bit-tech.net

Fermi Testing Update

Fermi Testing Update

Following our reviews of the GeForce GTX 480 and GTX 470 there's been some confusion over some of the benchmark results, in particular those from our Colin McRae: Dirt 2 benchmark. It seems that websites testing with the demo version rather than the full game were reporting GTX 480 performance far in excess of the HD 5870 - a surprise considering in most other games the GTX 480 offers at best around a ten to fifteen per cent performance jump over the HD 5870.

The Dirt 2 benchmarking issue seems to stem from some review sites only using the demo version of the game. As it was released some months ago, the demo only recognises ATI hardware as DX11 compatible, with the GTX 480 and GTX 470 cards defaulting to DirectX 9, even though the hardware .ini file says otherwise. With one brand's card running with DX9 and the other DX11, the results obviously aren’t going to be directly comparable. There were even rumours of the Nvidia cards not actually employing dynamic ambient occlusion in the game, even when it was forced on.

Fermi Testing Update
Click to enlarge - Dirt 2 is a great looking game, but has someone been sneaky with in game performance?

We tested with the full retail version of Dirt 2 though, editing the game’s hardware settings .ini to switch between resolutions and AA settings having set all the ingame detail settings to max. We don’t even go so far as to trust the in game benchmark’s reporting system, instead recording frame rates using FRAPS. Our results still heavily favoured the GTX 480 and HD 470 though, more than other review sites out there, prompting questions from our readers as to whether our results were valid. At the time of testing we were unaware of the DX 9/DX 11 demo issues and despite using the full version of the game and taking every care in our results, we wanted to make sure our numbers were on the money. As a result, we’ve done some follow up testing, so check out the graph below.

Colin McRae: Dirt 2

1,920 x 1,200 4xAA, 16xAF, DirectX 11, Maximum Detail

  • Nvidia GeForce GTX 480 1.5GB
  • ATI Radeon HD 5870 1GB
  • Nvidia GeForce GTX 480 1.5GB - No Driver AF
  • ATI Radeon HD 5870 1GB - No Driver AF
  • Nvidia GeForce GTX 480 1.5GB - DX 9
  • ATI Radeon HD 5870 1GB - DX 9
  • Nvidia GeForce GTX 480 1.5GB - No AO
  • ATI Radeon HD 5870 1GB - No AO
    • 79
    • 70
    • 64
    • 55
    • 85
    • 76
    • 73
    • 63
    • 113
    • 101
    • 78
    • 67
    • 85
    • 74
    • 66
    • 58
0
25
50
75
100
125
Frame Per Second
  • Average
  • Minimum

As you can see, the difference in our results is that we force 16x Anisotropic Filtering on at the driver level, to get the optimum image quality. While disabling this improved performance on both the GTX 480 and HD 5870, enabling it causes the HD 5870 to take a much larger performance hit, at least at 1,920 x 1,200 with 4xAA, 16xAF which we’ve tested here. With AF forced to 16x the 480’s average frame rate is some 23 per cent faster than the HD 5870 but with it disabled (as many other sites have tested) the lead drops to just 16 per cent, more in line with the generally agreed 10-15 per cent performance jump offered by the GTX 480 in DX11 tests.

You can also see the huge differences apparent in the DX9 performance of the two cards, with the GTX 480 performing superbly when forced to DX9 with an average frame rate of 113fps and a minimum of 101fps. Meanwhile the Radeon gains comparatively little performance when switched from DX11 to DX9, jumping from 64fps average and 55fps minimum to 78fps average and 67 minimum.

We also wanted to make sure that dynamic ambient occlusion was working for both cards, and as you can see, both experience notable improvements with it disabled.

What’s important to remember though is that benchmark results from any given source are only as useful as those alongside them. Due to differences in test systems, testing methodology or, as we’ve found here, tendency to force settings in the driver to ensure they’re used it’s useless to compare our results to those from other tech sites. We’re confident of both our testing’s accuracy and relevance and want you to be too.

Related Reading

Nvidia GeForce GTX 470 1,280MB Review
Nvidia GeForce GTX 480 1,536MB Review
Fermi card on stage wasn't real
ATI Radeon HD 5870 Architecture Analysis