So in the end it appears the AMD Radeon graphics performance on the open-source driver comes down to a change with CPUfreq and its ondemand governor. Now rather than CPU frequencies jumping frequently between their lowest and highest steppings, the CPU is staying in its highest power state for longer periods. There's nothing Radeon-specific, but if you have a newer CPU with many different frequency steppings and a mid-to-high-end GPU where you'll be more likely to be CPU bottlenecked is where the graphics performance will then tend to be better with Linux 3.12 as shown by the 10-way AMD Radeon Linux GPU comparison.
As stated already, in the benchmarks done of Intel "Ivy Bridge" and "Haswell" graphics on Linux 3.12 I haven't seen any significant performance changes like there were with the Radeon graphics. Nouveau Gallium3D benchmarks are in the process of happening and I hope to have those done in the next day or two to see how the performance changes. However, rather than the open-source drivers gaining ground against the closed-source AMD and NVIDIA graphics drivers, it's also likely the closed-source drivers will do much better now on 3.12 too thanks to this CPUfreq governor improvement.
From my Linux 3.12 kernel testing thus far I haven't seen any other particular workloads seeing much in the way of faster performance like there were with the Radeon graphics aside from some file-system performance regression fixes. Linux 3.12 testing though is still ongoing from the range of systems at my disposal.
This finding also reiterates somewhat I've long been pursuing and that's daily or per-commit testing of the mainline Linux kernel code. It's incredible that such changes and sometimes regressions still get in and leaves kernel developers baffled with real-world user-facing workloads. This is far from the first time I have spotted high-profile regressions within Linux that went unnoticed for days or even weeks. With the Phoronix Test Suite, Phoromatic, and OpenBenchmarking.org I have already developed the tools to provide automated kernel testing. It's easy to do testing on per-commit or daily (or per any other context) basis and with any number of test profile(s) that can be easily extended. I've done daily Linux kernel benchmarks going back to 2010 and that went on for over one year in demonstrating the Phoronix automated benchmarking capabilities for Linux.
That fully automated routine testing of the kernel sadly ended though due to hardware/budget constraints and the occasional time commitments when there would be a kernel panic or the system would fail to boot with various commits early on in each development cycle. However, as I shared last week, I'm actually in the process of restoring automated testing of the Linux kernel, Mesa, and compilers. Your support in this matter is appreciated by any means possible whether it just be feedback, a company looking to donate hardware to this test farm, a PayPal contribution, etc. This kernel performance change could have been easily spotted weeks ago had this tracker been running. With SteamOS and the Linux desktop growing, it's only becoming more pressing and urgent that such testing takes place.
Last but not least, if you appreciate this timely testing of the Linux kernel, hardware drivers and reviews, and other key open-source projects with benchmarks you can't find anywhere else, please consider subscribing to Phoronix Premium or making a PayPal tip. Besides contributions going to support Phoronix.com and my writing of thousands of Linux/open-source articles annually as the only main author of Phoronix.com, it also goes to help out the time I invest on the GPLv3 version of the Phoronix Test Suite as its lead developer. This allows for carrying out of new benchmarks, developing new features, better advance its system monitoring capabilities, etc. Commercial support and custom engineering services are also available for PTS. You can also follow @Phoronix and @MichaelLarabel on Twitter or like us on Facebook.
Next on the testing agenda will be investigating the Nouveau driver performance on Linux 3.12 and delivering a new AMD Radeon open-source vs. closed-source Linux graphics driver comparison. There will also be new power consumption / power efficiency (performance-per-Watt) benchmarks of Linux 3.12 given the CPUfreq governor change as well as other CPU governor benchmarks.