NVIDIA Releases Fleet Intelligence to Provide End to End Visibility and Operational Integrity for Large Scale Data Center GPU Deployments Across Blackwell and Hopper Platforms
NVIDIA has now officially released Fleet Intelligence, a managed service offering end to end, round the clock visibility for its customers’ at scale GPU deployments. This service leverages a low footprint host based agent to report hardware telemetry to a central cloud portal. The software has been designed, according to an NVIDIA technical release, to handle both heterogeneous hardware and power constrained scenarios. It is offered as free to users of NVIDIA data center hardware, such as their Vera Rubin, Blackwell, and Hopper architecture platforms.
Fleet Intelligence concentrates its monitoring efforts on three main operational components, including inventory visualization, health alerts and system integrity monitoring. IT departments can be confident in an IT department auditing the agent on any system due to its open source release on GitHub. When integrated with systems like NVIDIA Data Center GPU Manager and GPUd, the agent can also report power utilization trends, thermal behavior, and performance issues that arise due to issues in system hardware. Memory bandwidth and the stability of interconnects can be analyzed to proactively monitor for component failure before system wide downtime can occur.
"the platform provides end to end visibility and catches early warning signs of hardware fatigue across their Blackwell and Hopper clusters."
Chuan Li, Chief Scientific Officer for Lambda
The thermal behavior and power usage of the cluster will also be analyzed to allow clusters to adhere to strict data center energy budgets. The detection of hot spots or airflow inefficiencies will be logged as measures of to mitigate early wear out of components. Direct alerts are sent via channels like email or Slack when the software detects an XID signal, ECC error, or retired page. Maintaining consistent drivers, BIOS versions, and firmware updates on more than 100,000 accelerators is crucial for ensuring 100 percent reproducible results from each computational element.
An attestation system verifies the trustworthiness of all GPUs within a given fleet. The Fleet Intelligence agent utilizes a proprietary Attestation SDK to get hardware metrics at runtime which can then be digitally signed using certificates rooted in NVIDIA’s own system. This attestation data can be checked against the Reference Integrity Manifest for deviations that indicate possible tampering with the vBIOS or firmware on the device. Attestation can be scheduled as part of an ongoing compliance audit or on demand to maintain the integrity of the fleet.
The attestation system currently supports Vera Rubin and Blackwell hardware. For users managing their own infrastructure, NVIDIA has a global dashboard on their NGC platform that displays the utilization of their fleet across various compute zones. Remedial suggestions for fixing hardware issues can be made based on historical error patterns. The agent only performs read operations and thus does not modify the host configuration as it gathers telemetry data, crucial for 2026 ROI with continued scaling.
