1、驱动准备
从NVIDIA网站（https://nvid.nvidia.com/ 需有购买NVIDIA账号登陆访问）下载对应驱动包或者第三方下载
2、NVIDIA常用链接
显卡和驱动版本匹配查询地址：https://docs.nvidia.com/grid/gpus-supported-by-vgpu.html
| GPU1,2, 3 | vGPU 17 | vGPU 164 | vGPU 15 | vGPU 14 | vGPU 13 | vGPU 12 | vGPU 11 | vGPU 10 | vGPU 9 | vGPU 8 | vGPU 7 | vGPU 6 | vGPU 5 | GRID 4 | GRID 3 | GRID 2 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| NVIDIA A800 PCIe 80GB | – | – | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A800 PCIe 80GB liquid cooled | – | – | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A800 HGX 80GB | – | – | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A100 HGX 80GB | – | – | ✔ | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A100 PCIe 80GB | – | – | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A100 PCIe 80GB liquid cooled | – | – | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A100X | – | – | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A100 HGX 40GB | – | – | ✔ | ✔ | ✔ | ✔ | ✔10 | – | – | – | – | – | – | – | – | – |
| NVIDIA A100 PCIe 40GB | – | – | ✔ | ✔ | ✔ | ✔ | ✔10 | – | – | – | – | – | – | – | – | – |
| NVIDIA A40 | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A30 | – | – | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A30X | – | – | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A16 | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A10 | ✔ | ✔ | ✔ | ✔ | ✔ | ✔9 | – | – | – | – | – | – | – | – | – | – |
| NVIDIA A2 | ✔ | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA H800 PCIe 80GB6 | – | – | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA H100 PCIe 80GB | – | – | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA L40S4 | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA L407 | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA L208 | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA L20 liquid cooled9 | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA L46 | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA L28 | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA RTX 5000 Ada4 | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA RTX 5880 Ada | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA RTX 6000 Ada7 | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA RTX A6000 | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – |
| NVIDIA RTX A5500 | ✔ | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – | – | – | – | – |
| NVIDIA RTX A5000 | ✔ | ✔ | ✔ | ✔ | ✔ | ✔9 | – | – | – | – | – | – | – | – | – | – |
| Quadro RTX 8000 | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – | – | – | – |
| Quadro RTX 8000 passive | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – |
| Quadro RTX 6000 | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – | – | – | – |
| Quadro RTX 6000 passive | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – | – | – | – | – | – |
| Tesla V100 | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – | – |
| Tesla T4 | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔11 | – | – | – | – | – |
| Tesla P100 | – | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – |
| Tesla P40 | – | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – |
| Tesla P6 | – | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – |
| Tesla P4 | – | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – | – |
| Tesla M60 | – | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ |
| Tesla M10 | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | – | – |
| Tesla M6 | – | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ | ✔ |
| GRID K2 | – | – | – | – | – | – | – | – | – | – | – | – | – | ✔ | ✔ | ✔ |
| GRID K1 | – | – | – | – | – | – | – | – | – | – | – | – | – | ✔ | ✔ | ✔ |
服务器和显卡兼容性查询地址：https://www.nvidia.cn/data-center/resources/vgpu-certified-servers/
英伟达最全vGPU 链接：http://vgpu.com.cn/可以查看所有相关的文档
NVIDIA Grid驱动版本匹配地址：
https://docs.nvidia.com/grid/get-grid-version.html
3、开启vsphere ssh功能
在vSphere平台下安装vGPU驱动之前需要打开ESXi的SSH服务，方便来传输vib驱动安装包以及后面一些命令行的操作。
ESXi需要加入vCenter，vGPU相关的一些配置都需要在vCenter下来修改。
前置设置，ESXI上的所有虚拟机关闭电源；并将ESXI置于维护模式；在BIOS 中启用 SR-IOV，然后在 ESXi 的 PCI 硬件配置中允许 SR-IOV。为所需的显卡启用 SR-IOV 后，重新启动系统。
4、上传驱动包
使用WinSCP把驱动压缩包中Host主机目录下相关文件传到ESXi的/tmp目录下，根据驱动版本不同，文件夹下文件也不同，对于目前的长周期支持分支vGPU 13.x（ESXi 7.0 推荐使用，不支持ESXi 8.0）,Host主机仅有一个NVD-VGPU开头的驱动包，对于最新的产品分支vGPU 15.x及之后（支持ESXi 8.0） ，还会有一个nvd-gpu-mgmt-daemon开头的文件，Host主机下的NVD-VGPU开头的驱动zip文件不需要再次解压，请把文件直接拷过去。
对于nvd-gpu-mgmt文件，在最新的15.3版本中做了2次打包，所以nvd-gpu-mgmt-daemon_525.125.03-0.0.0000_21816754-package.zip这个文件需要解压,然后把解压得到的nvd-gpu-mgmt-daemon_525.125.03-0.0.0000_21816754.zip文件拷过去，注意我们需要的安装文件是不带package结尾的，之前的版本如15.2不用解压直接拷过去就行，最新16.0版本驱动也直接可用，安装之前先检查一下。
5、主机安装显卡驱动
5.1、安装驱动之前先检查一下GPU是否被正常识别到，SSH到ESXi，使用命令：lspci | grep NVIDIA 来检查。
5.2、服务器需要在BIOS里开启SR-IOV和Monitor/Mwait功能
登录ESXI网页将主机进入维护模式，开始安装驱动
对于vGPU 13.x：esxcli software vib install -d /tmp/NVD-VGPU*.zip
5.3、对于vGPU 15.x及之后需要执行俩次命令：
esxcli software vib install -d /tmp/NVD-VGPU*.zip
esxcli software vib install -d /tmp/nvd-gpu-mgmt-daemon*.zip
5.4、等待几分钟，提示安装完成，然后将显卡切换为直通模式，重启一下主机验证驱动是否正常。注意：这一步必须要设置， 否则无法进行下一步的操作，笔者曾经在这里付出了真金白银的惨痛代价。
重启完成后输入nvidia-smi
5.5、特别注意2点：1、ECC模式是否显示off 2、默认情况下vSphere会使用vsga模式，而不是vGPU模式，所以能看到最下面xorg条目，这个我们后面改。
如果看到nvidia-smi信息里面ECC区域不是off状态，而是0，这说明你现在的ECC是启用的状态，并不是所有vgpu都可以使用ECC，如果你不能确认你的环境是否需要启用ECC，那建议先关闭ECC
5.6、使用命令来关闭ECC，注意，启用或者关闭ECC都需要重启主机。
执行命令关闭ECC，关闭后需要重启主机 nvidia-smi –e 0
5.7、驱动安装完成后将主机退出维护模式。
6、vCenter配置显卡共享模式
6.1、登陆vCenter，定位到主机-配置-图形，编辑主机图形设置，可以看到默认是共享模式，也就是vsga，我们需要切换到“直接共享”才能使用vGPU。
6.2、默认为共享，修改成直接共享，并重启xorg服务。
更改成功以后就没有xorg条目了，可以开始使用vGPU了！
7、分配虚拟机显卡资源
7.1、编辑虚拟机，新增PCI设备，可根据授权来分配显卡Q B A
显存分配必须一致，比如一个主机下所有虚拟机显存都是2Q 或者4Q，
不能虚拟机1是2Q，虚拟机2是4Q，显卡资源调度会有问题。
7.2、GPU显卡选择类型如下：
每一种物理GPU支持的vGPU类型都不一样。VGPU显卡性能主要以分配的显存及显卡类型进行划
分，即虚拟机所分配的虚拟显卡型号。VGPU有几种后缀：
1.Q系列虚拟GPU类型针对设计师和高级用户。（vDWS，虚拟工作站）
2.B系列虚拟GPU类型针对高级用户。（vPC，虚拟pc）
3.A系列虚拟GPU类型针对虚拟应用程序用户。（VAPP虚拟应用，有点类似于远程应用）
8、系统安装驱动包
8.1、Win10系统挂载好显存后，开机安装显卡驱动包
安装虚拟机驱动，注意一定要使用安装包内对应版本的驱动（使用不同版本可能会有问题），截图示例win10系统驱动包
安装驱动后，重启虚拟机，在控制台界面，看到“黑屏”，这是正常的。使用远程桌面RDP协议登录
