Interoperability
IO and platform requirements
The following IO versions are required for the AIFO integration:
Infrastructure Observability version 2026.4.1 or later must be installed and configured.
Each AI Factory host must run the AIFO Collector service. The Collector gathers metrics from the host and exposes them for polling.
On each AI Factory host, the nv-hostengine service must be running and accessible from the Collector container.
The Collector retrieves metrics from NVIDIA GPUs by using NVIDIA tools and packages that are installed on the host.
The AIFO Collector runs in one of the following environments:
As a container in a Linux container environment, such as Docker or Kubernetes
Directly on a Windows host, as a Windows service (for example, on a Hyper-V server)
Linux GPU software requirements
Component | Required version |
|---|---|
NVIDIA-SMI | 570.133.20 |
Driver | 570.133.20 |
CUDA | 12.8 |
NVML | Same as driver (570.133.20) |
dcgmi | 4.2.3 |
Windows GPU software requirements
Component | Required version |
|---|---|
NVIDIA-SMI | 581.80 |
NVML | 581.80 |
Driver | 581.80 |
CUDA | 13.0 |
Note
Windows does not support dcgmi, so the available metrics are more limited than with the Linux Collector.