At 720p resolution, we see the stock GTX 680 holding a 20% lead over the stock HD7970. When they are max overclocked (albeit different clocks), the difference narrows to 9% which shows that the HD 7970 could actually be faster if the clocks are normalized. All the other cards are just spectators in this battle at the top, falling almost 50% behind.
At 1080p, the GTX 680's lead is narrowed to 13% at stock and just 4% overclock for overclock vs the HD 7970. Clearly the lack of memory bandwidth or shader efficiency is telling here, but the GTX 680 is leading nonetheless.
Unigine Heaven 3.0 (DX11)
With tesellation disabled, the extra memory bandwidth and better raw muscle probably helped HD 7970 to keep up with the GTX 680 in this notoriously fillrate heavy benchmark. Both the top dogs are head to head when overclocked while the rest are too asthmatic to keep up.
When tessellation is set to extreme, the GTX 680's superior algorithms puts distance between itself and the HD 7970. Notice the almost 50% drop here for the AMD cards with reference to the above diagram.
Either there is something wrong with NVIDIA's OpenCL driver implementation or the Kepler floating point units are bonked – the GTX 680 is even behind the old Fermi GTX 580! We see this happening on other OpenCL benchmarks that we've informally tested as well and will follow up with a separate piece on this.
We see a similar situation here with the compute mystery but this time with DirectX 11, and the results are significant here as they affect ambient occulsion performance in modern titles like DiRT 3 and Battlefield 3 (tested on the next page).
ShaderToyMark 0.20 (OpenGL Pixel Shader Test)
The GTX 680 regains the lead here but not by much. Once the cards are overclocked the HD 7970 surges in front.
SPECviewperf 11 (OpenGL Workstation Tests)
We threw in a workstation benchmark here (OpenGL heavy). No wins for the GTX 680 except in the catia test. Both vendors would pretty much like these type of users to buy Quadros and FireGLs, but not everybody can afford these thousand dollar cards that are essentially the same hardware as consumer grade boards.
Description of application tests (taken from SPEC.org):
The snx-01 viewset is based on traces of the Siemens NX 7 application. The traces represent very large models containing between 11- and 62-million vertices, which are rendered in modes available in Siemens NX 7.
The tcvis-02 viewset is based on traces of the Siemens Teamcenter Visualization Mockup application (also known as VisMockup) used for visual simulation. Models range from 10- to 22-million vertices and incorporate vertex arrays and fixed-function lighting.
The sw-03 viewset was created from traces of the graphics workload generated by the Solidworks 2009 SP2 application from Dassault Systemes.
The proe-05 viewset was created from traces of the graphics workload generated by the Pro/ENGINEER Wildfire™ 5.0 application from PTC. Model sizes range from 7- to 13-million vertices.
The maya-03 viewset was created from traces of the graphics workload generated by the SPECapc for Maya 2009 benchmark.
The ensight-04 viewset represents engineering and scientific visualization workloads created from traces of CEI's EnSight 8.2 application.
The catia-03 viewset was created from traces of the graphics workload generated by the CATIA™ V5 R19 and CATIA V6 R2009 applications from Dassault Systemes.
The lightwave-01 viewset was created from traces of the graphics workloads generated by the SPECapc for Lightwave 9.6 benchmark.