Here's what happens when you sacrifice CPU for GPU

AFTER NVIDIA'S recent rant about GPU being more important than the CPU, it was Intel's turn to fire back. Naturally, the chip maker reckons the CPU is the be-all end-all. Yes, if you don't run 3-D graphics, the usual application multitasking and multi-threaded performance, will benefit from adding more CPU cores. All apps will also benefit from upping the GHz or, for that matter, the memory speed and capacity.

On the other hand, games are increasingly sensitive to the GPU power provided and, once physics also moves to the 3-D cards, the dependence is bound to grow. If you reduce the CPU part, is it really the case that the graphics shine? How much will the scores vary?

We built a simple set-up to try out the 3-D gaming benchmark response in such situation: the reference board this time was the new Asus Striker II Extreme using Nvidia Nforce 790i Ultra chipset. As mentioned sometime ago, this is the only non-Nvidia mobo design for the 790i Ultra as of now, which has miffed some memory vendors since, according to one of them, DDR3-2000 doesn't seem to run on this board but it does on the Nvidia reference design. Maybe they should try to liquid cool the North Bridge, as Asus provided all the connectors for it?

Two OCZ ReaperX heat-piped 2GB DDR3 DIMMs provided 4GB RAM at DDR3-1600 CL8 speed: not bad for 2GB DDR3 modules. An OCZ PC Power & Cooling 860W power supply fed the assembly, running hot in 35C or 95F Singapore's humid heat. This PSU is a great test bed, since it's very compact size-wise, yet has all the power and connectors to feed twin 9800GX2 - or, in the future, twin GTX280 or 4870X2 units - at one go.

The CPU was just a dual-core Intel Wolfdale Penryn running at 3.2 GHz / FSB1600 - not a big deal, one would argue: these CPUs can do another 25 per cent without a hitch, especially on high-FSB boards where their multiplier lock doesn't get in the way of the GHz. Nevertheless, it is kind of slow coach compa red to the freeze-cooled 4.27GHz quad-core QX9770 reference setup on the X38 chipset. This CPU choice should accentuate the GPU-related benchmark result differences when switching various graphics cards around.

The new 3Dmark Vantage, with its more realistic and far more taxing test, some imposing Crysis-like stresses on the whole system, was here for the first time. Generally, it does seem to differentiate between and more sensitive to the CPU, GPU, memory and overall settings far beyond its predecessor.

And, that's exactly what we did - a pair of Asus GeForce 9800GX2 cards was here on default 600 MHz GPU clocks, as well as a pair of newer Asus 9800GTX TOP, factory overclocked to 775MHz GPU and 2.3GHz memory. We ran both single and dual-card setups on Windoze Vista with 3Dmark06, 3Dmark Vantage, and PCmark Vantage too (more on that last one in the next story). Here are the first results:

For some further fun, we also plugged in Asus Ageia Physx card to check the impact on the 3Dmark Vantage CPU core.

Oh boy, doubled CPU core overall, just by adding this card? It'd be interesting to see the impact once Physx runs on the 9900GTX, sorry GTX280, next month.

As you can see here, the 3Dmark Vantage scores are less CPU dependent, and more GPU dependent, overall. They scale far, far better in the quad-GPU configuration compared to the older 3Dmark 06, and even the usual dual-GPU, SLI or single card, looks better. Unfortunately, the Extreme Dual 9800GTX SLI scores are still not there at the press time due to consistent crashes.

Would it be a good measure of new-generation game dependence on GPUs? Yes, if you consider 3Dmark Vantage to be a good peformance comparison reference. On the other hand, if your games include any significant amount of AI, for instance simulating a city battle between two large armies with individual soldier and weapon behaviour, the CPU will very much play just as important role - such games are increasingly well multi-threaded too. Over here, somehow even a dual core CPU seems to have a very high - translate as overdone? - rate, while only a quad GPU setup has a GPU score overwhelming the CPU score on the balance scale.

In summary, if you cut down on your CPU but keep an expensive GPU, yes the frame rates in 3Dmark Vantage will still look playable, but somehow there seems to be less consistent frame rate curve when you hit scenes with more AI content. Also, there is obvious taper-off with the quad-GPU dual 9800GX2 setup - even though dual-channel DDR3-1600 RAM is sufficiently fast to feed both cards well, the CPU side, at that points, seems to need a jump to quad core to up the scores further.

With two 9800GTX TOP cards in SLI, almost all 3Dmark06 WUXGA 1920x1200 and 3Dmark Vantage Extreme 1920x1200 runs crashed at some point - 3Dmark06 always in Deep Freeze (the very last test, darn!) and 3Dmark Vantage at the middle of Jane Nash. Changing drivers, board BIOS, etc. didn't help, and the SLI patch was in Windoze Vista from the very start anyway. Only after about 10 runs, and doing it in cooler night weather, did we get the results for 3Dmark06, while the Vantage Extreme is still not there.

We're now re-running the whole shebang with a quad-core Intel 45nm darling, at exactly the same 3.2GHz FSB1600 speeds. It will take another day or two to go through the motion - just in time for Computex. µ