{{ $t('productDocDetail.updateTime') }}: 2026-01-05

Function Description

This function supports the passthrough mode or automatic or manual division for the GPU cores of a server. By default, automatic division is used. GPU resources can be divided into Q-series and C-series vGPUs.

The Q series is applicable to virtual workstations that require the performance and functions of the Quadro technology in creative and professional scenarios.

The C series is applicable to compute-intensive servers for AI, deep learning, high performance computing (HPC), and more scenarios.

Precautions

1. For graphics cards supporting the passthrough mode, the passthrough mode is used by default. For graphics cards supporting vGPUs, the automatic mode is used by default, and you can select different division modes as needed.

2. Generally, two graphics cards can be passed through to a VM. For T4 or 2080Ti, a maximum of eight graphics cards can be passed through. You cannot allocate multiple vGPUs to a VM.
3. NVIDIA vGPU Software 10.2 and 13.2 are supported. For more information about the compatible versions, visit the NVIDIA official website. If the latest HCI is installed or HCI is upgraded to 6.8.0, you need to import the GRID driver again and restart the host for the driver to take effect.
4. Hot migration is not supported for VMs on which GPU resource allocation has been performed in the HCI console.
5. In vGPU scenarios, the video memory resources of graphics cards are allocated exclusively, and the computing resources of graphics cards are scheduled by a scheduler. Currently, the following three schedulers are supported: Best Effort, Equal Share, and Fixed Share.

Best Effort: If multiple VMs run on a graphics card, computing resources are preempted based on workloads.

Equal Share: If multiple VMs run on a graphics card, computing resources are allocated evenly.

Fixed Share: If multiple VMs run on a graphics card, computing resources are allocated based on the percentage of video memory allocated to a VM.

6. A server cannot use graphics cards of different types, and different nodes in a cluster can use graphics cards of different types.
7. If graphics cards A and B are inserted into two neighboring PCI slots, and the cores of graphics card B are damaged, the cores of graphics card B will be identified as belonging to graphics card A. In this case, replace the damaged graphics card promptly.

Prerequisites

The VM using the graphics card has been shut down before the GPU resource division mode is changed.

Procedure

1. Go to Nodes > Graphics Cards to view the realtime allocation of GPU cores and driver version information.

2. Click Edit in the Operation column of a graphics card.

①Edit the name of the graphics card.

②Set the mode to Passthrough or Division. If you select Division, and the graphics card has multiple cores, you can divide the video memory resource based on the number of cores.

③Specify whether to enable ECC memory. If it is enabled, GPU stability can be improved.

④Specify the division mode. GPU resources can be can be divided into Q-series and C-series vGPUs.

⑤Select the GPU computing scheduler.

3. Click OK. Then, you can view the graphics card configuration information on the Graphics Cards tab.

4. Click Change Server Driver to import the GRID driver.

If HCI is upgraded from 6.2.30 to 6.8.0, you need to import the GRID driver again (10.2 and 13.2 versions are supported) and restart the host for the configuration to take effect. Therefore, you are advised to perform the upgrade first, import the driver, and restart the host.

5. Wait for the driver upload to complete.
6. Select the node to install the driver and wait for driver distribution to complete.
7. Click Enter Maintenance Mode and restart the node for the driver configuration to take effect.