Introduction to GPU monitoring

Introduction

Supported GPU vendors

Supported Redfish collectors for out-band GPU monitoring

Supported OSes for in-band GPU monitoring

How to specify out-band vs in-band GPU monitoring

Terminology

Introduction

Entuity supports out-band and in-band monitoring of the Graphics Processing Units (GPUs) in your managed servers. In previous versions of Entuity, certain server hardware collectors could perform out-band monitoring of GPUs, but the GPUs were classified and displayed in the UI as CPUs. Now, the GPUs are classified and displayed as GPUs. With out-band GPU monitoring, Entuity discovers the GPU identity and inventory (name, vendor, model, serial, firmware) and the GPU component status (temperature, power) for each out-band GPU.

Note, historical data recorded against the old CPU objects is not migrated.

To discover more detailed information about GPUs, including incidents, events, and thresholds, Entuity added support for in-band GPU monitoring. With in-band GPU monitoring, Entuity uses the command-line tools and utilities of each supported GPU vendor run over SSH on the managed server. With in-band GPU monitoring, Entuity discovers the GPU model, its partitions, utilization, instances, interconnects, fabric switches, memory, and clocks for each in-band GPU.

Supported GPU vendors

Entuity supports monitoring GPUs from the following vendors:

• AMD
• Intel
• Nvidia

Supported server hardware collectors for out-band GPU monitoring

The following server hardware collectors can discover GPU information for out-band GPU monitoring:

• Dell iDRAC
• Gigabyte
• HP iLO
• IBM/Lenovo
• Intel
• SuperMicro

Supported OSes for in-band GPU monitoring

The GPU vendor CLI tools and utilities that Entuity uses for in-band GPU monitoring are supported on the following operating systems: 

• AMD GPUs - Linux 
• Intel GPUs - Linux and Windows (for Intel Arc, Flex, and Max GPUs)
• Nvidia GPUs - Linux and Windows

How to specify out-band vs in-band GPU monitoring

A server can be monitored out-band through its management controller (BMC) and in-band through its operating system. When adding a server to your network in Entuity, you can choose to either perform out-band or in-band monitoring of GPUs, not both. 

To specify out-band GPU monitoring, add a server to your network as you normally would, and select a server hardware collector that can discover GPU information (see above). 

To specify in-band GPU monitoring, add a server to your network as you normally would, and select an OS monitoring Connection Method and specify a GPU Connection Method. 

For more information on how to configure GPU monitoring, including prerequisites for doing so, see this article.

For more information on adding a server device to your network, see this article.

Terminology:

The following terminology is used for GPUs in Entuity:

• GPU - A physical graphics processor in a managed (or, monitored) server. Metrics for the whole GPU are reported on this object.
• GPU instance - a partition of one GPU with its own memory allocation. A GPU can contain several GPU instances. Entuity shows each one under its parent GPU.
• GPU compute instance - a further division of a GPU instance. It isolates compute resources for a workload and uses the memory of its parent GPU instance. Entuity shows these under that GPU instance. They are displayed only on GPUs that support this second level of partitioning.
• GPU link - a high-speed hardware communication pathway that lets GPUs share data and memory directly with each other or with the CPU.

• Fabric switch - a specialized, high-speed networking or silicon crossbar device that connects multiple GPUs together.