F8x86_64 on the Acer Ferrari 3400LMi
10.3.1 Simple benchmark
In order to compare different settings we need some kind of benchmark tool. A simple tool that comes with the glxutils package found in most distributions is glxgears. It may be rough, rudimentary and lacking a lot of hype features, but it is present and sufficient for our needs.
An interesting exercise is to explicitly disable hardware acceleration and comparing the result with the default xorg.conf. Without hardware acceleration glxgears clocks in around a modest frame rate of 150 FPS, compared with approximately 2050 FPS for the default configuration. This is quite a difference, and it verifies that the radeon driver really does its job.
10.3.2 Optimization
The default values for most settings work well and there is no needed to modify xorg.conf. Furthermore, most of the other settings are correctly autodetected, such as AGP 8x and memory. All this is reported by the X server in its log file /var/log/Xorg.0.log during startup.
The options of most interest for performance are AccelMethod (default XAA), AGPMode (autodetected), ColorTiling (default on) and EnablePageFlip (default off). The first three are correct by default, but due to instability in rare cases EnablePageFlip is disabled by default.
Turning on the option EnablePageFlip in xorg.conf reveals no flaws. I have not noticed any glitches with this option on my setup so I feel confident in recommending it. Then the obvious question How good is it? It is good!
glxgears gives us an indication. By enabling EnablePageFlip performance is increased from 2050 FPS to about 3270 FPS. A significant boost for tweaking one single option.
You only need to add one line to the Device section in the default xorg.conf to boost 3D performance:
Section "Device" |
|
|
Identifier | "Videocard0" |
|
Driver | "radeon" |
|
Option | "EnablePageFlip" | "1" |
EndSection |
|
|
10.3.3 Other observations
During the optimization procedures a few interesting observations were made. Primarily the CPU speed does not seem to matter. Most tests were run with the CPU frequency first set to 800MHz, then repeated at 2000MHz. No significant difference related to the CPU speed was noticed.
Furthermore, the size of the configured virtual screen does not affect
20
