For those looking to purchase a low or mid-range graphics card for use with the open-source graphics drivers -- rather than being bound by NVIDIA's proprietary driver or AMD's Catalyst -- here's a comparison of nine different discrete graphics cards when benchmarked by the open-source drivers.
When using the open-source Linux graphics drivers found "out of the box" on Linux desktops, generally the lower-end graphics cards perform the best since these drivers don't take advantage of all optimizations and capabilities found with the higher-end graphics cards. These lower-end graphics cards generally have less core/memory frequency steppings, power management, and other features to worry about, especially in the case of the Nouveau driver that's a reverse-engineered implementation of NVIDIA's official graphics driver.
Aside from the lower-end graphics cards generally working better, for both NVIDIA and AMD, hardware that's usually two to three generations old also does the best in terms of open-source support. This open-source support is generally well tested and mature by the time of being two to three generations old. The sweet spot at the moment seems to be the Radeon HD 5000/6000 series on the AMD side and the GeForce 9 series for Nouveau. However, there are exceptions, with regressions frequently being spotted at Phoronix. It was just earlier this week that I discovered the Radeon 3000 IGP to be broken and went undetected into a stable release.
Testing was done with Mesa Git and the Linux 3.8 stable kernel while using the latest xf86-video-ati and xf86-video-nouveau stable releases. In terms of regressions this time around when trying to facilitate a low-end NVIDIA/AMD Linux GPU comparison on the open-source drivers, the problems encountered all happened with the NVIDIA graphics cards. As I've said many times, the game that the Nouveau driver runs the best at is that of Russian Roulette; regressions of hardware support, driver features, and/or performance is all too common between releases. The Radeon HD 4000/5000/6000 graphics cards at least worked without any blatant failures.
- The GeForce 8 series (at least the GeForce 8500GT, 8400GS, and 8600GT) there were problems for most of the Linux OpenGL benchmarks attempted. Most commonly during the GeForce 8 series problems there were PGRAPH kernel error messages pertaining to the "TLB flush idle timeout fail" and/or GPU lock-ups.
- The GeForce 9500GT graphics card was working for the most part, but some rendering corruption was spotted within the Reaction Quake 3 game. Other OpenGL games appeared fine and the GeForce 9 series support overall seems among the best.
- Re-clocking is still unreliable for the GeForce 9600GSO graphics card.
- The GeForce GT 220 used to work fairly well on Nouveau, but now mode-setting fails for it on the Linux 3.8 kernel with no monitor signal.
- The GeForce GT 240 continues to show a checkerboard memory problem after mode-setting, a problem that's long been happening for this graphics card.
- The GeForce 400/500 "Fermi" and GeForce 600 "Kepler" graphics cards still have no re-clocking support yet so the graphics cards with their core, shader, and memory clocks can actually run at their rated speeds.