AMDGPU Performance Tests With New WattMan-Like Settings, Power Capping

With the recent stable debut of the Linux 4.17 kernel, one of the most common performance test requests coming in has been for checking out the Radeon WattMan-like support that was introduced with the Linux 4.17 AMDGPU code for recent generations of Radeon graphics card. Here are some benchmarks of that and on a somewhat related note also some Linux gaming benchmark results when carrying out some power capping tests to restrict the graphics card to a given Wattage.

WattMan is AMD's new approach for power management with fine-grained control over GPU voltage, engine clocks, memory clocks, fan speed, and temperature. They've been offering a lot of GUI-based information and controls via the Radeon Software Windows driver while under Linux 4.17 is the first time seeing such controls exposed under Linux, albeit through the CLI.

The new functionality is exposed via /sys/class/drm/card0/device/pp_power_profile_mode where there are predefined modes for 3D full screen, power savings, video, VR, compute, and a custom mode. You can write to that file with the desired index for changing modes.

When first trying it out on Linux 4.17 and 4.18-based kernels, there were errors when trying to write to /sys/class/drm/card0/device/pp_power_profile_mode and from some searching others also ran into this issue. Before toggling the pp_power_profile_mode, you first need to write "manual" to the /sys/class/drm/card0/device/power_dpm_force_performance_level.

Complementary to those test runs, I also did some basic tests when manipulating the power cap as another test I've been meaning to work on. The power cap to limit the GPU to an effective power draw can be done by writing the desired value to /sys/class/drm/card0/device/hwmon/hwmon1/power1_cap while also seeing the power1_cap_max for the maximum allowed value and power1_cap_min for the minimum power value that can be set.

With testing from the Linux 4.18 Git kernel as of this week on an RX Vega 56 and RX 580 graphics cards, adjusting the pp_power_profile_mode didn't appear to have much impact either under load, on the overall AC system power consumption, or the GPU temperature for that matter. Adjusting the power1_cap did have an obvious affect as you can see from the results.

Take this early data as you wish while I continue exploring with AMDGPU power tests. You can find the RX 580 data including AC system power usage and thermal and performance-per-Watt data as well as similar metrics on the Radeon RX Vega 56.