AMDGPU_TOP

amdgpu_top is tool that display AMD GPU utilization, like umr or clbr/radeontop or intel_gpu_top.
The tool displays information gathered from performance counters (GRBM, GRBM2), sensors, fdinfo, and AMDGPU driver.

Simple TUI (like nvidia-smi, rocm-smi)	TUI	GUI

Quick links

• Usage
  • Options
  • Commands for TUI
  • Example of using JSON mode
  • fdinfo description
• Installation
  • Packages
  • Build from source
    • Distribution specific instructions
• Binary Size
• References
• Translations
• Alternatives

Dependent dynamic libraries

• libdrm
• libdrm_amdgpu

Usage

cargo run -- [options ..]
# or
amdgpu_top [options ..]

# Run TUI mode
amdgpu_top

# Run GUI mode
amdgpu_top --gui

# Run SMI mode
amdgpu_top --smi

# Dump AMDGPU info
amdgpu_top -d

# Dump AMDGPU info and gpu_metrics
amdgpu_top -d -gm

# Dump AMDGPU info and pp_table
amdgpu_top -d --pp-table

# Dump AMDGPU info in JSON format
amdgpu_top -d --json

# Dump DRM info
amdgpu_top --drm-info

# Decode gpu_metrics file
amdgpu_top --decode-gm <path>

# Decode gpu_metrics file and output in JSON format
amdgpu_top --json --decode-gm <path>

Options

FLAGS:
   -d, --dump
       Dump AMDGPU info. (Specifications, VRAM, PCI, ResizableBAR, VBIOS, Video caps)
       This option can be combined with the "-J" option.
   --list
       Display a list of AMDGPU devices.
   -J, --json
       Output JSON formatted data.
       This option can be combined with the "-d" option.
   --gui
       Launch GUI mode.
   --smi
       Launch Simple TUI mode. (like nvidia-smi, rocm-smi)
   -p, --process
       Dump All GPU processes and memory usage per process.
   --apu, --select-apu
       Select APU instance.
   --single, --single-gpu
       Display only the selected APU/GPU
   --no-pc
       The application does not read the performance counter (GRBM, GRBM2)
       if this flag is set.
       Reading the performance counter may deactivate the power saving feature of APU/GPU.
   -gm, --gpu_metrics, --gpu-metrics
       Dump gpu_metrics for all AMD GPUs.
       https://www.kernel.org/doc/html/latest/gpu/amdgpu/thermal.html#gpu-metrics
   --pp_table, --pp-table
       Dump pp_table from sysfs and VBIOS for all AMD GPUs.
       (only support Navi1x and Navi2x, Navi3x)
   --drm_info, --drm-info
       Dump DRM info.
       Inspired by https://gitlab.freedesktop.org/emersion/drm_info
    --xdna
       Dump XDNA NPU info.
   --dark, --dark-mode
       Set to the dark mode. (TUI/GUI)
   --light, --light-mode
       Set to the light mode. (TUI/GUI)
   --gl, --opengl
       Use OpenGL API to the GUI backend.
   --vk, --vulkan
       Use Vulkan API to the GUI backend,
       and use APU/iGPU for GUI rendering if it is available.
   -V, --version
       Print version information.
   -h, --help
       Print help information.

OPTIONS:
   -i <usize>
       Select GPU instance.
   --pci <String>
       Specifying PCI path. (domain:bus:dev.func)
   -s <u64>, -s <u64>ms
       Refresh period (interval) in milliseconds for JSON mode. (default: 1000ms)
   -n <u32>
       Specifies the maximum number of iteration for JSON mode.
       If 0 is specified, it will be an infinite loop. (default: 0)
   -u <u64>, --update-process-index <u64>
       Update interval in seconds of the process index for fdinfo. (default: 5s)
   --json_fifo, --json-fifo <String>
       Output JSON formatted data to FIFO (named pipe) for other application and scripts.
   --decode-gm <Path>, --decode-gpu-metrics <Path>
       Decode the specified gpu_metrics file.

Commands for TUI

key
g	toggle GRBM
r	toggle GRBM2
v	toggle VRAM/GTT Usage
f	toggle fdinfo
n	toggle Sensors
m	toggle GPU Metrics
h	change update interval (high = 100ms, low = 1000ms)
q	Quit
P	sort fdinfo by pid
M	sort fdinfo by VRAM usage
G	sort fdinfo by GFX usage
M	sort fdinfo by MediaEngine usage
R	reverse sort

Example of using JSON mode

$ amdgpu_top --json | jq -c -r '(.devices[] |
  (.Info | .DeviceName + " (" + .PCI + "): ") +
  ([.gpu_activity | to_entries[] | .key + ": " + (.value.value|tostring) + .value.unit] |
  join(", ")))'

Output

AMD Radeon RX 6600 (0000:03:00.0): GFX: 13%, MediaEngine: 0%, Memory: 4%
AMD Radeon Graphics (0000:08:00.0): GFX: 0%, MediaEngine: 0%, Memory: null%
AMD Radeon RX 6600 (0000:03:00.0): GFX: 15%, MediaEngine: 0%, Memory: 5%
AMD Radeon Graphics (0000:08:00.0): GFX: 0%, MediaEngine: 0%, Memory: null%
AMD Radeon RX 6600 (0000:03:00.0): GFX: 3%, MediaEngine: 0%, Memory: 2%
AMD Radeon Graphics (0000:08:00.0): GFX: 0%, MediaEngine: 0%, Memory: null%
...

fdinfo description

fdinfo for the AMDGPU driver shows hardware IP usage per process.

VRAM

GTT

Graphics Translation Tables.

KFD

The process of using the AMDKFD driver.

GFX

GFX engine.

Compute/COMP

Compute engine.
The AMDKFD driver dose not track queues and does not show them in fdinfo.

DMA

DMA/SDMA (System DMA) engine.

Decode/DEC

Media decoder.
This is not show on RDNA 4.

Encode/ENC

Media encoder.
This is not show on RDNA 4.

VCN, Media

Media engine.
From VCN4, the encoding queue and decoding queue have been unified.
The AMDGPU driver handles both decoding and encoding as contexts for the encoding engine.

JPEG

JPEG decoder.

VPE

Video Processor Engine.
ref: https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/amd/vpelib/README.md?ref_type=heads

Installation

Packages

• Releases
  • .deb (generated by cargo-deb)
  • .rpm (generated by cargo-generate-rpm)
  • .AppImage (generated by cargo-appimage)
• AUR
  • amdgpu_top
  • amdgpu_top-bin
  • amdgpu_top-git
• OpenMandriva to install run sudo dnf install amdgpu_top
• Nix
• Solus to install run sudo eopkg it amdgpu_top

Build from source

cargo install amdgpu_top

# or

git clone https://github.com/Umio-Yasuno/amdgpu_top
cd amdgpu_top
cargo install --locked --path .

without GUI

cargo install --locked --path . --no-default-features --features="tui"

Distribution specific instructions

Debian/Ubuntu

sudo apt install libdrm-dev

Binary Size

Features	Size (stripped)
json	~1.1M
tui	~1.5M
json, tui	~1.7M
json, tui, gui	~18M

References

• Tom St Denis / umr · GitLab
• Mesa3D
  • src/gallium/drivers/radeonsi/si_gpu_load.c · main · Mesa / mesa · GitLab
• AMD Documentation
  • R6xx_R7xx_3D.pdf
  • CIK_3D_registers_v2.pdf
  • MI200 performance counters and metrics — ROCm Documentation
• https://github.com/AMDResearch/omniperf/tree/v1.0.4/src/perfmon_pub
• https://github.com/freedesktop/mesa-r600_demo
• radeonhd:r6xxErrata
• Linux Kernel AMDGPU Driver
  • libdrm_amdgpu API
    • /drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
  • amdgpu_allowed_register_entry
    • /drivers/gpu/drm/amd/amdgpu/{cik,nv,vi,si,soc15,soc21}.c

Translations

amdgpu_top is using cargo-i18n with Project Fluent for translation.
Please refer to pop-os/popsicle for additional supported languages.

Supported Languages

• en
• ja (partial)
• ka

Alternatives

If amdgpu_top is not enough for you or you don't like it, try the following applications.

• AMD_DEBUG=info <opengl application> or RADV_DEBUG=info <vulkan application>
  • Print AMDGPU-related information
  • https://docs.mesa3d.org/envvars.html#envvar-AMD_DEBUG
  • https://docs.mesa3d.org/envvars.html#envvar-RADV_DEBUG
• clbr/radeontop
  • View your GPU utilization, both for the total activity percent and individual blocks.
• Syllo/nvtop
  • GPUs process monitoring for AMD, Intel and NVIDIA
• Tom St Denis / umr · GitLab
  • User Mode Register Debugger for AMDGPU Hardware
• GPUOpen-Tools/radeon_gpu_profiler
  • for developer
  • Radeon GPU Profiler (RGP) is a tool from AMD that allows for deep inspection of GPU workloads.