tuned-amdgpu

Hacky solution to integrate AMDGPU power profile control in tuned with Ansible

Takes a list of existing tuned profiles and creates new ones based on them. These new profiles include AMDGPU power/clock parameters

An attempt is made to discover the active GPU using the 'connected' state in the DRM subsystem, example:

$ grep -ls ^connected /sys/class/drm/*/status | grep -o card[0-9] | sort | uniq | sort -h | tail -1
card1

Warning: This is only tested with RX6000 series GPUs, it is probable that older AMD GPUs will not work properly. Use at your own risk!

Profiles

An example of the output/provided profiles follow

Output profile	Description
balanced-amdgpu-default	Includes the (assumed) existing balanced tuned profile. Only adjusts the GPU power limit (typically lower). Clocks/voltage curve remain the default.
desktop-amdgpu-VR	Includes the (assumed) existing desktop tuned profile. Adjusts the GPU power limit, clocks, and the voltage curve. Uses the predefined VR profile in the driver. See /sys/class/drm/card*/device/pp_power_profile_mode
latency-performance-amdgpu-custom	Includes the existing latency-performance tuned profile. Like the existing GPU profiles (eg: _VR)), this also adjusts the GPU power limit, clocks, and the voltage curve. This differs by using the custom profile in the driver. This opens up further tweaking of the power/clock heuristics through the driver (currently manual). see: pp-dpm

Note: This is non-exhaustive, see the variables base_profiles and amdgpu_profiles below for the (default) sources of the merged profile mapping

Notable variables

These are the variables you're likely to want to change. They are defined in playbook.yml

Variable	Description	In-playbook
gpu_clock_min	Sets the minimum (dynamic) GPU clock (in Mhz) for the non-default amdgpu profiles	700
gpu_clock_max	Sets the maximum (dynamic) GPU clock (in MHz) for the non-default amdgpu profiles	2600, results in 2.6GHz (rounded); mild overclock
gpumem_clock_static	Sets the static memory clock for the GPU (in MHz). This is not the effective data rate. That is a multiple of this depending on the type of VRAM. To avoid flickering this does not change dynamically with load.	1050, results in just over 1GHz; mild overclock Actual effective clock depends on this being multiplied against the data/pump rate of the GDDR? GPU memory
gpu_mv_offset	GPU core voltage offset. Takes +/- some integer in millivolts. Can be used to both over and under volt.	-10 (undervolt 10mV or 0.01V)
base_profiles	List of base tuned profiles to clone in the new AMDGPU profiles. Defaults based on Fedora	balanced desktop latency-performance network-latency network-throughput powersave virtual-host
amdgpu_profiles	Dictionary mapping the AMDGPU power profiles found in /sys/class/drm/card*/device/pp_power_profile_mode and custom power limits. For each item, two keys: pwrmode and pwr_cap_multi. pwrmode maps to the number assigned in /sys above. pwr_cap_multi is a multiplier against board power capability. Must be a float, eg: 0.5 for 50%	default: pwrmode: 0 pwr_cap_multi: 0.75 # 75% relatively safe default VR: pwrmode: 4 pwr_cap_multi: 0.8 # 80%, likely slight boost custom: pwrmode: 6 pwr_cap_multi: 1.0 # 100%, full GPU board capability # warning: significantly increased heat