FutureMark unveiled today 3DMark 06, just a little bit more than one year after the previous version. It is surprising because the time gap between the two previous versions was 2 years. We don't use 3DMark in our surveys because we prefer basing our tests on results obtained in several games. We have however to admit that 3DMark is now the reference benchmark for the industry and that it has an important impact. This is the reason why we have looked at it.
First, since only one year passed between the 2006 and 2005 version, there hasn't been any DirectX evolution. Technologies employed but also test scenes are identical (except for a fourth scene which is new). In other words, shaders 2.0 and 3.0 are included just like the 2005 version. Graphic effects evolved however and are now much more complex. Scenes generally also include more details. Graphic tests are divided in two groups.
SM2 tests
The first one uses an improved version of the3DMark 2005 first two scenes and represents, according to FutureMark, the shaders 2.0. These scenes however also use optimised shaders 3.0 for cards which support them, but all that is rendered in the first two scenes can be done with basic shaders 2.0. It is also true for the pixel shaders 2.x with a profiles for the GeForce FX (ps 2.a close to the ps 3.0) and for the Radeon X800/X700 (ps 2.b, closer to the ps 2.0).
Shadow rendering has been strongly improved and in these tests it uses hardware acceleration when it is available. It means using a D24X8 type surface and PCF filtering (permits to mix 4 shadow samples in a way that is adapted to shadow filtering) for NVIDIA but also DF24 surface and fetch4 function for the Radeon X1000 (except for the X1800 which doesn't support those 2 functions). Fetch4 recovers four non filtered samples in one cycle instead of 4 distinct accesses, but filtering must always be done via the pixel shader. This is what 3DMark06 do. For other cards which don't support it, 4 samples are recovered from a R32F surface (the first one that is just above the other two). It means more important bandwidth and memory consumption and slightly longer pixel shader.
HDR/SM3 Tests The next 2 tests are named shaders 3.0/HDR and need these functions to run. HDR actually means blending FP16 for FutureMark. FP16 filtering is optional and emulated in the pixel shader for cards that don't support it. The first one of these tests is an improved version of the third test of 3DMark05. More complex shaders and HDR rendering are included. The fourth test is a new test that takes place on a quite empty pack-ice.
In these tests there is a more in-depth use of the shaders 3.0 with dynamic branching (we don't know if they are use during strong granularity operations or if they are only used when it isn't too much of a problem for the GeForce 6 and 7), partial derivative, high number of instructions etc. It makes them inaccessible for cards that don't support the shaders 2.0 or 2.x. These 2 scenes beneficiate too from improved shadows based on a similar technique to the first two tests but uses a filtering in the shader based on 16 samples (PCF and fetch4 aren't used).
CPU and theoretical testsMajor innovation of this new 3DMark, 2 CPU tests are now included in the global rate. It makes even more difficult the graphic cards comparison unless if you use SM2 and SM3/HDR under scores. This choice is justified by FutureMark, by the fact that 3DMark is a "gamer" benchmark and so it is logical to associate CPU scores. This is something that must make the CPU manufacturer, which desperately try to convince gamers that dual core CPUs is what they need, very happy.
The 3DMark05 theoretical tests are still there and now come along with two new ones. The first, the Shader Particles Test calculate the physique for a high number of pixels and require for this operation the shaders 3.0 but also the Vertex Texture Fetch that ATI doesn't support in the Radeon X1000 thanks to a skilful trick with the DirectX specifications. Only the GeForce 6 and 7 can run this test for now. The second new test, Perlin Noise, represents procedural texturing and uses a very complex shader: 447 mathematic instructions on top of the 47 texture accesses (almost 500 instructions).
Score calculation :SM2.0 = 120 x 0.5 x (SM2 GT1 fps + SM2 GT2 fps)
HDR/SM3.0 = 100 x 0.5 x (SM3 GT1 fps + SM3 GT2 fps)
CPU = 2500 x Sqrt (CPU1 fps x CPU2 fps)
For global score it is even more complicated. An intermediate global value is calculated for SM2 and SM3/HDR is calculated differently for SM2 cards (and SM3 without HDR like the GeForce 6200) and SM3 cards:
GS for SM3.0 cards = 0.5 x (SM2S + HDRSM3S)
GS for SM2.0 cards = 0.75 x SM2S
Global score = 2.5 x 1.0/ ((1.7/GS + 0.3/CPU Score)/2)
A few results :
This new 3DMark particularly appreciates the GeForce 7800. The GTX 256 MB model is right behind the X1800 XT even though the gap was much bigger in the 2005 version. Will the Radeon X1900, which will be announced very soon, will be able to take back the first place?
Even if the GeForce 7300 Go was announced in November and the rest of the range two weeks ago, NVIDIA chose today to unveil the desktop equivalent. This new cars is higher clocked than the GeForce 7300 Go (power consumption restrictions are less important outside of the mobile world). 
