NVIDIA details Minerva’s job-to-hardware graph and MCP investigation interface

NVIDIA’s [October 26 technical post on Minerva](https://developer.nvidia.com/blog/how‑nvidia‑minerva‑turns‑telemetry‑chaos‑into‑peak‑gpu‑yield/) explains how its AI‑operations platform connects a workload to the physical and software resources behind it. Starting with a job ID, an engineer can follow the relationship through GPUs, hosts, network interface cards, switches and logs, according to the company’s description. That is a different starting point from checking a GPU dashboard and a network dashboard separately. The workload provides the thread joining the observations: which resources the job occupies, where they sit in the topology and what evidence is attached to them.

Minerva [collects metrics, logs and topology](https://developer.nvidia.com/blog/how‑nvidia‑minerva‑turns‑telemetry‑chaos‑into‑peak‑gpu‑yield/) from GPUs, networking equipment, storage, schedulers and AI workloads. NVIDIA describes a knowledge graph that maps the relationships among infrastructure, scheduling and workload elements and updates those relationships as the environment changes. The graph can be navigated in either direction. An operator starting from an interface can trace its host, network path and affected jobs; an engineer starting from a job can work back toward the relevant devices and logs. That relationship layer is the practical addition to the raw telemetry described in the post. NVIDIA also describes correlating events that share a dependency path into an incident with the affected topology and workloads attached. Grouping related observations is a way to organize an investigation. It does not, by itself, demonstrate that the system has correctly identified the physical cause of a particular failure.

The [post describes Sherlock as Minerva’s native AI agent](https://developer.nvidia.com/blog/how‑nvidia‑minerva‑turns‑telemetry‑chaos‑into‑peak‑gpu‑yield/), grounded in live telemetry and the knowledge graph. An operator can ask about a job’s behavior during an interval, changes in component state or other nodes in the same failure domain. NVIDIA says Sherlock can provide supporting evidence and recommend remediation. Minerva’s Model Context Protocol server lets external agentic tools query the same operational context. This gives an outside assistant a route to the telemetry and relationships used by the platform, rather than requiring a separate inventory of the environment. A recommendation still needs to be assessed against the evidence behind it. Querying that context and executing a mitigation are separate parts of the operating model. NVIDIA says mitigation can remain subject to human approval, while selected actions can be automated according to local policy and the risk of the action. The description does not establish that every installation permits an agent to drain nodes or change infrastructure without approval.

NVIDIA [describes Minerva’s development alongside DSX OS](https://developer.nvidia.com/blog/how‑nvidia‑minerva‑turns‑telemetry‑chaos‑into‑peak‑gpu‑yield/), including plans to integrate selected health and resilience components. It says Minerva capabilities beginning with the data collector will become open‑source parts of DSX OS. That is a stated future direction, not a statement that the whole Minerva platform is already open source. An operator interested in the collector should distinguish that roadmap commitment from an already released component and check the applicable release and deployment documentation before building around it. The post does not supply a dated open‑source release or a license for the entire platform. NVIDIA’s headquarters in Santa Clara, California, photographed August 4, 2018. Historical company context, not a Minerva deployment or 2026 product image. Photo: [Coolcaesar](https://commons.wikimedia.org/wiki/User:Coolcaesar); source perspective/crop adaptation by [Jacek Halicki](https://commons.wikimedia.org/wiki/User:Jacek_Halicki), [via Wikimedia Commons](https://commons.wikimedia.org/wiki/File:NVIDIA_Headquarters.jpg), [CC BY‑SA 4.0](https://creativecommons.org/licenses/by‑sa/4.0/). Resized and cropped for display; adaptations remain under the same licence.