What's New | NVIDIA Minerva User Guide

What's New in NVIDIA Minerva 1.1.0

NVLink management: Register NMX-C and NMX-T services, view the NVLink inventory, and create and manage GPU partitions with the new NVIDIA Minerva NVLink API, which uses the same paths as the NetQ NVLink (NMX-M) API for these operations. See NVLink Management.
Vera Rubin NVLink telemetry: NVIDIA Minerva collects NVLink switch telemetry from Vera Rubin racks through NMX-T, the same way it does for GB racks. See Connect to Data Sources.
High-Frequency Telemetry (HFT, beta): Collect switch counters at intervals as short as 100 microseconds from Spectrum-4 and later switches running Cumulus Linux 5.18 or later. Start a collection from the UI or the HFT API. NVIDIA Minerva configures the switches, stores every sample, and returns each switch to its previous configuration when the collection ends. See High-Frequency Telemetry.

BGP sessions: BGP sessions are modeled in the network graph and shown in Topology, so you can correlate BGP state changes with fabric anomalies. The new BGP Operations dashboard tracks session health and recovery. See BGP Sessions and BGP Operations.
Job topology with anomalies and events: Select a job to see its topology and the events and anomalies that occurred during the job, with servers shaded by event and anomaly count. See Job Topology.
Topology performance: The topology loads faster and stays responsive on large, dense fabrics. Labels and icons are hidden as you zoom out so you can scan the whole fabric, and auto layout keeps nodes arranged in tiers. See Topology.
North-south pod groups: Switches labeled fabric_type=north-south are placed in their own pod groups, named Pod-NS-<pod>, so north-south and east-west pods with the same number stay separate. See Assign Device Roles and Groups.
Adaptive routing congestion debugging: A new AR Congestion Debug dashboard puts the Cumulus Linux adaptive routing counter next to buffer occupancy, carrier changes, ECN marking, and drops, so you can tell fabric congestion apart from link instability. You can also ask Sherlock about it. See Adaptive Routing Congestion.

Interface metric retention: NVLink interface metrics now have their own table, interface_nvlink_v1, with its own retention setting. By default, switch and NIC interface metrics (interface_v1) and NVLink interface metrics are each kept for 10 days. See Data retention in Install on a Single Node.

Sherlock GA: Sherlock is generally available. It can add a line chart, bar chart, heatmap, or table to its answers, and you can pin, rename, and rate conversations and continue an investigation where you left off. You can also open Sherlock from the Trends page to ask about the view on screen. See Ask Sherlock.
NVIDIA Minerva MCP GA: The NVIDIA Minerva Model Context Protocol (MCP) server is generally available. AI assistants that support MCP can query NVIDIA Minerva data, read-only, after a user authorizes them. See Connect AI Assistants with NVIDIA Minerva MCP.

Anomaly feedback view: A new default Trends view shows the feedback users have given on anomalies over the last 30 days. Trends bar charts can now use a single dimension. See Give Anomaly Feedback.
Dashboard sections: Group dashboard widgets into sections that you can collapse and expand. See Organize a Dashboard into Sections.

OpenTelemetry notification endpoints (alpha): Send notifications as OTLP log records to an OpenTelemetry receiver, such as an OpenTelemetry Collector. See OpenTelemetry Endpoints.

Kubernetes installation: You can now install NVIDIA Minerva on a Kubernetes cluster you own and operate. The NVIDIA Minerva operator, installed with Helm, deploys the platform and collector stacks together from three custom resources. See Kubernetes Installation.
ARM64 support: NVIDIA Minerva now runs on ARM64 (aarch64) hosts as well as x86-64 (amd64) hosts.
UI usage statistics: Platform Admins can collect statistics on how NVIDIA Minerva UI features are used and export them for analysis. Collection is off by default. See Collect Usage Statistics.
High availability: In a multi-node deployment with shared NFS storage, NVIDIA Minerva can keep running when a node fails. See Install on Multiple Nodes with High Availability.
Ceph deep storage: NVIDIA Minerva can store historical time-series data in a Ceph cluster that you run on your own hosts, through the Ceph Object Gateway's S3-compatible API. Choose s3-compatible at the installer's time-series storage prompt. See Use Ceph for Deep Storage.

To install NVIDIA Minerva 1.1.0, see Get Started. To upgrade a Docker Swarm deployment from 1.0.x, see Upgrade NVIDIA Minerva on Docker Swarm.

A license is now required. NVIDIA Minerva 1.1.0 does not include a built-in evaluation license. A new installation, or an upgrade of a 1.0.x deployment that runs on the built-in evaluation license, starts in the invalid state until you install an evaluation or production license from the NVIDIA Licensing Portal. See Manage Licenses.
New installation method: Kubernetes. Kubernetes is now available as an alternative to Docker Swarm. See Kubernetes Requirements.
New installation option: high availability. To install NVIDIA Minerva with high availability, see Install on Multiple Nodes with High Availability.
MySQL 8.4. NVIDIA Minerva 1.1.0 uses MySQL 8.4. When you upgrade from 1.0.x, the existing MySQL 5.7 data is converted to MySQL 8.4 the first time you run ./svm stack-up, so make sure there's enough free disk space first. See Upgrade NVIDIA Minerva on Docker Swarm.
MCP clients might ask for authorization again. After the upgrade, MCP clients might ask for authorization once.

For the open issues in this release, see Open and Fixed Issues.

Last updated: November 03, 2026