Conclusion

So let us take a deep breath and rethink the facts we have experienced on the last pages. Firstly, we pay the tool Furmark some attention. It is an application which produces a high load on the GPU and seems to have no competition at the moment. The long-serving and popular 3DMark06 is still used often to measure the power consumption but it basically cannot stand up to Furmark in this discipline. It depends on the graphics card but 3DMark06 regularly stays 20-30 % behind and this is not a new insight. For years this fact has been approved regarding to measurements in our game-benchmarks. Anyway, it is not quite comfortable to take games for measuring the power consumption of graphic cards. Either the scenes can be readjusted very badly or the tester has to wait for suitable scenes in the benchmark. Accordingly, 3DMark06 with its benchmark options and loops is in comparison very comfortable.

But Furmark is even more comfortable since it produces a constant high load which doesn't let the graphics card time to cool down again. This affects the temperature and indirectly the power consumption. The producers certainly haven't expected an application which creates such high requirements - at least not in public, although we suspect similar tools at AMD and NVIDIA. AMD even reacts by putting a protection on the Catalyst, which should artificially reduce the strains of Furmark (renaming the exe file will avoid this problem).


The exceeding of own "TDP"-designations and serious violation of the PCI Express Specification demonstrates that neither AMD nor NVIDIA did anticipate the Furmark. With their Dual-GPU graphic adapters both put their worst instincts on display. The AMD Radeon 4870 X2 draws more than 200W out of the 8-pin connector which was specified for 150W and also NVIDIAs GTX 295 behaves inglorious by drawing 104W out of the 6-pin connector which was designed for 75W. Our measurements show: when it is about presentation of performance in the high end market segment both manufacturers are on their limits or rather exceed them massively. In contrast to that with mainstream adapters (where the most money is made) AMD and NVIDIA do not violate their specifications, mostly they even go below them.

You say Furmark is not "real world"? Possible. But is 3DMark "real world"? Or CPU stress tools like Core2MaxPerf, Intel BurnTest or Prime95? Like them Furmark is a synthetic benchmark and definetively a stress tool! It is virtually certain a piece of software comparable to internal test tools of AMD and NVIDIA. In times of CUDA and more demanding games than 3DMark's sequences one wants too know how much power is drawn in the worst case and if the system is still stable. Thats what stress tools are for.

Generally both companies appear to have no upper limit in power consumption for their high end cards. Even in the idle mode (where efficiency is important for the electricity bill) only few models can satisfy the customer. Merely the GeForce GTX 285 can score some points within the high end group with a consumption of 28W in idle mode. In the middle class the AMD Radeon 4670 demonstrates with 8W in idle mode that a lower consumption is possibe and nevertheless delivers enough rendering power for the casual gamer. To sum the things up: we will see Green-IT developments reach the high end graphic adapters market segment too. But there are not much signs indicating that this will happen in the near future.

Finally we would like to thank our good old friend Wolfgang Mattes for his help on translating this article.

[ls], 29. Januar 2009

Comment this Article in our Forum. English postings are welcome, too.