Sign in with
Sign up | Sign in
Your question

PCI Express Scaling Analysis

Last response: in Graphics & Displays
Share
March 27, 2007 12:11:37 PM

What is the real importance of PCI Express for graphics? See how ATI Radeon X1900 and Nvidia GeForce 8800 high-end graphics cards perform at reduced PCIe link speeds.
March 27, 2007 12:29:16 PM

For some reason, most pictures refuse to show up. I am using Internet Explorer 7.
March 27, 2007 1:24:11 PM

All the pictures worked for me until I got to the benchmark results... pages 8, 9, and 10.

-TyShoe
Related resources
March 27, 2007 1:33:19 PM

Quote:
For some reason, most pictures refuse to show up. I am using Internet Explorer 7.

same here. Pretty much all the benchmark pics ain't showing up. Boooourns
March 27, 2007 1:53:01 PM

Works on firefox, no go on IE7
March 27, 2007 2:04:07 PM

Quote:
Works on firefox, no go on IE7

Damn, I'm only allowed to use IE7 on my work computer
March 27, 2007 2:06:20 PM

I think that the absence of dual-card solution running is x8 mode is very disappointing. There are no gamers in the world that are planning to tape their PCI-E cards to run at fewer lanes; on the other hands, thousands of gamers have dual-X8 solutions in use right now! Also, I'm forever hearing forum threads about, "will such and such a motherboard run OK with only two X8 lanes, or should I pay extra for a board with dual X16?"
March 27, 2007 2:17:30 PM

Blargh. *kills IE*

I saw one benchmark... :( 
March 27, 2007 2:19:18 PM

must be one of those firefox myths :lol:  :p 
March 27, 2007 2:27:27 PM

Quote:
I think that the absence of dual-card solution running is x8 mode is very disappointing. There are no gamers in the world that are planning to tape their PCI-E cards to run at fewer lanes; on the other hands, thousands of gamers have dual-X8 solutions in use right now! Also, I'm forever hearing forum threads about, "will such and such a motherboard run OK with only two X8 lanes, or should I pay extra for a board with dual X16?"


I agree completely. Should have at least thrown in a dual-x8 vs. dual-x16 benchmark. That's really the only item of importance.
March 27, 2007 2:42:40 PM

I am a bit confused as to why the conclusion is written from an application standpoint as in "application x benefits from x16 operation". Isn't that kind of a moot point? pretty much any application would benefit from more speed. Shouldn't the conclusion be written from the standpoint of the cards that are being benched, as in "with newer cards, its clear that they are now able to saturate 88 lanes, and will benefit from x16 lane operation".

I also agree on the lack of dual x8 lane testing. That pretty much defeats the purpose of this whole test, since there are few (if any) single slot x8 solutions out there and nobody is going to restrict themselves to x8 just for fun.

All in all, not the finest work I have seen on Tom's. Good start though, just needs some polish to provide us with what we really want.
March 27, 2007 3:14:42 PM

Yeah, I hate it when an article about apples fails to talk about oranges. It's really dissapointing. :oops: 
March 27, 2007 3:20:14 PM

Quote:
I think that the absence of dual-card solution running is x8 mode is very disappointing. There are no gamers in the world that are planning to tape their PCI-E cards to run at fewer lanes; on the other hands, thousands of gamers have dual-X8 solutions in use right now! Also, I'm forever hearing forum threads about, "will such and such a motherboard run OK with only two X8 lanes, or should I pay extra for a board with dual X16?"


I agree completely. Should have at least thrown in a dual-x8 vs. dual-x16 benchmark. That's really the only item of importance.

Indeed. I was considering some nforce 650 board. I doubt i will ever go SLI, but if still would like to know about the performance gain or loss due to SLI and insufficient PCIe links. Apart from that i really liked the article. It pretty much sums up what has been discussed here and there on the forums for months.
March 27, 2007 3:33:38 PM

The difference between nVidia and Ati here is most likely mainly a difference between 320MB and 512MB of memory. If you can't fit everything in video memory the card will start swapping textures to system ram, and it is here the bus bandwidth is important. So if you plan on going x8 SLI, make sure the cards have plenty of memory.
March 27, 2007 3:36:02 PM

I'm running a 8800 GTS on an AsRock Mobo that is only has a x4 PCI-express. But my lcd has a 1280x1024 resolution.

I wonder if is has the same importance at that resolution? It would also have been nice to see what impact it had on minimum/maximum fps compared to the average. If the drop in the average fps is only due to a reduction in the maximum fps, then you might not notice that big of an impact.
March 27, 2007 3:38:27 PM

I was really looking forward for the performance results for dual x8 for the 8800gts.

I suppose you can draw the conclusion that x16 would be better, but x8 is still adaquate with little performance loss. I doubt I will get my AW9D to do sli anyway (despite those strange rumors of 'hacked' BIOS)

And I havent bought the 8800gts yet anyway; waiting out on ATIs offerings.
March 27, 2007 3:58:39 PM

Quote:
I think that the absence of dual-card solution running is x8 mode is very disappointing. There are no gamers in the world that are planning to tape their PCI-E cards to run at fewer lanes; on the other hands, thousands of gamers have dual-X8 solutions in use right now! Also, I'm forever hearing forum threads about, "will such and such a motherboard run OK with only two X8 lanes, or should I pay extra for a board with dual X16?"


Word.
March 27, 2007 4:11:37 PM

I have to agree with the guys posting before me, there should have been a comparison of a dual card solution running at each speed.

However I'm happy you at least did the test as it shows x16 is more than just a marketing hype.
March 27, 2007 4:25:40 PM

Well not read it yet, but hmmm let me wage a guess.

It barely makes any difference to fps what mode you use.
But it will make a difference to load times.

Consider this.
The most amount of memory on the best vid card at the moment is 768 (lets not be picky with sli stuff.)
The idea of a graphics card is to take strain off other components.

So why would you combard a bus with 8GB/s of data?
Where is the card going to put all the info you are sending it?

When playing games the bus should pretty much only be sending instructions. Not huge textures etc. They should be loaded at map start or w/e.
That's what the cards ram is for. So how on earth would you use 8GB/s i don't know.
March 27, 2007 4:28:14 PM

Have to agree that more testing configs would have been nice, but good start anyway. Sort of a shame to see x16 already starting to be limiting. Wonder how super high rez massive texture games will turn out later this year with lower mem cards running on x8 lanes. As mentioned card memory is an important factor and anyone going SLI/Crossfire is going to want lots of memory. Also wonder how this will affect GPGPU type acceleration. Think I read somewhere the first bearlake boards(i.e. intel G35? chipset) not having dual x16 but later (P35?) with the having it also with PCIe2.0
March 27, 2007 4:30:04 PM

Another vote for a Crossfire/SLI test, here. No real person would ever run a single card at less than x16 PCIe anyway, so testing that is pointless. But the difference between x8/x8 and x16/x16 (if there is one) is actually quite pertinent. Certainly a comparison of x16/x16 SLI withy x8/x8 is highly relevant.

In fact, there are one or two P965 boards that offer x16/x4 Crossfire, I think; that would also be a good combination to test, see how it compates to x8/x8 on 975X (or RD600).

Granted, there's no easy of testing x16/x16 Crossfire. (Hmph).
March 27, 2007 4:37:39 PM

This guy sucks as a reviewer... Check other threads from his reviews.. he manages to screw up about every test he puts out there.
March 27, 2007 5:14:52 PM

Just stop trying.

Read the darn article before posting. Your conclusions are incorrect, which isn't surprising.

That 8GB/s is full duplex, so in your theory you should only have to use 4GB/s... :?
March 27, 2007 5:18:41 PM

man, those bandwidth hungry cards, i wonder how much the x2900xtx will suck up

hopefully pci express doesnt die so fast like agp did

agp kept changing its sizes like every year
March 27, 2007 5:32:50 PM

Quote:
man, those bandwidth hungry cards, i wonder how much the x2900xtx will suck up

hopefully pci express doesnt die so fast like agp did

agp kept changing its sizes like every year


Oh really? Please provide examples. AGP 8x hasn't changed for years.

And yeah, PCIe will be phased out for PCIe 2.0 in 2008.
March 27, 2007 5:41:09 PM

Quote:
Just stop trying.

Read the darn article before posting. Your conclusions are incorrect, which isn't surprising.

That 8GB/s is full duplex, so in your theory you should only have to use 4GB/s... :?


Yep i know its full duplex, I meant to write 4.

I read it over now that the pics are showing. Quake 4 doesn't seem to care. Cod2 does. Would be interesting to see how newer engines perform.

The way i see it is the card should hold most the data, and each frame should only be sending small amounts of data such as vertices info.

Isn't this why we use accelerated hardware. I can't see why cod2 performs so badly, maybe they need to make greater use of the card's onboard ram. I'd hardly say the graphics in cod2 are better than q4.

Or maybe it's OGL vs DX?
March 27, 2007 6:05:06 PM

Hmmm...
this was not abot sli, it was about what is the importance of the amount of the lines. The article proves that the number of lines actually do count, in some aplications.
Why to thest this in anyway? In the beginnig od pci era, there was not much difference between agp and pci. Now it seems to be so that some newer aplication prefer vider bus.
March 27, 2007 6:13:15 PM

Quote:
Hmmm...
this was not abot sli, it was about what is the importance of the amount of the lines. The article proves that the number of lines actually do count, in some aplications.
Why to thest this in anyway? In the beginnig od pci era, there was not much difference between agp and pci. Now it seems to be so that some newer aplication prefer vider bus.


Well, it actually in a sense IS about SLI, since the only time that 99% of people would be worrying about using a smaller lane would be in an SLI situation. That is why we are upset that there was no dual testing done. This whole article becomes pretty much synthetic with very little real world applicability when you don't compare the affects of having two cards together.
a b U Graphics card
March 27, 2007 6:40:48 PM

Quote:
Well not read it yet, but hmmm let me wage a guess.


Why don't you just read it instead of all the theoretical typing which is invalidated by the results? Probably a better idea.

Quote:
The most amount of memory on the best vid card at the moment is 768 (lets not be picky with sli stuff.)


Actually it's 1.5GB, and who'd be picky about SLi? It doesn't add memory, it copies it for most apps. Sounds like another review you guessed about instead of reading.
March 27, 2007 6:45:46 PM

Quote:


Well, it actually in a sense IS about SLI, since the only time that 99% of people would be worrying about using a smaller lane would be in an SLI situation. That is why we are upset that there was no dual testing done. This whole article becomes pretty much synthetic with very little real world applicability when you don't compare the affects of having two cards together.


In that you are right. This has very little real life value. It's more viable to academic debate... I like it somewhat... It has to be my old age and academic education ;-)

There has been guite a lot test with 8 + 8 and 16 + 16 sli graphics at several web sites. So that would be nothing new. Most of them have proven that there is very small difference.
But you are guite right. This test have very little real life value.
a b U Graphics card
March 27, 2007 6:53:28 PM

Quote:
Another vote for a Crossfire/SLI test, here. No real person would ever run a single card at less than x16 PCIe anyway, so testing that is pointless. But the difference between x8/x8 and x16/x16 (if there is one) is actually quite pertinent. Certainly a comparison of x16/x16 SLI withy x8/x8 is highly relevant.


Why?
I'm insterested in single card multi-lane configurations, not SLI, so to me it's pointless and not relevant.
The review answered a far more more common question adn far more common misstatement than whether 8X+8X is as good as 16X+16X, of whicih there are already a ton of MoBo reviews to answer that banality. The difference is so small as to be similar to the difference between chipset A vs B of the same lane configuration.
March 27, 2007 6:58:58 PM

Quote:
Well not read it yet, but hmmm let me wage a guess.


Why don't you just read it instead of all the theoretical typing which is invalidated by the results? Probably a better idea.

Quote:
The most amount of memory on the best vid card at the moment is 768 (lets not be picky with sli stuff.)


Actually it's 1.5GB, and who'd be picky about SLi? It doesn't add memory, it copies it for most apps. Sounds like another review you guessed about instead of reading.

What card has 1.5GB? Mainstream of course.

Take a look at Q4. Shows that there is plenty of bandwidth in lower speeds.
a b U Graphics card
March 27, 2007 7:08:27 PM

This isn't completely directed at you hergieburbur you were just the last to comment in this in a similar way.

Quote:

Well, it actually in a sense IS about SLI, since the only time that 99% of people would be worrying about using a smaller lane would be in an SLI situation. That is why we are upset that there was no dual testing done. This whole article becomes pretty much synthetic with very little real world applicability when you don't compare the affects of having two cards together.


Actually it's not, it's seems pretty clear what they are looking at (other than the lack of pictures), and perhaps if more of you took your heads out of your collective SLi/Xfired A$$E$ you'd figure out there are other areas that this specific test becomes extremely relevant to, like external graphics solutions, perhaps you'd stop whinning about your own already tested scenarios. I've been waiting for a detailed review like this (with better pictures though) for a while, which shows pretty much anything plugged into the ASUS external solution will suck compared to anything on the LASSO solution. There's already more than enough 8+8 and 16+16 reviews out there from when nV brought out SLi32 and ATi brought out X3200.

And if you SLi/Xfire fans though about the future of VPU physics you'd see how this applies very well to a 2 card configuration on the older MoBos for 8+8 configuration. SLi/Xfire numbers wouldn't matter to that.

While I don't disagree that it would be nice to add multi-card numbers to the review, their lack doesn't negate it's utility, just reduces it's usefulness to you and your particular situation.

Perhaps if more people looked at reviews for the information they DO provide instead of asking why their personal little world wasn't addressed maybe you'd all get more out of them and understand more than Bungholiomarks. :roll:

Next time they run a non-laptop review remind me to come in and comment on how irrelevant it is to the growing trend of PC users! :evil: 
March 27, 2007 7:21:26 PM

OK Then. I done been pwned.

For the record, I don't personally care about either SLI/Crossfire. The only use I have for my crossfire board is for two discrete cards, and the x8/x16 debate won't affect my needs in the least.

I admit I just skimmed the article and didn't really pick up on some of the other relevant situations. My Bad there.
a b U Graphics card
March 27, 2007 7:35:27 PM

Quote:

What card has 1.5GB? Mainstream of course.


Why 'mainstream of course'? You said BEST, not mainstream, whcih of course the GTX isn't either. You also didn't qualify what it is the BEST at. And for the SpecViewPerf tests, that would be the new Quadro, followed by the old 1GB Quadro, and 1GB FireGLs.

And as for the SLi part, yeah best we forget that. :wink:

Quote:
Take a look at Q4. Shows that there is plenty of bandwidth in lower speeds.


Yes, but it also shows that more is still better.
And until you do more investigating you can't tell if it's the benefit of no texture compression versus load/swap that's the trade-off.

Once again, guesses on your part.
a b U Graphics card
March 27, 2007 7:38:32 PM

Quote:
OK Then. I done been pwned.


That wasn't my goal, and I tried to make it clear it's not directed at you, just frustrated at the others really, you just were the only one I saw on page 2. Sorry to focus my reply on you.

And for me this is a great revelation that the ASUS solution with 1X ExpressCard connection will be extremely hampered compared to the 8 - 16X Lasso solutions.
March 27, 2007 7:45:13 PM

I know, but you are right. That was my way of admitting my mistake.
March 27, 2007 7:56:09 PM

Quote:
even though i use the 8x 8 method i ain't that interested in it. as i said in my earlier post i would have just liked them to show something? of course i haven't been able to see the benchies but either way did i miss the explanation of the results? if i did can someone show me?

what did this article actually show us?

when people ask if 16x really amtters and why, i cannot link them to this article as it does not tell us anything as far as i can see.

what exactly is the purpose of this article? it is not clear to me that is for sure.


An attempt to finally show that PCIe 16x really matters years after it was introduced. :wink:
a b U Graphics card
March 27, 2007 8:02:15 PM

I agree with you, more information would be better. But to do saturation tests you need more synthetic tests like the Serious Magic and others that are only meant to stress the bandwidth of the slot. Unfortunately they can tell you that say the PCIe HSI bridged verion of an AGP chip in the old days has theoretically less bandwidth, but the practical implications were low.

The tests here are trying to show the impact of the game performance, where it's about not just the bandwidth but the actual use of that available bandwidth.

I don't think the review showed us alot of what we didn't already expect, but some of the 4X results were pretty interesting considering the assumptions of 8XAGP which would have the two way performance of 4X PCIe.

I don't think the article will show why 16X matters so much as why 1-4X may be limiting. I think for what you're looking for more tests are needed, and if they were to add them, then I'd love to see the memory size impact alongside the throughput. But I'll take the 1-8x info for the external and Physics applications.
March 27, 2007 8:08:58 PM

Quote:
I'm running a 8800 GTS on an AsRock Mobo that is only has a x4 PCI-express. But my lcd has a 1280x1024 resolution.

I wonder if is has the same importance at that resolution? It would also have been nice to see what impact it had on minimum/maximum fps compared to the average. If the drop in the average fps is only due to a reduction in the maximum fps, then you might not notice that big of an impact.



UHHHH wich asrock mb are you using? the dual vsta 775llg with both pci-e and agp??
March 27, 2007 8:14:01 PM

I heard tomorrow Patrick Schimd will be using an experimental beta motherboard with a next generation card that utilizes an x32 interface.

I hope he can confirm that x32 is indeed faster than x8. I've been wondering for a while if that's how it works.
March 27, 2007 8:46:23 PM

Quote:
This guy sucks as a reviewer... Check other threads from his reviews.. he manages to screw up about every test he puts out there.


true, but i thought this one was a lot better though. Onlything missing really is a Dual x8 vs Dual x16.
a b U Graphics card
March 27, 2007 8:49:54 PM

Yeah I agree, especially on letting Cleeve do more.

I think that in conjunction with the high end and investigation into the low end would also be useful, because I doubt that having Ultra vs High textures in the X1300/GF7300 range would matter as much about the 8x-16X as it does in the higher end because it would be pointless really due to lack of core strength.

Memory size differences would be more inteesting also for big extremes around the game limits.

The Bandiwdth is handy for textures swaps, and the memory size is handy for not haing to uncompress textures, but the impact f these can be minimal in some games/situations. Also the settings in game allow for flexability, and while the Ultra vs High setting in Q4 and D3 are technically discernable, I don't know any reviewer who has said that they are all that noticeable, and so for cards like that likely not an issue compared to lower resolution, and AA, etc.
a b U Graphics card
March 27, 2007 10:29:03 PM

Quote:
you know i think i figured out this article. this is clearly only about what happens when you limit the amount of PCI-e lanes. clearly the how and more importantly the why does not matter to the authors.


Yes, that did seem to be the intention in their intro, it's no so much as why do we need 16X, but looking at some titles, what is the impact of less leanes. It's more theoretical right now than anything, like our old AGP 4X vs 8X and 128MB vs 256MB discussions, but it does have appplication beyond just the pure investigative realm. I think the portion you're looking for would be the follow-up. Now that we've seen that there is a difference... why is there a difference. And then investigate that. There are lots of possible reasons for the differences, including the aforementioned texture swap-outs, etc. However I don't think that was the goal yet, but a good idea for a follow-up. However you would need to get into some DEV tools IMO to truely muck around in such a way as to increase things separately. It would be like doing an amperage test to test PSUs just to see how they handled the loads, not why they have to handle such loads.

Quote:
now, the higher the tick rate the higher the bandwidth usage obviously, too much info for the server to handle and server based lag kicks in, the same as the CPU.


Yes, and that would be the case if the entire game could be stored in video memory like some machine language games/demos can be done. They don't require swapping out much information. And it would be the same with games if you stayed in a single space and only moved within that sace held in memory (like rotating 306 degrees) you will still have visual stress on the card, but little need to communicate across the PCIe bus other than to get point in space references. So the only thing that the memory would do is back buffering and such for AA calculations, etc.

Quote:
my question is that if instead of the games used you threw in some more CPU intensive ones(take your pick) then what happens to the bandwidth.


Likely very little. It's not used in the same way you're thinking (at least what I get from your descriptions). The CPU overhead in most games is usually not the small graphics related part handled by the CPU (unless not supported or offloaded by the VPU) most of the CPU intensive parts involve other portions of the game like mapping, AI, audio, etc.

Quote:
am i missing something here or was there no point to this article expect to show some benchmarks which have been done before except now we have a better CPU?


Well it would likely be CPU independant to some extent when you go down to the 1X and 4X scenarios. CPU and HD speeds are involved of course, of course going way down it's going to slow things down for all the other parts handled by the CPU. It's a question of the multipier effect of all these bottlenecks.
March 27, 2007 11:07:13 PM

Quote:
The difference between nVidia and Ati here is most likely mainly a difference between 320MB and 512MB of memory. If you can't fit everything in video memory the card will start swapping textures to system ram, and it is here the bus bandwidth is important. So if you plan on going x8 SLI, make sure the cards have plenty of memory.


That could explain a lot here. I would like to see the 640mb added to the test as well as a couple 256mb cards. Good observation Zorlius.

When I saw how little difference there was between 8x and 16x for the x1900 I wondered if that kind of scaling would be similar on my x1600pro (which, incidently, I've been thinking about moving to an 8x slot due to layout constraints in my computer). But it's only 256mb.
March 28, 2007 1:09:29 AM

on a single card the max memory at the moment is 1gb so your both wrong Leadtek QuadroFX 5500 is all i can say.
March 28, 2007 1:53:24 AM

I would really enjoy reviewing the benchmark data but the only benchmark image that I can view is image026.png. All others fail...

I am using IE7, fully "up-to-date".
March 28, 2007 2:05:20 AM

Try firefox, works fine.
a b U Graphics card
March 28, 2007 2:49:43 AM

Man I just want pschmid to post a 3 page flame against everyone in the forums. About the fact that no one fully appreciates his time and effort and stuff. I think that after the amount of abuse his reviews get here, he deserves to give some back.

You can do it!
March 28, 2007 4:01:04 AM

Quote:
Man I just want pschmid to post a 3 page flame against everyone in the forums. About the fact that no one fully appreciates his time and effort and stuff. I think that after the amount of abuse his reviews get here, he deserves to give some back.

You can do it!


I just read a quote today about how you should never answer your critics unless they happen to be right...
!