PCIe And CrossFire Scaling: Does Nvidia's NF200 Fix P55?
Table of contents
- 1. When Does 16/2=32?
- 2. Test Settings And Boards
- 3. Benchmark Results: Crysis
- 4. Benchmark Results: Far Cry 2
- 5. Benchmark Results: Tom Clancy’s H.A.W.X.
Our recent analysis of three-card CrossFireX on P55-based motherboards demonstrated that the bandwidth restrictions of the third slot’s first-gen PCI Express (PCIe) x4 interface was the most likely cause of dramatic performance drops in some games, compared to an otherwise-identical dual-GPU configuration. The LGA 1156 interface’s native support for 16 PCIe 2.0 lanes (running at 5 Gb/s) across a maximum of two devices leaves the P55 chipset’s slow 2.5Gb pathways as the only means for hosting additional devices. Many of our forum members would label the attempts made by several manufacturers to add a third x16-length slot to LGA 1156 platforms as failures of epic proportions.
But what if the third slot didn’t have to rely on the platform controller hub to facilitate lackluster throughput? Certainly some other type of data hub could be devised to allow the processor’s 16 lanes to be divided three ways, with the equivalent bandwidth of five 5.0Gb lanes feeding each card, right? Alright, so the cards can't really operate in x5 mode. But surely an add-in device, if smart enough, could spread that bandwidth across eight lanes. Because the Lynnfield-based Core i7/i5's PCIe controller is only able to host two devices, the add-in part would have to present itself to the CPU as a single component, negotiating data traffic to any connected graphics cards using its own logic.
Fortunately these types of devices, called PCIe bridges, already exist. Nvidia has long used its nForce 200 bridge to “transform” the PCIe 1.1-equipped 680i SLI chipset into the PCIe 2.0-equipped 780i. While many of the 780i’s critics pointed to the northbridge as a bottleneck, the NF200 proved itself extremely capable of managing the bandwidth disparity.
And yet our previous CrossFire examination began with a PCIe performance analysis that showed a 4% slow-down between x16 and x8 PCIe 2.0 slots. Wouldn’t dropping to the equivalent of five lanes of bandwidth per card result in an even more dramatic performance decrease? Two motherboards that arrived for our recent Extreme Motherboard Shootout gave us the opportunity to find out.
Score One for the X58. The NF200 should help the P55 however but at an added cost. ALl in all interesting to see that Nvidia has been busy making this for the onboard and motherboard vendors instead of focusing on the GF100. I am sure they are really strategizing and letting AMD/ATi get the upper hand for a short time as their target market for business is also reaching a wider market in motherboard vendors. Then when GF100 comes out, ppl will see the type of quality GPUs that are meant to please the eye and take on the best and most demanding apps/games/HD content.
With dual CrossFire/SLI (where it really matters most) it is kinda hard to fix something that ain't broken.
Nice gains with three way though.
Pretty sure if you have $1200 to dump on GPU's you would have gotten the proper platform in the first place...
MSI wins again.
well there you go, high end x58/1366 systems with a high end price offer the higher performance, and pushing a mid-range 1156 platform to high end performance requires extra exotic equipement/chips to do it - no new news here
With dual CrossFire/SLI (where it really matters most) it is kinda hard to fix something that ain't broken.Nice gains with three way though.Pretty sure if you have $1200 to dump on GPU's you would have gotten the proper platform in the first place...
Exactly my point +1000
Testing should have been done with 5670's
Just Give us the Dman FERMI already
It proves again that the x58 is stock the best performer albeit not very much in comparison with the buffed up p55 gf200 chipset. But in practical mode, I don't see the big wins with triple vs. dual gpu's. Moreover I probably won't even go dual videocards because of the relative high fps which don't matter much, 60+ is more than enough for any game on a flatscreen monitor. 30+ is mostly enough for most single player games.
It DOES however prove that the chip gives a reasonable gain if games are going to demand more then they do now. The only games available that can really max out anything are crysis, WIC, total commander and that's pretty much all.
For the rest pretty good article though
I prefer to use the remaining 4x PCI-e 2.0 lanes of the X58 for a Fusion-io Duo drive. Can you imagine?
I bow to you Crashman. This was the granddaddy of all p55 reviews and I openly admit to standing corrected. I honestly didn't believe the NF200 could pull it off, but I was obviously wrong.
I wish this review had come a week earlier, I would have gone 1156 over 1366 but at least it finally showed up. The last vestiges of X58's hold over P55 have been erased in my mind.
Now you can move on to figuring out SATA 6Gb/s for us. I know, I know I'm never pleased. Great work.
...It DOES however prove that the chip gives a reasonable gain if games are going to demand more then they do now...
Don't be so quick with judgement. This may hold true for a single monitor, but many of us are already running 3x monitors in Eyefinity and even more will be in the near future. With 3x 1680x1050's a single 5870 is pushed well beyond it's capability to run many games even on medium settings.
I would have gladly bought an MSI Trinergy had I believed the NF200 would do what it claimed. I just didn't believe P55's x16 ondie pcie could pull it off.
who gets tri gpu cards, and not a X58 platform? this is like trying to build the eiffel tower, while leaving out the concrete base to hold it up....
who gets tri gpu cards, and not a X58 platform? this is like trying to build the eiffel tower, while leaving out the concrete base to hold it up....
Considering the performance difference, anyone could. There's no benefit in buying a X58 for gaming; just read the benchmarks.
If these motherboards where cheaper, buying a X58 for gaming would be a waste of money. But right now its price $100,00 over a cheap P6T X58, still give the big brother a fighting chance.
who gets tri gpu cards, and not a X58 platform? this is like trying to build the eiffel tower, while leaving out the concrete base to hold it up....
Also with other benifits - 6 core processors shortly, and you can pack in 50% more ram in a 1366 socket system thanks to an extra memory channel
hence overall its the same rule:
Performance = 1366
Mainstream = 1156
Decisive results, good article.
...but sadly, irrelevant to anyone (minus the .002% of PROFESSIONAL gamers) who has any sense of value and realizes that 3x high-end GPUS and the substation to run them are not a good use of money.
Well, for those who like to save a few bucks (it there's any save at all, lol) they can get a cheaper platform for tri-way-whatever they want now. May not get the features from it's big brother (X58), but at least they gave another option for a gamer to choose.
But yeah, hard to accept that when you have money for 3 video cards, you'll spend on a cheaper MoBo/Chipset for them.
Is the RAM/CPU/MoBo combo way cheaper on those platforms with the NF200 than on a X58 chipset? I think there's where the question should be headed, anyway. A Core i5 750 OCed, 8 Gigs with low timings and 3-way vids should do the trick for most of us, right? Even, a 3-way is kinda a lot under 1900x1200 =/
Cheers!
Also with other benifits - 6 core processors shortly, and you can pack in 50% more ram in a 1366 socket system thanks to an extra memory channelhence overall its the same rule
So now we're down to justifying 1366 over 1156 to only a 1k future cpu, or 1k in ram that only a handful of enteprise users would ever need? If you're thinking that you'll wait till Gulftown becomes affordable then you're looking at 18 months at which point you'll buy a new mobo to go along with it anyways.
If you haven't noticed, i7-750 edges the 920 in gaming, and the 860 edges the 920 in everything else. The last vestige the 1366 could claim performance in was tri/quad gpu solutions. That's now been erased as well.
Good prof of concept, but frankly still ridiculous; anyone who can afford a 25x16 monitor or 19x12 and 3 power gpus like a 5870 which need the bandwidth and an appropriate cpu to run all that likely wouldn't skimp on the mobo and chipset. imo.
Testing should have been done with 5670's
Probably tongue in cheek, but relevant; can the traces of these mobos handle three cards that don't use PCIE power connectors?