NVIDIA GeForce GTX 680 "Kepler" On Linux

Getting back on track, here are the much-anticipated GeForce GTX 680 Linux results using the NVIDIA binary driver. Initial benchmarks of the NVIDIA GeForce GTX 680 from the initial open-source Nouveau NVE0/Kepler code is complete and will be published in the next few days. The graphics cards used for this GTX 680 Linux comparison included a Radeon HD 5830, Radeon HD 6870, Radeon HD 6950, and Radeon HD 7950 on the AMD side. The Radeon HD 7950 was tested at its factory-overclocked speeds for the particular card (900/375MHz) and then at its reference clock speeds (800/1250MHz) and a third time when overclocked further to 1025/1450MHz.

The GeForce GTX 680 graphics up next to its distant sibling, the NVIDIA Tegra 3 "Cardhu" development tablet.

The comparison hardware on the NVIDIA side was the GeForce 9800GTX, GeForce GTX 460, and the GeForce GTX 550 Ti. The graphics hardware used in the comparison was limited to what hardware was available (which as aforementioned is limited on the NVIDIA side) and for both the AMD and NVIDIA GPUs the hardware went back a few generations since the Radeon HD 5000 / GeForce 9 series continues to be popular with some Linux users because that's where there is mature open-source driver support for those not wishing to use the proprietary drivers. Due to no Kepler overclocking support under Linux with NVIDIA's driver, the GeForce GTX 680 could only be benchmarked at stock speeds.

The testing was done from an Ubuntu 12.04 LTS "Precise Pangolin" x86_64 installation while using the Catalyst 12.3 and NVIDIA 295.33 binary drivers. Besides running a number of OpenCL and OpenGL benchmarks, the power consumption and GPU core temperatures were also monitored during the testing process and there are also performance-per-Watt graphs.

This power consumption was done using a USB-based WattsUp power meter monitoring the overall system AC power usage and automatically polled via the Phoronix Test Suite (just set the MONITOR=sys.power and PERFORMANCE_PER_WATT=1 environment variables prior to test execution). All benchmarking was done via the Phoronix Test Suite / OpenBenchmarking.org automated testing platform.