This past Friday we had delivered benchmarks comparing the performance of the open-source Radeon driver against the new closed-source fglrx driver from AMD. These benchmarks had just looked at the AIGLX performance when using Compiz on an Ubuntu 7.10 desktop. In all of the benchmarks except one, the fglrx driver had carried a staggering lead over the open-source competition. In addition to these Compiz benchmarks, on the same system we had also ran some additional benchmarks to see for gaming and 2D rendering how the two ATI Linux drivers compare.
The system once again was running Ubuntu 7.10 "Gutsy Gibbon" with the Linux 2.6.22 kernel and X server 1.3, but with these benchmarks, the Compiz effects were disabled during testing. The hardware included a PCI Express ATI Radeon X800XL 256MB graphics card, Intel Pentium D 820 (2.80GHz dual-core), 2GB of DDR3-1333 memory, and an ASUS P5E3 Deluxe (Intel X38) motherboard. We had used Enemy Territory and GtkPerf as our vehicle for benchmarking the two drivers, since both benchmarks are compatible with the current Radeon driver. The ATI driver used was fglrx 8.42.3.
The last time we had compared these two drivers was back in February with X.Org 7.2: ATI Open v. Closed Drivers. During that examination, Fedora 7 Test 1 was used with the fglrx 8.34 driver, and of course, that was running off AMD's old driver code-base. Mesa and the Radeon driver have also seen a few improvements since that point. Most notably, Mesa 7.0 does contain a number of improvements. In those earlier benchmarks, the Radeon X800XL with the open-source driver was over twice as slow as the X800XL with the fglrx 8.34 driver. However, in these benchmarks we had not looked at the 2D GTK+ performance.
For our Enemy Territory benchmarks, we had completed the timed-demo tests with a resolution of 800 x 600 and 1280 x 1024. For the GtkPerf tests, which as the name implies tests the GTK+ performance, we had reported the results for GtkComboBox, GtkCheckButton, GtkTextView - Add Text, GtkDrawingArea - Text, and the total time needed to run all GtkPerf tests. Each GtkPerf test was run 1,000 times.