Twin Asus ATI HD5970 Dual-GPU cards in Crossfire

WE'VE SEEN that users buying fast GPUs usually tend to install them on systems having fast CPUs, too. A speedier processor or two will not only help serve the processed polygons, primitives and textures faster to the hungry GPU(s) but, on top of that, the usually greater memory bandwidth will help access all that directly from the memory system via PCIe when the graphics cards need it.

Now, if you were to get the fastest desktop GPU combination available to date, it would most likely be a pair of AMD ATI Radeon HD5970 cards, each with two GPUs and dual 1GB banks of fast GDDR5 memory. Until Nvidia gets out a dual GPU version of the Fermi on the desktop, the HD5970 - and its rumoured faster successor next quarter, the HD5990 - are expected to rule the roost in raw GPU power. Put two of these in Crossfire, and there should be an 8 TFLOPS single-precision floating-point computing monster in your PC.

The Asus HD5970 cards used here are standard reference type units. Asus does have voltage tweak utilities for overclocking, however we stayed away from pushing the cards even further here as this twin GPU contraption is known to be quite hot even without speeding it up any further.

So, assuming we choose a set of platforms with full dual x16 PCIe v2 slots available, and standardise the operating environment - in this case, it was Windows 7 64-bit with somewhat old but reliable ATI Catalyst 10.1 drivers, and use that as a baseline, what will we get when we compare 3D Mark Vantage results? Will a faster CPU only result in faster CPU scores, or will it also impact the GPU only results?

Here I put together three test configurations. These Asus ROG mainboards from the X48 chipset generation were among the best, if not the best, out there, and still command respect:

1) Asus Rampage Formula X48 with Intel quad-core Core 2 Extreme QX9770 at 3.2GHz FSB1600, 4GB G.Skill DDR2-1200 memory and two full x16 v2 PCIe slots:

2) Asus Rampage Extreme X48 with Intel quad-core Core 2 Extreme QX9770 at 3.2GHz FSB1600, 4GB Geil DDR3-1600 memory and two full x16 v2 PCIe slots:

3) The upcoming speed and cost champion, EVGA SR2 with two Intel six-core Xeon X5680 CPUs at 4GHz, 12GB Kingston HyperX DDR3-1600 memory and four full x16 v2 PCIe slots:

Note that all boards were using the default standard PCIe speed settings, which combined with default clock GPU operation should give as much as possible an 'apples to apples' comparison.

So, three times more cores, triple the number of memory channels and roughly 25 per cent higher per core speed at the very least. How will it impact the results?

Here they are, first the X48 DDR2.

Then, its Extreme DDR3 sibling. Notice that despite the subtle difference that having that bit of extra bandwidth from the memory to the CPU, even though the CPU FSB can't really make use of it, still results in higher CPU scores, but somehow lower GPU scores.

And last, we move to the dual overclocked Xeon on EVGA SR2 with its twelve cores running at 4GHz as well as four x16 PCIe slots.

As you can see, this must be one of the record combined CPU and GPU 3Dmark Vantage scores around anywhere right now, whether looking at the Performance 1280 x 1024 mode or the Extreme 1920 x 1200 mode with all the bells and whistles turned on. Also, note that the GPU score portion alone jumped by some 10 per cent just from the triple to quadruple peak CPU performance jump.

An interesting exercise, both for benchmarking and for the poor Thermaltake ToughPower 1000 PSU that had to feed the EVGA mainboard with twin 4GHz six-core fast Xeon CPUs and dual HD5970 GPUs throughout all the tests. Of course, since EVGA will be offering Nvidia dual SLI, tri SLI and quad SLI setups with up to GeForce GTX480 graphics cards on them, why not contemplate a similar multi-GPU setup that can then be done using the high end ATI graphics cards as well. It would be a great choice offer mainboards with the best of either vendor's graphics cards.

In the follow on feature, we will look at the performance in the same setup using AMD Phenom X4 processors as well as Intel's single CPU Core i7 processors with the X58 chipset. µ