The Current State Of Radeon Power Management
The tests in this article are using several Radeon HD 4000/5000/6000 series graphics cards while monitoring several key metrics when running the open-source Radeon and proprietary Catalyst drivers. The Catalyst driver was tested in its stock mode while the Radeon DRM driver was tested in the default, low, mid, high, and DynPM modes. Monitored in the process were the following:
- The performance of the test itself in each of the forced power
management modes.
- The CPU usage (as a percent) of the system in each of the modes.
- The GPU frequency.
- The GPU operating temperature as reported by OverDrive / the i2c adapters.
- The GPU usage. The Catalyst driver exposes the GPU usage for modern Radeon graphics
processors as a percent of utilization. However, this is not calculated on the
open-source driver side. What the Phoronix Test Suite measures for the GPU usage
via the Radeon DRM is the number of fences handled per second.
- The GPU voltage. The Catalyst driver does not expose any real-time voltage information,
but the Radeon DRM driver does expose this as part of its power management interface.
Missing:
- What is missing from this article is actually the system power
consumption. While it is one of the most important measurements, the WattsUp USB
power meter generally used in the Phoronix tests for desktop power consumption
appears to have gone kaput. Only null packets are being read from the device,
so until the device can be replaced or other power equipment obtained those measurements
are not available. However, there is at least other interesting numbers in this
article to make it worthwhile.
- Unfortunately, there is also not any GPU fan speed information that is widely
exposed across Radeon graphics cards.
The graphics card testing included the Radeon HD 6950, Radeon HD 5770, Radeon HD 4870, and Radeon HD 4670 graphics cards. We begin the tests on the next page with the Radeon HD 6950 "Cayman" using the open-source driver power management options and the Catalyst blob. During this testing, swap buffer wait was disabled from the xorg.conf, but the color tiling support wasn't overridden due to some stability problems experienced on the Ubuntu 11.10 x86_64 + Linux 3.2 + Mesa 7.12-devel (git-1bb59b3) configuration used. The same Intel Core i5 2500K system was used for the duration of the testing. All of the tests were facilitated by the Phoronix Test Suite with the results and system logs then being archived on OpenBenchmarking.org.