nvtop

Added support for the following hardware accelerators/GPUs:

• More Adreno GPUs thanks to the panfrost linux driver
• Mac GPUs (tested on M1 and M2 by the community)
• Huawei Ascend accelerators

Bugfixes and improvements:

• Don't call handlers for GPUs that are marked as not to be monitored
• Fix crash when the configuration file path could not be discovered through user environment variables (XDG_HOME etc)

Bugfixes:

• AMGDPU: GPUs other than id 0 should now be correctly monitored. Thanks to @CabbageSong
• Fixed a bug in GPU selection logic through the F2 menu where the wrong GPUs would still be displayed. Thanks to @cfwen for pointing me in the right direction

Improvemenrs:

• Support for ADRENO GPUs through the msm driver. Thanks @Sonicadvance1 and @robclark
• Better AMDGPU device name query for newer GPUs thanks to @amdsc21 and @paklui.
• "hjkl" directional keys. Thanks @Euro20179

Bugfix:

• Other process than nvtop would end up hidden instead of nvtop itself.
• AMDGPU processes that use graphical and compute resources within the same DRM context would erroneously only report compute usage.

New for this version:

• Intel GPU support
  • Partial for this release due to pending kernel patches. See the Intel section of the README for more information
    • Works (any): dectect GPU with name and clock speed
    • Works on Linux kernel >= 5.19: GPU usage, encode/decode usage, processes running on GPU.
    • Not supported yet: Memory (GPU and processes), Memory clock, fan, power, RX/TX data exchange rate.
  • Uses libsystemd (requires sd-device) or libudev (on systems without systemd or sd-device) to detect the devices and extract some GPU metrics
• Support for detecting integrated graphic cards (Intel and AMD)
• A new "GPU select" window in the setup window (F2) to enable/disable monitoring of select GPUs
  • This setting can be saved to be persistent after a restart
  • This replaces the "-s" and "-i" command line options
• An new option to hide nvtop's own process from the process list (enabled by default)
• Support messages on startup to inform about limitations on old kernels
• Provide an Appstream metainfo file for nvtop

Fix:

• Workaround to detect AMD GPUs which are always reporting 100% usage through the DRM driver and show the sum of the processes usage instead

New this update:

• snap support thanks to @Nightmayr
• Standalone application through an AppImage
  • No compilation needed, download and run (see README)!
  • Built inside an Ubuntu 18.04 docker (should run almost everywhere but please report any issue).
• Logo and .desktop file (required for an AppImage). The logo is not the best since I am not very talented ...
• AMDGPU compute task reporting in the process list.

Fix and improvements this update:

• Fix warnings from misuse of ncurses print function. Thanks @felixonmars
• Go through /proc only once per update, independently of the number of GPUs.
  • Less resources used on systems with multiple AMD GPUs.
• Fix inconsistencies coming from AMDGPU fdinfo interface (e.g., greater than 100% usage in some occasions)

Help nvtop:

I am working on adding Intel support for the next release. If someone has (or plans to buy) a dedicated Intel GPU, please drop a comment on the #106 thread if you are willing to help by sharing some udev and fdinfo dump of your system.

New this update:

• Second are shown on the graphs time axis.
• AMDGPU support for the Linux kernels >=5.19
  • Support for the updated fdinfo interface (seems to be a name harmonization between multiple drivers, i.e., AMD and Intel)
  • More precise GPU usage metrics
    • Per-GPU gfx, encode and decode time usage in nanoseconds allows for a precise percentage usage computation independently of the driver update rate (a GPU time usage column for the process list will follow)

Fix:

• Segfault while retrieving AMDGPU device handles and no device node could be opened.
• Compilation warnings

Thanks to all contributors!

New this update:

• AMDGPU encode and decode metrics full support
• Getting GPU metrics from the process information for older AMDGPU (when the DRM interface does not support it)

Fixes and improvements:

• Fix interface artifacts when the terminal is maximized
• Plot support for up to 64 accelerators
• Fix crash for some terminal sizes when the number of GPU was greater than 1
• Multiple interface positioning algorithm fix
• Memory allocation overflow checks
• Switched to linear array allocation strategy to reduce the memory footprint
• Unit testing using gtest (provides a test target when -DENABLE_TESTING is set or during Debug build)

Minor bugfix release following AMDGPU support.

• Fixed compilation error on systems with older libdrm versions.
• Extract PCIe generation and width from a more reliable source for AMDGPU.

New in this release:

• AMDGPU support

Bugfix:

• Removed a printf to stderr that could break the ncurses layout.

Minor release with some bugfix before updating the Debian package.

New this release:

• Per process GPU, encode and decode usage metrics are available as process columns.
• A setup window (F2) to configure the interface.
  • General options
    • Color support
    • Update interval
  • Devices display options
    • Temperature scale
    • Auto hide of encode/decode
  • Chart options
    • Update direction
    • Which metric is displayed for each device
  • Processes options
    • Toggle ascending/descending
    • Sort by
    • Which field are displayed
• Save user preferences to a configuration file (F12)

Miscellaneous:

• Rewrite of GPU information extraction in preparation for potential multi vendor support (see #106).
  • The build no longer requires the NVML header
  • The build doesn't link to libnvidia-ml anymore, instead loads the symbols at runtime.
• Rewrite and small optimization of the interface drawing code.
  • Chart positioning algorithm
  • Line plot
  • Avoid unnecessary ncurses clean/draw calls

This is a minor release including:

• Bugs fixes where the interface would hang or fail to allocate and crash on systems with many GPUs.
• UI updates: plot direction, PID order and units consistency.
• Showing the power draw even when the upper limit is not provided by the driver.
• Manpage update.
• Installation instruction update.

New in this version:

• Automatic layout of the interface
• Each GPU has its own plot for multi-GPU setup (when there is enough room on screen)
• The host CPU and memory utilization are added to the process list
• Horizontally scrollable process list to see the whole the command line

Fix multi-device support for plot.

New in this version:

• Maximum GPU and memory line plot
  • Can be disabled with option -p or --no-plot.
• Encoding end decoding meters auto-hide if they were not used recently
  • Always visible with option -E -1
• The multi-device information header now automatically takes advantage of the horizontal space available
• Fahrenheit option
  • Activation with option -f or --freedom-unit
• A process name now comes with its arguments
• Process name cache
  • Disable cache with the option -N or --no-cache

• Process list and command options are scrollable

New options

• Select / Ignore GPU by ID
• Monochrome

Bugfix

• Build
• Multi-GPU checks
• Background color = Terminal background color

• CMake build update
  • Search for NVML header and lib is more robust
  • Provide -DNVML_RETRIEVE_HEADER_ONLINE=True option to download nvml.h from NVIDIA repo.
• README update with distribution installation instructions.

• Man page
• Command line options
  • Version
  • Help
  • Refresh rate
• Interactive options
  • Sort by column
  • Kill selected process
• Some refactoring

First release of the nvtop monitoring tool.