IN THE HIGH END ARENA the AMD Radeon HD6990 and Nvidia Geforce GTX590, both dual GPU cards, now share the top performance position in PC 3D. Both cards have very high power usage, large size and lots of RAM. The 3GB of memory on the GTX590 and the 4GB on the HD6990 equal the average mainstream PC main system memory capacity.
The Radeon HD6990 is particularly interesting as there has been barely any clock frequency sacrifice over the single GPU HD6970, despite fitting two of these on a single card and adhering to all the power limitations. The only other performance penalty (and very slight at that - below one per cent) is the PLX PCIe bridge chip on each card, required to enable two GPUs to share one PCIe x16 interface to the host computer.
In applications beyond gaming, multi GPU setups like this are usable in workstations and even in supercomputing application niches, where each GPU can run a task, handling accelerated threads or specific routines that make sense to run on them. After all, the single HD6990 can offer nearly 6TFLOPs in single precision and 1.5 TFLOPs in double precision. However, running a single thread across multiple GPUs efficiently is right now still very hard to do as there is no clear shared memory model between two or more GPUs due to the PCI express bus communication. Once this is solved, though, there will be many more suitable applications for multiple GPU usage.
But back to the gaming world. Since new AMD Catalyst drivers did improve multi GPU Crossfire scaling quite a bit last year, it was an interesting question of how high scores would be achieved in the usual 3Dmark benchmarks with two HD6990 cards, with four GPUs in total. Of course, with that many high end graphics processors there has to be sufficient CPU horsepower to feed them.
So we combined the new Intel Core i7 990X CPU, running at 3.46GHz with the trusted Gigabyte X58A UD9 mainboard, the well featured top end overclocking reference and possibly the most expensive high-end single processor mainboard ever. It has two NF200 Bridge ASICs from Nvidia to enable four full x16 PCIe slots, resulting in a little larger than usual size. The 6GB of GEIL DDR3-2133 gamers' memory, Intel X25-M 160GB SSD and Antec 1200W high end power supply, are easily capable of feeding the high-end hardware. The 64 bit Windows 7 install used default settings in the drivers and benchmark test applications.
Here are the 3Dmark Vantage and 3Dmark 11 test results. They speak for themselves:
As you can see, the combination of four GPUs and six CPU cores on a good motherboard can - even at default mode without any overclocking - achieve very high scores. This is up to some three times the score of a single HD6970 on the same processor platform, a very reasonable scaling benefit in either test.
Yes, the benchmark is just a benchmark, but it still shows well the scaling results and the platform's performance potential. Now, the issue of heat does come up, so you'd better ensure that the case is spacious enough with strong airflow to match, as graphics cards these days create way more heat than the processors. µ