SLI & CFX PCIe Bandwidth Perf. - x16/x16 vs. x16/x8 @ [H]

FrgMstr

FrgMstr

Just Plain Mean
Staff member
Joined
May 18, 1997
Messages
55,601
SLI & CFX PCIe Bandwidth Perf. - x16/x16 vs. x16/x8 - Have you wanted to space your SLI or CFX video cards farther apart on your motherboard to allow for better airflow? Do you have a motherboard that will not support a x16/x16 PCIe configuration? We put x16/x16 and x16/x8 PCIe SLI and CFX configurations head to head and show you what sacrifice there is to be made.
 
May you run triSLI and triFire next? Or is that just me asking for too much? :p

I like how, even at insane res, very little changes from 16/16 to 16/8 :)
 
Great job as always.
probably going to run 16x/8x on my P6t now simply because it spaces out nicer.
 
jeremyshaw said:
May you run triSLI and triFire next? Or is that just me asking for too much? :p

I like how, even at insane res, very little changes from 16/16 to 16/8 :)
Click to expand...

If you have the cash to buy 3-way SLI you should get the "right" mobo. But I would suggest that our next set of tests will really find that out anyway doing x16/x16 and x8/x8.
 
This error is fixed. Look for the follow-up this week. - Kyle
 
Last edited by a moderator:
Thank you for answering this age old (relative to tech) question.

Does this mean PCI Express 3.0 is just marketing bullshit? It doesn't sound like anything can come close to even half the speed of PCI Express 2.0.
 
IRSmurf said:
Thank you for answering this age old (relative to tech) question.

Does this mean PCI Express 3.0 is just marketing bullshit? It doesn't sound like anything can come close to even half the speed of PCI Express 2.0.
Click to expand...

I don't see video cards taking advantage of PCIe 3.0 for a good while, but there are a lot of solid uses for PCIe 2.0 in enterprise systems. PCIe 3.0 is a bit different from 2.0 in the way data is passed, but it is being projected that it will be 2X the bandwidth of 2.0. Don't see much use for 3.0 on the desktop for a good while in reality.
 
I would love to see some 3x cfx action too. It seems to work much better with 10.7a drivers. nobody reviews 3x cfx 5770s which at 150 are a decent price. It was easier for me to spend 150 in 3 shots than buy a 5870, and 2 5770 seem to be close in perf to a 5870.
 
Im soo happy Kyle and the crew did this. Im opening my case tomorrow and spacing out my sli gtx 480 sc + setup. WOW. This is great news.
 
nice, already seen this several times from different sites but nice to see an authoritative H articale showing it. 8x has been plenty for a long time now and likely for the foreseeable future.
 
Not surprised. From what I've seen on other sites even at x8 on both, or even x4, the drop is minor enough that you'd have to benchmark to notice it.
 
I asked a question related to this article on these forums ages ago and no one could help me. Since this article deals with the same thing i'll ask it here. My mobo supports CF but not SLI. It is also PCI-E 1.x only not 2.0. So my slot setup for CF is 16x/8x but translated to 2.0 numbers that's only 8x/4x. Am i going to get hurt by bandwidth by doing that? Or should i just wait until i revamp my entire system and get a new mobo?
 
thanx so much for this comparison. I rely on your articles rather than other fan boy sites.
 
This may even open up the options on motherboards a little for those whom had previously required a greater # of x16 slots now that we see that we can run a card in an x8 slot with only a very slight performance hit. And that being a hit so small that 99% of us wouldn't be able to see it visually at anything less than insane eyefinity resolutions.

That 460 SLI setup is becoming more and more tempting.
 
Very nice. Always enjoy taking a look at these x16 vs x8 reviews. Looking forward to next weeks x16/x16 vs x8/x8 results.
 
Damn.
I might really have to consider buying a second gtx470 as a viable upgrade path now...
Also, is the 5970 considered xfire? Might be worth adding it to your tests considering it cost about as much as the setups you are testing.
 
Doesn't this also tend to help prove that current video cards/slots don't fully saturate or use up all the available bandwidth they can?
 
After the first couple of games it becomes evident that there's no difference between 16x/16x and 16/8x. I'd probably stopped right there, but you did the extra mile and tested extreme resolutions in the cards that most of us would pick.

So either you have lots of spare time or you like to back your findings with [H]ard evidence.
 
Arg! 460SLI is besting the 5870 CFX...starting to feel like buying two 5850s was a waste.
 
Good lord the difference in scaling (SLi vs XFire) really stick out in this review. Also it is interesting to see that in the x16/x8 configurations the minimum and average FPS was one to two FPS higher than the x16/x16; while the max FPS was one to three FPS lower.
 
this is good news in the fact that we can now space our cards finally, but bad news in knowing we arnt really utilizing the full bandwidth capabilities yet.
 
It's not bad news. Real bad news would be knowing we ARE utilizing all the bandwidth and that version 3 wasn't going to be out for a while.
 
Thanks guys for putting in the hardwork for this review, but I'm not surprised in the least by the results. Definitely looking more forward to the 8x8 review.
 
I'm also curious on how the 5970 would perform on x16 compared to x8. Or do a quick test on the Asus Ares if you still have it.
 
I find these types of articles and comparisons very interesting and informative. These are the kinds of things I read, that while seemingly simple, really add a lot to my overall computer knowledge, and often come into play when making buying decisions. Thank you Brent and Kyle.

Edit: Really looking forward to the 8x/8x vs 16x/16x article.
 
2 x 5970 comparison would be interesting, albeit for a small number of readers; but if you have the inclination and equipment available showing the effect of PCIE bandwidth on two 5970 cards in crossfire would be great.

Thanks.
 
Well... i guess that answers the question of whether I will go P55 or X58. :D

Then again... maybe I will wait for P6x or X6x and see what they have to offer. ;)
 
The reason discrete graphics cards come standard with their own onboard pools of local ram, from 512mb to 2048mb, is because the gpu can texture much, much faster from its local ram than it can across the much slower PCIex8/x16 system bus. Ideally, it shouldn't matter whether the system bus is x8 or x16, as the gpu should always be texturing out of its local ram.

Depending on gpu driver and hardware capabilities, though, and depending on how the game software itself wants to operate--this isn't always the case. For instance, if a GPU driver is engineered to more-than-normally rely on cpu performance then exceptions to the rule might be more noticeable and plentiful.

If a more thorough test between x16 and x8 overall frame rate performance is desired, I would recommend testing with nothing except 512mb cards. In this case the gpu will be forced to texture out of system ram much more often than from local ram (since there won't be enough local ram), and then one can more accurately gauge the performance differences between x8 and x16. The more local ram available to a gpu, however, the less likely the speed of the system bus will matter.
 
Only thing that might use PCIE 3.0 is SSDs. For that I say bring it on!

Thanks for the info on video card performance though.
 
Synthetic benchmarks take the greatest toll when running at x8 (or lower).

Other than that, who cares.
 
Well, I've learned over the years that you can refund yourself the cost of hardware up to 75% sometimes even higher.

You just have to watch the enthusiast sites, the tech news places, and forums to formulate a window in which to sale.

Right now, the window is open for the 5870. You can very easily get $325 - $350 for a used card today and into the next few months before this product line refreshes or they release their new 6xxx models. After that happens, prices will mostly likely drop to $225 - $250 for these cards. From what I have read, there are millions out there.

I was able to sell one of my 5870's and move over to a pair of 460's for around $100 out of pocket. I suggest others think of doing the same to save money and maintain performance.

I ordered two of the ASUS ENGTX460 DirectCU/2DI/1GD5 GeForce GTX 460 (Fermi) 1GB 256-bit GDDR5 cards from newegg.

http://www.newegg.com/Product/Product.aspx?Item=N82E16814121391

These cards just need a simple bios flash to turn them into the more expensive TOP models. 775 / 1400 - 1500?
 
Last edited:
Looks like this will solidify my decision to go SLI with a a pair of SC 480s from EVGA. Thanks a ton Kyle! :)
 
SolidBladez said:
Synthetic benchmarks take the greatest toll when running at x8 (or lower).
Click to expand...

No doubt about this because you are getting a flood of data that is being pushed as fast as possible with synthetics/canned timedemos. I have seen the canned benchmark x16 vs x8 comparisons and I even talked to NVIDIA about this before we started the research for the article and all NV had was canned timedemo benchmarks, that is why I thought it was important for Brent and I to put our 2 cents in on real world gaming. Our data shows even less of a difference than synthetics as we thought it would. That made me interested even more in doing X16/x16 vs. x8/x8.
 
I was always searching for the answers to this question. I am running two XFX 5870 XXX's on a Gigabyte 790FX board using the first and second slots 16x/16x and my first card wasn't able to get any proper air flow because the slots were too close together so I was having to manually turn the fan up on the cards to keep them cool.

After reading this I have now spaced my cards apart running 16x/8x and I don't notice any difference with any game that I play 1920x1200 and the cards are running max 60 degrees default fan speeds.

Thanks for the info!
 
Nice work and good job well written. and im sure this helped a bunch of people now.. Keep it up [H]ard.
 
You must log in or register to reply here.
Back
Top