NVLink management lets you use NVIDIA Minerva to manage the NVLink fabric in your AI factory. Through the NVIDIA Minerva NVLink API, you can register the NVLink switch services, view the NVLink inventory, and create and manage GPU partitions. NVLink management is separate from the NVLink dashboards, which monitor NVLink telemetry.
NVLink management is available on Docker Swarm installations only. Kubernetes installations don't include it yet.
Use NVLink management to split the GPUs in your NVLink racks into partitions for different workloads. The NVIDIA Minerva NVLink API uses the same /nmx/v1 paths as the NetQ NVLink (NMX-M) API for service registration, inventory, partitions, operations, and settings. Switch profiles use /minerva/v1. The NetQ NVLink KPI, metrics, health, firmware validation, webhook, certificate, switch bring-up, image, support package, and switch upgrade endpoints aren't available.
How NVLink Management Works
NVLink management uses three services:
|
Service |
Role |
|---|---|
|
NMX-C (controller) |
Runs on the NVLink switch trays and controls the NVLink fabric. |
|
NMX-T (telemetry) |
Runs on the NVLink switch trays and provides NVLink telemetry. |
|
NMX-M (manager) |
Runs on the NVIDIA Minerva Platform Stack. It provides the NVIDIA Minerva NVLink API and registers with NMX-C and NMX-T. |
Before You Begin
-
NVLink racks with GB200, GB300, or Vera Rubin systems.
-
An on-premises Docker Swarm installation of NVIDIA Minerva: either a single-node installation, or a platform with remote collectors. The collectors must be able to reach the NVLink switch trays. See Get Started and, for remote collectors, Install Collectors for a Hybrid Deployment.
-
Configure an API key for the NVIDIA Minerva NVLink API. See Configure an API Key.
Set Up NVLink Management
-
Register the NMX-C and NMX-T services. See Register NVLink Services.
-
Check the GPUs in the NVLink fabric. See Query GPU Inventory.
-
Create GPU partitions. See Manage GPU Partitions.
Last updated: