The OpenGL Speed & Performance-Per-Watt From The Radeon RX 480 To Radeon HD 4850/4870

With the Radeon RX 480 Linux review now being out of the way and our various other RX 480 Linux benchmarks, the latest results I have to share with being a benchmarking fanatic are RX 480 results with high-end AMD GPU tests of each generation going back to the Radeon HD 4850/4870 (RV770) days. This article has high-end GPUs from the RX 480 to RX 200, HD 7900, HD 6900, HD 6800, HD 5800, and HD 4800 series compared side-by-side with the latest open-source Radeon Linux graphics driver code. Not only is the raw performance being looked at but the system power consumption was also being polled in real-time for looking at the performance-per-Watt too. For any other benchmarking fanatics curious about the Radeon GPU evolution over the past eight years (RV770 launch in 2008), here are the numbers to enjoy.

All of these tests were done with the latest open-source AMD Linux driver stack: Linux 4.7 Git and Mesa 12.1-dev git-6b0ac95 along with the latest user-space components (and LLVM 3.9 SVN) via the Launchpad Padoka PPA. Compared to the proprietary driver that has long abandoned older generations of GPU support, the latest open-source code continues to be maintained for all of these legacy GPUs going well before the Radeon HD 4000 series. The Radeon HD 4000 series was used as far back as testing for still providing reasonable OpenGL 3.3 support and not being too weak for comparing to modern GPUs.

The graphics cards tested for this OpenGL/perf-per-Watt comparison were the Radeon RX 480, R9 285, HD 7950, HD 6950, HD 6870, HD 5830, HD 4870, and HD 4850. The high-end GPUs used of each generation was basically limited to the relevant x80/x90 series hardware I had available: unfortunately, I have better coverage of older AMD GPUs back when AMD used to send over more review samples than they do now, but at least they did seed the RX 480 this time around rather than having to buy it. One exception to benchmarking the x80/x90 hardware I had available was needing to omit the Radeon R9 290 from benchmarking since its performance is currently in a heavily regressed state under the Linux 4.7 kernel that was used for testing. The R9 290/390 regression has yet to be addressed upstream.

For benchmarking this range of GPUs going back eight years, mostly OpenGL 2/3 games/benchmarks were utilized. With the pre-GCN open-source support, only the Radeon HD 5800 and Radeon HD 6900 series have full OpenGL 4 support right now (due to the other GPUs missing FP64 support -- or emulated support -- in the R600 Gallium3D driver). Besides needing to focus on OpenGL 3.3 and below for these benchmarks to test the full range of GPUs, the benchmarks couldn't be too demanding either for being able to push out more than a few frames per second on a HD 4850/4870. Thus the test selection I ended up with that also fit my strict benchmarking requirements included Counter-Strike: Global Offensive, Team Fortress 2, Dota 2, Insurgency, Portal, Tesseract, GpuTest, and Unigine Heaven. For the HD 5800/6900 and above, for some OpenGL 4 numbers were Metro 2033 Redux and Metro Last Light Redux.

In addition to fully-automating each benchmark in a 100% repeatable manner, the Phoronix Test Suite was also recording the system power consumption as each OpenGL test was being executed. The Phoronix Test Suite was polling the AC system power consumption via a WattsUp Pro USB power meter. For some additional figures, the GPU temperature was also being charted in real-time via the Phoronix Test Suite.

Let's take a look at these fun results from the HD 4850/4870 to the RX 480. But before getting too far, if you appreciate all of the unique Linux benchmarking work done at Phoronix, please consider hopping into our premium program to support our site, view the ad-free version of the site, view multi-page articles on a single page (such as this longer piece!), and benefit from other features.