Crossfire x8 + x8 versus x16 + x16
The X38 chipset allows two graphics cards running in a dual-card configuration to each use an x16 connection. Since this Intel chipset only supports a dual-card setup consisting of ATI cards, our testing is limited to the fastest of ATI’s current crop of cards.
Also, we limited our benchmarks to the resolutions at which Crossfire configurations showed pronounced performance gains on our old 975BX2 motherboard using a 2x8 configuration. In all, our benchmark suite usually consists of 35 test resolutions, but we’re working with a reduced sample of only 13 here.
In Crossfire mode, both the Radeon HD3870 and the HD2900 XT show a performance improvement of 12 percent. The Radeon HD3850 benefits the most, gaining 15 percent in performance.
Now, before you go and denounce our previous tests as being unfair towards ATI, let’s put those results in perspective. The performance gains of between 12 and 15 percent only apply to 13 of 35 tests. Factoring this change into the overall result of all 35 resolutions, we are left with a performance boost of between 6 and 7.7 percent. Finally, we have to take a realistic look at where this performance boost actually manifests itself by looking at the individual results, which you can find in our benchmark charts in this article.
We saw the biggest performance increase of nearly 20 percent in Call of Duty 4 with a Radeon HD3850 Crossfire configuration On the other hand, Call of Duty 4 already ran silky smooth even on an x8 link. We can only really call the 1920 resolution with FSAA a real win on the HD3850, with the frame rate going from 53 to 63 fps. The rest of the resolutions already easily ran at more than 60 fps or 100 fps on all three dual-card configurations to begin with. Nonetheless, we did clock a new record. The Radeon HD2900 XT delivered 215.6 fps in Crossfire mode, where the Geforce 8800 GTS 512 reached “only” 176.7 fps on a 680i-based motherboard.
In Doom 3, performance increases by 5 percent. In other words, you get 110 instead of 105 fps. Half Life 2 Episode 2 saw a performance increase of 9 percent on the new HD38x0 cards, going from 64 to 71 fps or 75 to 80 fps. In World in Conflict, finally, the performance improvement was only measurable, though not tangible. Specifically the game ran 1 to 3 fps faster.
Overall, the change in platforms results in a speedup of only a few percent. Games that ran smoothly before run a little faster, while slower games such as Crysis or Flight Simulator X don’t yet benefit from Crossfire optimizations in the WHQL drivers. Currently, going from x8 to x16 for Crossfire really only offers a tangible payoff in higher resolutions. At the highest resolution of 1920 x 1200 pixels, only a single test enjoyed a really meaningful speedup as a result of the faster interface, going from 53 to 63 fps. The remaining resolutions were already running smoothly at more than 70 or 100 fps, so any improvements here only help the overall performance rating, not real-world playability.
all very nice but like every article before this one, you do not mention the causes of bandwidth uses. so, like all your previous article's here on toms it is absolutely useless for anything other than saying"ooh, look at the fancy graph, aint they pretty" because apart from graphs the article has damn all to do with testing anything.
I found it useful, I now know I can save money by not upgrading just yet. That's more beer money in my pocket, that has to be useful.
What I have, serves my needs. Besides, these articles tend to be aimed at the games-playing mentally-preadolescent set, so they can't get too in-depth.
Erm.. am I missing something?
If one card increases by about 7% going from x8 to x16, then each card in a crossfire setup would increase by about 7% going from 2x8 to 2x16. Therefore the overall performance would improve by 7%. Why would you expect it to increase by more ("twice as much, which we would have expected based on the single card results")? If anything it should increase by less, due to increased loads on the chipset/system memory/processor etc.