Sign in with
Sign up | Sign in
Your question

Crossfire meets PCI Express 2.0 - More Lanes, More Frames?

Last response: in Graphics & Displays
Share
January 23, 2008 12:56:37 PM

How much of a performance advantage does PCI Express 2.0 offer over PCIe 1.0a? How much performance do you lose on an x4 or x8 connector? How much faster is a Crossfire setup with two full x16 connectors? Read on for the answers!

http://www.tomshardware.com/2008/01/23/crossfire_meets_pci_express/index.html
January 23, 2008 1:59:21 PM

Nice project, thanks for doing that research. Also, nice work with discussing the error. It is a relevant variation in data that people overlook when making conclusions.
January 23, 2008 2:28:07 PM

Would have been nice if they didn't limit the testing to INTEL only boards. I'm interested in seeing how much of a performance increase if any I got going to my GA-MA790FX-DQ6 which also sports PCI Express 2.0 and dual x16 slots for Video cards which house my pair of 3870s.
Related resources
January 23, 2008 3:22:00 PM

Good conclusion for the interface vs. "real world" results... however I have issue with this paragraph on the last page:

Quote:
Another thing we saw is that a dual x16 configuration is not as effective for Crossfire as a single x16 slot is for a single card. As mentioned above, a single card loses between 7 and 8 percent performance when operating in an x8 slot. Moving from an x8+x8 setup to an x16+x16 connection also only yielded a 7.7 percent frame rate increase - not the doubling we might have expected based on the single card results.

b/c you are talking in percentages there, you cannot expect a "doubling" of percentage with two cards. If a single card gets "between 7 and 8 percent" and the dual cards show a 7.7 percent increase (also between 7 and 8) then you have a linear scale between the single and dual card config. This means that the x16 interface IS as effective for both single and dual cards.

You got a doubling of the increase in frames, but NOT the percentage. "Doubling" in this case means the percentage stays the same from single to dual cards but the raw increase is doubled. This is a 100% increase on raw numbers (doubled), but the increase on the difference percentage from one interface to the other should stay exactly the same.

Example: if 100 cows produce 50 pounds of crap eating grain, and 75 pounds eating grass (a 50% increase from 50 to 75) and 200 cows produce 100 and 150 pounds respectively (also a 50% increase) you can see that the increase from grain to grass is doubled when the cow numbers are doubled, but the percentage of that increase stays the same. This means that when the cows are doubled, they produces twice the amount of crap... but the increase of that crap from grain to grass is exactly the same. (best analogy I could come up w/ on short notice ;)  )

The conclusion that the x16 interface is not as effective for CF is false. It is a near perfect linear scale, which means it is as effective.

Now, in real world gameplay does it really matter? no, the raw numbers are not that high... so your final conclusion about not being a worthwhile upgrade is correct. So is the conclusion that the increase could be anything outside of the interface...
January 23, 2008 4:20:50 PM

wow, well spotted!
January 23, 2008 4:46:30 PM

One thing to keep in mind... when discussing PCIe 2.0 remember that the 5GT/s lane speed is optional. Vendors can label a board or system as V2.0 compliant and NOT design in the higher speed I/O.

Look closely at the specs for both the system and your add-in card.
a c 143 U Graphics card
January 23, 2008 5:03:02 PM

sojrner said:

Example: if 100 cows produce 50 pounds of crap eating grain, and 75 pounds eating grass (a 50% increase from 50 to 75) and 200 cows produce 100 and 150 pounds respectively (also a 50% increase) you can see that the increase from grain to grass is doubled when the cow numbers are doubled, but the percentage of that increase stays the same. This means that when the cows are doubled, they produces twice the amount of crap... but the increase of that crap from grain to grass is exactly the same. (best analogy I could come up w/ on short notice ;)  )


The CEOs of AMD and nVidia would like to thank you for that wonderful analogy. I'm sure by now they have already mailed you some nice video cards as a reward. You don't mind if they use it in their presentations, do you? :lol:  :lol: 

January 23, 2008 5:06:13 PM

rofl, nope... not at all. Whatever I can do to help and contribute to the betterment of society. ;) 
January 23, 2008 7:41:47 PM

One thing i would like to point out that didn't seem to get mentioned is that we don't need to wait for GeForce 9xxx or ATi 4xxx. They ahve it with the new X2 cards that are coming out. 2 GPUs on the same board will use double the bandwidth. Hence they need the 2.0 interface. This testing would be well to do again with the dual GPU cards once they hit the street. I bet in PCIe 1.0a the cards will loose a significant amount of performance over 2.0.
a c 143 U Graphics card
January 23, 2008 8:18:38 PM

Agreed, I want to see the HD 3870 X2 card tested on P35 and X38 ASAP.

Just guessing:
- the HD 3870 works at 92% of its normal speed if you put it in an x8 PCI-E 1.0a slot. This means it needs 50%*(100/92) of the bandwidth of a x16 PCI-E 1.0a. If the X2 needs twice the bandwidth of the HD 3870,
then it will need 100/92 of what the PCI-E 1.0a x16 slot offers, so it will also lose 8% compared to its best, i.e. compared to what it can do in a PCI-E 2.0 slot. But maybe the X2 doesn't need twice as much bandwidth as the HD 3870, and there won't be any penalty at all if you run it in a PCI-E 1.0a slot. I can hardly wait to see this actually measured...

Edit: or maybe the X2 needs three times the bandwidth of the HD 3870 because of some overhead, and it will stink big time in P35 motherboards. :) 
a b U Graphics card
January 23, 2008 8:50:14 PM

sojrner said:
Good conclusion for the interface vs. "real world" results... however I have issue with this paragraph on the last page:


Pffft, sure, if you use math, logic and reason !! :pfff: 

Math is so 2007, don't you know it's all about the size of numbers this year, not what they mean !?! [:wr2:4]

BTW, nice find, haven't had time to read the article yet, but now at least I can read it with the correct 'edit' already in my head.

Way to read the review(s) instead of just looking at the graphics/pictures. [:thegreatgrapeape:6]
January 23, 2008 9:32:56 PM

thanks grape... I could not resist that one when I read it. Blatant lack of logic when making a "logical" conclusion just bugs me. :p 

seems like today there is less and less true logic in the "news"... stats and reports showing "conclusively" that the environment is dead, westerners are evil violent turds and iranian fundamentalists are benign angels... I just can't tolerate it on a tech site where it is (or should be) run by effective geeks/nerds that should have math, logic and reason as their base foundation of knowledge.

...of course, maybe that's just me. ;) 
January 23, 2008 9:35:19 PM

My head hurts after reading that post but very good information.
January 24, 2008 3:42:08 AM

VERY Good article and should solve a lot of forum debates. Thanks!
January 24, 2008 9:30:47 AM

According to this AMD presentation:
http://download.amd.com/Corporate/SpiderPlatformPresent...
(Slide 20)

Going from PCIe1.1 to PCIe2.0 yields the following performance improvement:

Company of Heroes: 6% to 13%
Call of Juarez: 11% to 22%.
Lost Planet: 12%
World in Conflict: 6%

Maybe rechecking with Call of Juarez (apparently the best case scenario) would be nice.
January 24, 2008 10:15:33 AM

This article has been up for ages on the UK site. Its not like they had to translate it. Don't know it took so long to get it on the US site.
January 24, 2008 4:50:20 PM

Ok... so is anyone else rather irritated that they ONLY tested Intel boards rather than testing CF on the new AM2+ 790FX Chipset??? http://img.tomshardware.com/forum/uk/icones/message/ico...

That bothers me mainly for the reason that the 790GX chipset is custom-tailored for crossfire, while the x38 isn't.
Perhaps they should've done a comparison of % gains with similarly-clocked Phenoms and C2E's with CF and Tri-Sli respectively. That way they could get a full-compatibility comparison. Just a thought.
a c 143 U Graphics card
January 24, 2008 4:56:26 PM

I thought x38 was custom-tailored for Crossfire. It does offer dual x16 slots for it, better than P35. No clue what you mean by that. Is there anything that X38 doesn't have and it would make Crossfire better?

I totally agree that coverage of AMD could and should be better. It's always nice to have full info about both sides before you buy.
January 24, 2008 5:04:24 PM

Perhaps you're correct, I haven't fully reasearched the x38 chipset, I just know that 790fx is amd/ati so in my mind it's more appropriately suited for CF.

However, I'm rather astonished that in about half the benchmarks they didn't test CF of any sort... thanks for the useful info and comparison to SLI.
a c 143 U Graphics card
January 24, 2008 5:33:45 PM

Well, I've seen a 790fx mobo that allows 4 (FOUR) cards in Crossfire. Now that is serious stuff indeed :) 
January 25, 2008 2:26:12 AM

Those results depress me... and my pitiful CF system. 9 fps against 28 fps on WIC on 1920 with AA on... WTF.... this tells me that ATI has to make a driver for every game which is rediculous. the SLI systems get beat on 1 or 2 games, but scales to different games alot better.
January 25, 2008 4:27:39 PM

I HAVE a 790FX mobo that allows 4 cards in CF, here's how it works:

It has PCIe 2.0 which supports two full lanes at 16x each and divides them by two to get 4 8x lanes.
Basically dual original CF on one mobo. Still pretty impressive, albeit slightly impractical.

And don't be too depressed about the results, in case you didn't look too closely at the results, they did NOT test PCIe 2.0 CF on about HALF of the games.
February 7, 2008 5:06:53 PM

I do not see how this article answers the question of P35 Crossfire performance.
Sure, there is a test of x4 performance with a SINGLE card, but assuming that is indicative of how Crossfire would perform in a 16x/4x configuration seems incorrect due to one big fact: The Crossfire Bridge provides loads of bandwidth for communication between the two cards, so it's very possible the reduced bandwidth of the x4 slot wouldn't impact performance nearly as much when the bridge is in use as the x4 limitation does in a single card setup.
!