207 | description: These metrics refer to exporter/global instances. |
| 2208 | labels: |
| 2209 | - name: job |
| 2210 | description: job label from exporter metrics. |
| 2211 | metrics: |
| 2212 | - name: dcgm.exporter.health.status |
| 2213 | description: Exporter Health Status metrics. |
| 2214 | unit: state |
| 2215 | chart_type: line |
| 2216 | dimensions: |
| 2217 | - name: bind_unbind_event |
| 2218 | - name: dcgm.exporter.inventory.software |
| 2219 | description: Exporter Software and Firmware metrics. |
| 2220 | unit: value |
| 2221 | chart_type: line |
| 2222 | dimensions: |
| 2223 | - name: cuda_driver_version |
| 2224 | - name: driver_version |
| 2225 | - name: nvml_version |