As you all know, ATI has recently announced a new chipset with a fashionable marketing feature — support for two full PCI Express x16 graphics interfaces, of course with CrossFire support. Before that, the same step was taken by NVIDIA with its nForce4 SLI X16 chipset (we shall not mention nForce Pro in this context). Before that time, if you wanted to build a SLI/CrossFire system, you had to content yourself with a slower combination of graphics interfaces. Most products on the market (chipsets from NVIDIA, ATI) used the x8+x8 formula in this case — throughput of the only x16 interface was distributed equally between two slots (PCIEx16 form factor, of course).
But Intel chipsets, which were the first to offer PCI Express, hadn't this feature even in the second generation (i945/955). That was why some motherboard manufacturers used the non-standard x16+x4 solution. In this case the first graphics slot is supported by the graphics interface proper (x16), while the second one is connected to the southbridge instead of the northbridge and uses its four PCI Express lanes, initially intended for periphery. While an i945/955-based motherboard with the ICH7R southbridge still had two PCI-E lanes, solutions with the ICH7 southbridge or i915/925-based motherboards were deprived of the new interface completely. (However, there are practically no PCIEx1 expansion cards on the market so far, so users of such motherboards actually lost only integrated controllers (LAN, SATA RAID), connected to PCI Express.) Only the i975X allowed to divide the graphics interface into two slots as x8+x8, like in competing products.
We thought it would be interesting to check if there existed the difference in real applications between various chipsets and various organizations of SLI/CrossFire at least in current solutions. The fact is even video card manufacturers themselves do not promise noticeable gains from upgrading to x16+x16. They carefully mention 10% for top models under heaviest conditions. This article will be devoted to this very issue. We shall use two video cards based on ATI Radeon X850 XT (our lab did not have enough faster video cards at the time of these tests).
We used nearly all existing platforms, considering that NVIDIA chipsets do not allow (yet) to run two CrossFire-accelerators, even though they have a split graphics interface for two SLI-cards. Thus, the AMD system stands in lone conspicuity — we haven't seen similar products on chipsets from VIA or SiS yet. Note that we shouldn't compare its results with the three Intel systems directly, as processors are obviously different (the only selection principle was being a top model to slow down the tandem of video cards as little as possible). Moreover, we didn't make it our aim to maximize results of each platform — quite on the contrary, we tried to level them down. For this purpose we used motherboards supporting DDR2 with 4-4-4 timings, as the ECS PF22 Extreme motherboard couldn't manage lower values. However, the ECS PA1 MVP (2.0) refused to report the timings it started up with (it did not allow to set them manually either).
So, our today's task is to watch out for possible differences between a CrossFire Ready motherboard (ECS PF22 Extreme) and a CrossFire Ready chipset (i975X) as well as the second chipset — ATI Xpress 200 CrossFire Edition, which name speaks for itself. Results of the AMD system are given as a reference point, though we are still going to make certain conclusions upon them. But before jumping at games, we should evaluate "abstract" performance of each platform, in order not to confuse a possible effect of a better CrossFire implementation and memory performance advantage.
Indeed, we can clearly see a performance difference between the systems.
The i975X gains a couple percents relative to [a given representative]
i955X and up to 5% relative to the ATI chipset. Different CPU architectures
inevitably result in the difference between Intel and AMD platforms.
But there is one interesting moment here, which was difficult to prognosticate.
All systems with Pentium 4 EE had almost no reaction to CrossFire,
while results of a motherboard with Athlon 64 4000+ went down (yep,
you didn't expect that the second video card would accelerate archiving,
did you?!) by nearly 3%. That's not much, but it's a demonstration
of Hyper-Threading benefits: the second logical processor takes up
Catalyst AI completely to itself.
However, video encoding practically doesn't slow down in all cases. So Catalyst AI does not cost too much.
Now what concerns 3D applications. Perhaps we should remind you that not all games gain from distributed rendering — either due to driver bugs or because system performance is not limited by a video card, but by another component (processor, as a rule). So, we haven't published our thoroughly taken readings in SPECviewperf (enabling CrossFire results in up to two-fold slow-down) and in Painkiller (±0%); Unreal Tournament 2004 will illustrate the general situation:
Let's advert to those two games with a noticeable CrossFire effect.
Of course, we shall test processors with any modern video card at
640x480 with minimal graphics quality. So this situation resembles
the archiving test.
1024x768 mode with High graphics settings is actually a hump, where
CPU capacities go up to the loaded X850 XT; now we can already expect
some gains from CrossFire. Indeed, a little gain is demonstrated in
all cases, though these 3—6% are evidently insufficient to make
you buy a couple of video cards. The difference between CrossFire
effects in three chipsets for Pentium 4 is so small that we cannot
draw any serious conclusions.
And finally, 1600x1200 with maximum graphics quality settings is
a benefit night for the ATI solution; let's have a closer look at
the results. First of all, note that in a single video card mode the
three systems are obviously limited by the X850 XT performance. Increased
performance of the Gigabyte G1975X can be explained only by automatic
overclocking of the video accelerator. We disable such functions in
BIOS Setup during our tests, of course. But perhaps it's impossible
to do it in this Gigabyte model. Now if we speak of the difference
in CrossFire effect, we can see well that the only AMD platform gains
more than the others (+57%), the other solutions have the following
results: ATI Xpress 200 CFE — +50%, i955X — +53%, i975X
The 640x480 mode in FarCry is no different from this mode in Doom
3 — the gain is either zero or negative.
1024x768 mode, High — there is some positive effect, but it's
small (5—8%). But note the difference in minimum fps: CrossFire
always causes a little performance drop, so the game here is limited
by a processor. Differences between the systems are very small, we
cannot speak seriously of CrossFire performance gains.
And now the most interesting thing: maximum mode. The same performance artifact in the i975X (CrossFire gain is just 54% due to a slightly increased overall performance), the other results being almost the same (~62%). Even the CrossFire tandem is probably limited by its performance here, while CPU performance is excessive. But we cannot say the same about the minimum fps value, the system with Athlon 64 4000+ is evidently superior here: this value in CrossFire mode is higher by 32%, while the systems with Pentium 4 EE 3.46 GHz gain just ~22%. There is almost no difference between Intel systems.
Drawing a bottom line under our today's tests, we want to mention a number of observations:
Sergei Pikalov (firstname.lastname@example.org)
March 20, 2006
Write a comment below. No registration needed!