Syllo/nvtop

NVTOP stands for Neat Videocard TOP, a (h)top like task monitor for GPUs and accelerators. It can handle multiple GPUs and print information about them in a htop-familiar way.

Currently supported vendors are AMD (Linux amdgpu driver), Apple (limited M1 & M2 support), Huawei (Ascend), Intel (Linux i915/Xe drivers), NVIDIA (Linux proprietary divers), Qualcomm Adreno (Linux MSM driver), Broadcom VideoCore (Linux v3d driver), Rockchip, MetaX (MXSML driver), Enflame (Linux EFML driver), Tenstorrent (Linux tt-kmd driver). Rockchip, MetaX (MXSML driver), Enflame (Linux EFML driver), Iluvatar CoreX (ixML / libixml).

Because a picture is worth a thousand words:

• NVTOP Options and Interactive Commands
  • Interactive Setup Window
  • Saving Preferences
  • NVTOP Manual and Command line Options
• GPU Support
  • AMD
  • Intel
  • NVIDIA
  • Adreno
  • Apple
  • Ascend (only tested on 910B)
  • Iluvatar CoreX
  • VideoCore
  • Rockchip
  • MetaX
  • Enflame
  • Tenstorrent
• Build
• Distribution Specific Installation Process
  • Ubuntu / Debian
    • Ubuntu Impish (21.10) / Debian buster (stable) and more recent (stable)
  • Fedora / Red Hat / CentOS
  • OpenSUSE
  • Arch Linux
  • Gentoo
  • AppImage
  • Snap
  • Conda-forge
  • Docker
  • WSL2
• NVTOP Build
• Troubleshoot
• License

NVTOP has a builtin setup utility that provides a way to specialize the interface to your needs. Simply press F2 and select the options that are the best for you.

In the Chart section you can choose which metrics are plotted, including GPU and memory utilization, temperature, power, clocks, and the PCIe RX / TX load (the receive and transmit throughput as a percentage of the maximum link bandwidth).

You can save the preferences set in the setup window by pressing F12. The preferences will be loaded the next time you run nvtop.

NVTOP comes with a manpage!

man nvtop

For quick command line arguments help

nvtop -h
nvtop --help

NVTOP supports AMD GPUs using the amdgpu driver and the legacy radeon driver (legacy GPUs, limited support) through the exposed DRM and sysfs interface.

The radeon provides limited metrics compared to amdgpu.

AMD introduced the fdinfo interface in kernel 5.14 (browse kernel source). Hence, you will need a kernel with a version greater or equal to 5.14 to see the processes using AMD GPUs.

Support for recent GPUs are regularly mainlined into the linux kernel, so please use a recent-enough kernel for your GPU.

NVTOP supports Intel GPUs using the i915 or xe linux driver.

Intel introduced the fdinfo interface in kernel 5.19 (browse kernel source). Hence, you will need a kernel with a version greater or equal to 5.19 to see the processes using Intel GPUs.

Intel requires CAP_PERFMON or CAP_SYS_ADMIN capabilities to access the total memory usage, and an accurate GPU frequency. you can run sudo setcap cap_perfmon=ep $(which nvtop) to grant the necessary permissions or run nvtop as root.

The NVML library does not support some of the queries for GPUs coming before the Kepler microarchitecture. Anything starting at GeForce 600, GeForce 800M and successor should work fine. For more information about supported GPUs please take a look at the NVML documentation.

The extra GPU info bar (nvtop -i) shows the dynamic parameters NVML does not report, such as the XBAR, SYS and NVD clocks, and nvtop -X also shows the static specs (shader cores, L2 cache, execution engines) and the secondary HUB, HOST, DISP, MSD and UTILS clocks. The clock domains are read through the NvAPI library (libnvidia-api.so.1) that recent drivers install alongside NVML; without it the rest of the interface is unaffected and the bar takes no room on a GPU that reports nothing.

NVTOP supports Adreno GPUs using the msm linux driver.

msm introduced the fdinfo interface in kernel 6.0 (browse kernel source). Hence, you will need a kernel with a version greater or equal to 6.0 to see the processes using Adreno GPUs.

NVTOP includes some initial support for Apple using Metal. This is only supported when building for Apple, and when building for Apple only this vendor is supported.

APPLE SUPPORT STATUS

• Apple support is still being worked on. Some bugs and limitations may apply.

NVTOP supports Ascend through the DCMI API (tested on 910B3 hardware). It uses the logical-device API when available and falls back to the legacy card/device API. Read-only telemetry is reported when supported by the installed driver; unavailable values remain unset. Enable it with -DASCEND_SUPPORT=ON, and set -DASCEND_ROOT=/path/to/Ascend when CANN is installed outside the standard locations.

NVTOP supports Iluvatar CoreX GPUs through the ixML library.

The backend dynamically loads libixml.so from /usr/local/corex/lib, /usr/local/corex/lib64, or the default dynamic loader search path. The ixML runtime exposes an NVML-compatible API surface used by NVTOP to query device, power, PCIe, clock, temperature, memory, and process information.

NVTOP supports VideoCore (testing on raspberrypi 4B).

Supports GPU frequency, temperature, utilization, per-process utilization, GPU memory usage, and H264 decoding utilization.

On non-raspberry pi os, you need to use the linux-rpi 6.12.y kernel and above, and ensure the presence of the /dev/vcio device.

NVTOP supports Rockchip (testing on orangepi 5 plus).

Supports NPU frequency, temperature, utilization.

NVTOP supports MetaX (testing on MXC500) by MXSML LIBRARY.

For more information about GPUs please take a look at the METAX documentation

NVTOP supports Enflame GCUs (testing on Enflame S60, Enflame L300 and Enflame L600) by EFML LIBRARY

GCU, which refers to General Compute Unit, is a type of accelerator card that is used to perform general-purpose computing tasks just like GPGPU.

NVTOP supports Tenstorrent AI accelerators (Blackhole, Wormhole, Grayskull) through the tt-kmd kernel driver.

Supports temperature, power draw, AI clock, fan RPM, PCIe link info, and process listing. No external libraries required -- all data is read from sysfs, hwmon, and procfs.

Several libraries are required in order for NVTOP to display GPU info:

• The ncurses library driving the user interface.
  • This makes the screen look beautiful.
• For NVIDIA: the NVIDIA Management Library (NVML) which comes with the GPU driver.
  • This queries the GPU for info.
  • The NvAPI library (libnvidia-api.so.1), also from the driver, is used when present for the clock domains NVML does not report. It is optional.
• For AMD: the libdrm library used to query AMD GPUs through the kernel driver.
• For METAX: the MetaX System Management Library (MXSML) which comes with the GPU driver.
  • This queries the GPU for info.
• For Enflame: the Enflame Management Library (EFML) which comes with the GCU driver.
• For Iluvatar CoreX: the ixML runtime library (libixml.so) which comes with the driver.
  • This backend loads the library dynamically at runtime.

If your distribution provides the snap utility, follow the snap installation process to obtain an up-to-date version of nvtop.

A standalone application is available as AppImage.

sudo apt install nvtop

A PPA supporting Ubuntu 20.04 and newer is provided by Quentin Lienhardt that offers an up-to-date version of nvtop, enabled for NVIDIA, AMD and Intel.

sudo add-apt-repository ppa:quentiumyt/nvtop
sudo apt install nvtop

• AMD and Intel Dependencies

  sudo apt install libdrm-dev libsystemd-dev
  # Ubuntu 18.04
  sudo apt install libudev-dev

• NVIDIA Dependency

  • NVIDIA drivers (see Ubuntu Wiki or Ubuntu PPA or Debian Wiki)

• NVTOP Dependencies

  • CMake, ncurses and Git

  sudo apt install cmake libncurses5-dev libncursesw5-dev git

• NVTOP

  • Follow the NVTOP Build

A standalone application is available as AppImage.

• sudo dnf install nvtop

• sudo dnf install -y https://dl.fedoraproject.org/pub/epel/epel-release-latest-$(rpm -E %{rhel}).noarch.rpm
  sudo dnf install nvtop

• sudo dnf install -y epel-release
  sudo dnf install nvtop

• AMD and Intel Dependencies

  sudo dnf install libdrm-devel systemd-devel

• NVIDIA Dependency

  • NVIDIA drivers, CUDA required for nvml libraries (see RPM Fusion)

• NVTOP Dependencies

  • CMake, ncurses, C++ and Git

  sudo dnf install cmake ncurses-devel git gcc-c++

• NVTOP

  • Follow the NVTOP Build

A standalone application is available as an AppImage.

Build process for OpenSUSE:

• AMD Dependency

  sudo zypper install libdrm-devel

• NVIDIA Dependency

  • NVIDIA drivers (see SUSE Support Database)

• NVTOP Dependencies

  • CMake, ncurses and Git

    sudo zypper install cmake ncurses-devel git

• NVTOP

  • Follow the NVTOP Build

• sudo pacman -S nvtop

• sudo emerge -av nvtop

An AppImage is a standalone application. Just download the AppImage, make it executable and run it!

• Go to the release page and download nvtop-x86_64.AppImage

• # Go to the download location ** The path may differ on your system **
  cd $HOME/Downloads
  # Make the AppImage executable
  chmod u+x nvtop-x86_64.AppImage
  # Enjoy nvtop
  ./nvtop-x86_64.AppImage

If you are curious how that works, please visit the AppImage website.

• snap install nvtop
  # Add the capability to kill processes inside nvtop
  snap connect nvtop:process-control
  # Add the capability to inspect GPU information (fan, PCIe, power, etc)
  snap connect nvtop:hardware-observe
  # AMDGPU process list support (read /proc/<pid>)
  snap connect nvtop:system-observe
  # Temporary workaround to get per-process GPU usage (read /proc/<pid>/fdinfo)
  snap connect nvtop:kubernetes-support

Notice: The connect commands allow

A conda-forge feedstock for nvtop is available.

conda install --channel conda-forge nvtop

pixi global install nvtop

• NVIDIA drivers (same as above)

• nvidia-docker (See Container Toolkit Installation Guide)

• git clone https://github.com/Syllo/nvtop.git && cd nvtop
  sudo docker build --tag nvtop .
  sudo docker run -it --rm --runtime=nvidia --gpus=all --pid=host nvtop

Nvtop support in WSL2 mostly boils down to GPU/Accelerator vendor support of WSL2. Nvtop should work out of the box for the following configurations:

• NVIDIA: Please refer to NVIDIA's user guide to get started with CUDA on WSL2 to avoid overiding the driver exposed by WSL2 (from Windows) with the NVIDIA Linux drivers.

git clone https://github.com/Syllo/nvtop.git
mkdir -p nvtop/build && cd nvtop/build
cmake .. -DNVIDIA_SUPPORT=ON -DAMDGPU_SUPPORT=ON -DINTEL_SUPPORT=ON
make

# Install globally on the system
sudo make install

# Alternatively, install without privileges at a location of your choosing
# cmake .. -DNVIDIA_SUPPORT=ON -DAMDGPU_SUPPORT=ON -DINTEL_SUPPORT=ON -DCMAKE_INSTALL_PREFIX=/path/to/your/dir
# make
# make install

If you use conda as environment manager and encounter an error while building NVTOP, try conda deactivate before invoking cmake.

The build system supports multiple build types (e.g. -DCMAKE_BUILD_TYPE=RelWithDebInfo):

• Release: Binary without debug info
• RelWithDebInfo: Binary with debug info
• Debug: Compile with warning flags and address/undefined sanitizers enabled (for development purposes)

• The plot looks bad:
  • Verify that you installed the wide character version of the ncurses library (libncursesw5-dev for Debian / Ubuntu), clean the build directory and restart the build process.
• If nvtop exits with ncurses: cannot initialize terminal type ($TERM="unknown"), ensure that $TERM is set to a valid terminal type such as xterm-256color.
• Putty: Tell putty not to lie about its capabilities ($TERM) by setting the field Terminal-type string to putty in the menu Connection > Data > Terminal Details.
• NO GPU to monitor. for NVIDIA GPUs:
  • nvtop loads a shared library named libnvml.so (shipped with the NVIDIA drivers) to querry device information. If the library is not present, nvtop will not be able to monitor your NVIDIA device.
  • On WSL2, the installation instructions are slightly different since the driver is being exposed by Windows to the virtual machine (see instruction link in the WSL2 section). If you install the NVIDIA linux drivers inside WSL2, you may encounter a version mismatch error or nvtop may silently not work at all.

NVTOP is licensed under the GPLv3 license or any later version. You will find a copy of the license inside the COPYING file of the repository or at the GNU website <www.gnu.org/licenses/>.