Ubuntu 安装 NVIDIA 驱动和 CUDA (NVCC)

列出 GPU 信息

lspci | egrep -i "vga|3d|display"

或者：

# sudo apt-get install lshw
sudo lshw -C display

【重要】确保 Secure Boot 关闭

sudo mokutil --sb-state

输出应为 SecureBoot disabled。如果是 SecureBoot enabled，需要进入 BIOS 关闭。

运行：

sudo systemctl reboot --firmware-setup

会自动进入启动菜单界面，选择 EFI Setup （通常在最下面） 进入 BIOS 设置界面：

• 依次进入 Device Manager → Secure Boot Configuration → Attempt Secure Boot
• 将其设为 Disabled，也即从 [x] 改为 [ ]
• F10 保存，Esc 退出，回车确认修改设置，这样就会重启并进入系统

启动后，再在终端运行 sudo mokutil --sb-state，应当显示 SecureBoot disabled。

安装 NVIDIA 驱动

See: NVIDIA drivers installation | Ubuntu

• https://ubuntu.com/server/docs/nvidia-drivers-installation

查看驱动信息

cat /proc/driver/nvidia/version

输出形如：

NVRM version: NVIDIA UNIX x86_64 Kernel Module  575.64.03  Wed Jun 25 18:40:52 UTC 2025
GCC version:  gcc version 12.3.0 (Ubuntu 12.3.0-1ubuntu1~22.04.2)

如果没安装，会提示 No such file or directory。

或者运行：

# sudo apt install nvidia-utils-575
nvidia-smi

输出形如：

Tue Sep  2 16:17:39 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 575.64.03              Driver Version: 575.64.03      CUDA Version: 12.9     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   1  NVIDIA GeForce RTX 4090        Off |   00000000:41:00.0 Off |                  Off |
| 35%   35C    P8             27W /  450W |      43MiB /  49140MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

如果安装不正确，可能会提示：

NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver.
Make sure that the latest NVIDIA driver is installed and running.

这个一般是驱动版本不对造成的，所以可以用下面的方法安装设备推荐的驱动。

还有一种报错是：

Failed to initialize NVML: Driver/library version mismatch...

这个可以试试重启解决。

列出推荐驱动

sudo ubuntu-drivers devices | grep recommended

输出形如：

# driver   : nvidia-driver-535 - distro non-free recommended
# driver   : nvidia-driver-575 - distro non-free recommended      # latest on 2025-09-02
# driver   : nvidia-driver-580 - third-party non-free recommended # latest on 2025-11-24

See: NVIDIA-SMI has failed because it couldn't communicate with the NVIDIA driver - Stack Overflow

• https://stackoverflow.com/questions/42984743/nvidia-smi-has-failed-because-it-couldnt-communicate-with-the-nvidia-driver

列出全部可用驱动

# desktop
sudo ubuntu-drivers list

# server
# sudo ubuntu-drivers list --gpgpu

例如，Ubuntu 22.04 的可用驱动输出：

nvidia-driver-535-open, (kernel modules provided by linux-modules-nvidia-535-open-generic-hwe-22.04)                                                                              [30/49]
nvidia-driver-535, (kernel modules provided by linux-modules-nvidia-535-generic-hwe-22.04)
nvidia-driver-580, (kernel modules provided by nvidia-dkms-580)
nvidia-driver-580-open, (kernel modules provided by nvidia-dkms-580-open)
nvidia-driver-575-server, (kernel modules provided by linux-modules-nvidia-575-server-generic-hwe-22.04)
nvidia-driver-575, (kernel modules provided by linux-modules-nvidia-575-generic-hwe-22.04)
nvidia-driver-535-server, (kernel modules provided by linux-modules-nvidia-535-server-generic-hwe-22.04)
nvidia-driver-575-open, (kernel modules provided by linux-modules-nvidia-575-open-generic-hwe-22.04)
...

安装驱动

更新软件包列表：

sudo apt-get update

• 清华源可能会有 403 Forbidden 错误
• 建议使用中科大源，详见：Ubuntu 换国内源

直接安装推荐的驱动：

# sudo apt install nvidia-driver-535
# sudo apt install nvidia-driver-575   # latest on 2025-09-02
sudo apt install nvidia-driver-580   # latest on 2025-11-24

重装：

sudo apt install --reinstall nvidia-driver-580

用这种方式安装可以自动安装依赖的内核模块。

也可用 ubuntu-drivers 安装，不过可能会有问题，目前不推荐：

# desktop: recommended
# sudo ubuntu-drivers install
# desktop: specific version
# sudo ubuntu-drivers install nvidia:535

# server: recommended
# sudo ubuntu-drivers install --gpgpu
# server: specific version
# sudo ubuntu-drivers install --gpgpu nvidia:535-server
# server: install additional components
# sudo apt install nvidia-utils-535-server

重启生效

# sudo reboot
sudo shutdown -r now

查看驱动安装是否成功

查看内核模块：

dkms status

输出形如：

nvidia/580.105.08, 6.8.0-40-generic, x86_64: installed
nvidia/580.105.08, 6.8.0-87-generic, x86_64: installed

查看模块加载：

lsmod | grep -i nvidia

输出形如：

nvidia_uvm           2093056  0
nvidia_drm            139264  6
nvidia_modeset       1638400  5 nvidia_drm
nvidia              104091648  71 nvidia_uvm,nvidia_modeset
video                  77824  1 nvidia_modeset

查看内核驱动占用：

lspci -nnk

输出应当包含：

kernel driver in use: ....

卸载驱动

如果版本不对，或者需要重新安装，可以先卸载已安装的驱动：

sudo apt purge "nvidia-*" && sudo apt autoremove && sudo apt autoclean

安装 CUDA (NVCC)

【首选】通过 NVIDIA 源安装指定版本 nvcc

CUDA Toolkit 12.9 Downloads | NVIDIA Developer

• https://developer.nvidia.com/cuda-12-9-0-download-archive?target_os=Linux&target_arch=x86_64&Distribution=Ubuntu&target_version=22.04&target_type=deb_network
• https://developer.nvidia.com/cuda-13-0-1-download-archive?target_os=Linux&target_arch=x86_64&Distribution=Ubuntu&target_version=22.04&target_type=deb_local

cd ~/downloads
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
# sudo apt-get -y install cuda-toolkit-12-9
sudo apt-get -y install cuda-toolkit-13-0

在 .bashrc 或 .zshrc 中添加：

export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH

运行 bash 或 zsh 使配置生效。

【次选】通过 apt 安装 nvcc

sudo apt install nvidia-cuda-toolkit

See: How to install CUDA & cuDNN on Ubuntu 22.04

• https://gist.github.com/denguir/b21aa66ae7fb1089655dd9de8351a202

【次选】通过 conda 安装 nvcc

如果想在没有 root 权限的情况下安装 nvcc（比如在 conda 环境），可以运行：

conda install nvidia/label/cuda-11.7.0::cuda-toolkit

See: https://anaconda.org/nvidia/cuda-toolkit

查看 CUDA 版本

which nvcc

• 如果是通过 NVIDIA 源安装，并且设置了环境变量，则输出为：/usr/local/cuda/bin/nvcc
• 如果是通过 apt 安装，则输出为：/usr/bin/nvcc

nvcc --version

See: Different CUDA versions shown by nvcc and NVIDIA-smi - Stack Overflow

• https://stackoverflow.com/questions/53422407/different-cuda-versions-shown-by-nvcc-and-nvidia-smi

用 gpustat 查看 GPU 实时状态

安装：

pip install gpustat

运行：

gpustat -cpu -i

常见问题

NVIDIA-SMI Shows ERR! on both Fan and Power Usage

IOT instruction (core dumped)

原因是温度过高。可以尝试把显卡放到更凉快的位置，或者设置功耗限制和风扇速度。

This issue is due to the higher temperature.

First, you should reseat the question card to the coolest location of your workstation.

Second, set the power limitation [1] and fan speed [2] to ensure the peak temperature does not exceed 75C.

[1] Change them to 150W-to-200W sudo nvidia-smi -pm 1 sudo nvidia-smi -pl 150(200)

[2] https://github.com/boris-dimitrov/set_gpu_fans_public 402

Using these methods, I have restored two 1080ti cards which have the same issues.

NVIDIA-SMI Shows ERR! on both Fan and Power Usage - Graphics / Linux / Linux - NVIDIA Developer Forums

• https://forums.developer.nvidia.com/t/nvidia-smi-shows-err-on-both-fan-and-power-usage/68293/14

不过也有贴子提到，更新主板 BIOS 可以解决。（待验证）

To follow up, I solved this by upgrading my motherboard (ASUS PRIME X470-PRO) BIOS to the latest version.

RTX 4090 Fan state says "ERR!", performance is throttled - Graphics / Linux / Linux - NVIDIA Developer Forums

• https://forums.developer.nvidia.com/t/rtx-4090-fan-state-says-err-performance-is-throttled/264705/3