Description

• This feature helps customers migrate their businesses from the traditional physical architecture to HCI. As certain businesses, such as Geographic Information System (GIS), require GPU support, HCI needs to support GPU configuration.

• If a large number of applications require GPU support, this feature supports flexible use, on-demand allocation, and sharing of GPU resources to avoid resource waste or inflexible allocation.

• HCI can visualize the allocation, use, and adjustment of GPU resources for easier management.

Precautions

1. Using vGPU requires licensing from NVIDIA.
2. To use vGPU on your VM, you must install the NVIDIA GPU driver for HCI 6.9.0. You can obtain the driver from Self Services > Download > aCloud(HCI) at www.community.sangfor.com.
3. Two vGPU licenses are available for HCI: vWS and vCS licenses. The vWS license is used for professional graphics and compute scenarios, and the vCS license is used for compute-intensive scenarios, such as AI, deep learning, and scientific research.
4. The vCS license is unavailable to Windows VMs. The GPU partitioning granularity ranges from 4–32 GB.
5. Not all GPUs support vGPU. If your GPU does not support vGPU, you can use the passthrough mode to allocate GPU resources.
6. Generally, two GPUs can be passed through to a VM. For T4 or 2080Ti, a maximum of eight GPUs can be passed through.
7. In vGPU mode, you can check GPU compatibility in the HCI compatibility chart.
8. In passthrough mode, you need to download the NVIDIA reference driver of the matching version at https://www.nvidia.cn/Download/index.aspx?lang=cn.
9. HCI 6.9.0 and later versions support China's homegrown GPUs, but only in the passthrough mode.

• x86-compatible models: Cambricon MLU370-S4, Cambricon MLU270-S4, and Moore Threads MTT S2000

• Hygon GPUs are not supported.

• GPU usage and video memory usage cannot be collected if Moore Threads MTT S2000 is used on Windows.

• GPU usage and video memory usage cannot be collected if Huawei Atlas 300V Pro is used on openEuler.

Prerequisites

1. A cluster has been created, which consists of two GPU servers or one GPU server and one ordinary server. That is, at least one GPU has been installed. You can use your original GPU or purchase one from Sangfor.
2. 2D and 3D servers are supported. However, GPUs can be installed only on GPU servers.

Steps

Deploy the HCI environment. In terms of hardware, install the GPU on the server. Make sure that the power supply cable is well connected for the GPU to work (a dedicated power supply cable is required because the GPU consumes high power). As for software, when HCI is installed on the server, or HCI detects that a GPU has been installed in the cluster, the input-output memory management unit (IOMMU) option is displayed automatically. In this case, enable this option and restart the host for the setting to take effect.

=