Interesting experiments with different PCI-E speeds

[mjtdevries]

Tom's did some experiments on the influence of PCI-E speed on graphics cards.

http://www20.graphics.tomshardware.com/graphic/20041122/pcie-analyse-01.html

While most results are as expected, there are some unexpected differences between the ATI and Nvidia cards with games.

The 6800GT has a 25% performance hit in games when the PCI-E speed is reduced to x1. But the ATI cards hardly suffer in performance.

I am wondering what causes this difference?

There shouldn't really be much difference in the amount of data that each card moves over the bus, should there?
Is this the difference between a native PCI-E solution, and bridge chips?

 

[mczak]

Tom's did some experiments on the influence of PCI-E speed on graphics cards.

An interesting article indeed.

While most results are as expected, there are some unexpected differences between the ATI and Nvidia cards with games.

The 6800GT has a 25% performance hit in games when the PCI-E speed is reduced to x1. But the ATI cards hardly suffer in performance.

Not only in games actually. You can see in specviewperf that nvidia loses more performance than ATI when reducing PCI-E lanes.

I am wondering what causes this difference?

There shouldn't really be much difference in the amount of data that each card moves over the bus, should there?

Well, the results seem to indicate that nvidia indeed moves more data over the bus. It's hard to tell why, one reason I can think of is that nvidia might use larger fixed size buffers (for more efficient handling) for vertices (since it's usually not important to optimize for bandwidth usage). That's just a shot in the dark though.

Is this the difference between a native PCI-E solution, and bridge chips?

I don't think this would make a difference since the bandwidth usage shouldn't change. Actually though, maybe that's not true, if the GPU is using things like DME that might cause more bandwidth to be used than straight PCI-E transfer?

 

[Nick]

Why cares? I'm not going to run my card at PCI-Express 1X speed. Much more interesting is that 8X doesn't hamper SLI performance.

 

[trinibwoy]

Why cares? I'm not going to run my card at PCI-Express 1X speed. Much more interesting is that 8X doesn't hamper SLI performance.

Still interesting to note and understand the differences between ATI and Nvidia's use of PCIe bandwidth. It might have an impact down the road on something you do care about.

 

[Dave Baumann]

Given that 8 lanes is ~ equivelent to AGP8X in downstream bandwidth, this should come as much surprise as this is what current apps are coded to (and even then, people we were seeing little differences between 4X and 8X).

 

[Nick]

Still interesting to note and understand the differences between ATI and Nvidia's use of PCIe bandwidth. It might have an impact down the road on something you do care about.

If it ever has an impact, I'm sure NVIDIA will take care of it before it becomes an issue.

Congratulations if your Lada goes 10 km/h when you hardly push the gas, while my Porche still roars in place. It goes over 300 km/h when I push it to the metal.

 

[trinibwoy]

Congratulations if your Lada goes 10 km/h when you hardly push the gas, while my Porche still roars in place. It goes over 300 km/h when I push it to the metal.

Your analogy is off. You don't care why your porsche can do 300, only that it does. Some people like to know the why and how.

 

[Nick]

Your analogy is off. You don't care why your porsche can do 300, only that it does. Some people like to know the why and how.

I certainly understand the enthousiasm about these seemingly interesting results. But seriously, the cause could be anything. With the bandwidth only one PCI-Express lane delivers, not much data has to be transferred to get a 25% performance decrease compared to using sixteen lanes. It's pretty pointless to want to know 'why' the NVIDIA card transfers that tiny bit of extra data over the bus.

Besides, the Radeon suffers from the same slowdown in 3D Mark 2005. It just makes the guessing harder and more pointless.

My take: the NVIDIA driver attempts to balance memory usage, so it always keeps some video memory free for when it's suddenly needed. That could keep things running more smoothly, at the cost of an imperceptable average performance loss. The ATI driver might try to cramp as many textures in video memory as possible, and only swap when really required, like in texture-heavy applications such as 3D Mark 2005. More likely both drivers do a bit of both, the parameters were just chosen a bit different.

So it still doesn't matter. Tiny differences get exxagerated when looking at numbers produced in situations that were not and should not be anticipated on. What matters is the behaviour when using eight or sixteen PCI-Express lanes.

 

[Dave Baumann]

My take: the NVIDIA driver attempts to balance memory usage, so it always keeps some video memory free for when it's suddenly needed.

According to my testing, it always seems to be the case that NVIDIA falls off more in high memory situations; TR:AoD being one case where it refuses to operate at high resolution on a number of boards and more resolutions when AA is applied.

 

[mjtdevries]

Nick, this might seem hard for you to believe, but nobody in this thread has shown any 'enthousiasm' for these results.

They do not apply to normal use, so why would we?

We are not trying to bash Nvidia and I certainly don't want to start a Nvidia vs ATI flamewar.

It's just that these results are intruiging, because I don't see why there would be any difference at all.
Maybe it's not even the nvidia card we should be looking at, but the ATI card. Why does it loose performance in 3DMark05 and not in the other game benchmarks?

If people don't feel immediately think that 'their company' is attacked, we can have a meaningful discussion about this, and maybe learn something new about the cards. That is never pointless.

And if you still think it is pointless, than why not just ignore this thread?

 

[radar1200gs]

The discussion seems a little pointless to me.

Its seems you need 4 PCI-E lanes to get AGP class performance. Why would you go lower than that? Would you run an AGP 8x card in PCI (not PCI-E, just ordinary PCI) mode too?

 

[Nick]

Nick, this might seem hard for you to believe, but nobody in this thread has shown any 'enthousiasm' for these results. They do not apply to normal use, so why would we?

Well you're taking the words out of my mouth. There should indeed not be any enthousiasm for this. Not even interest, it's just artifically created circumstances with extremely little correlation to real-world situations. Tom's Hardware just included the numbers for 1X and 2X speeds because they could, and because 4X is the tilting point. That's the only conclusion that can be made.

We are not trying to bash Nvidia and I certainly don't want to start a Nvidia vs ATI flamewar. If people don't feel immediately think that 'their company' is attacked, we can have a meaningful discussion about this, and maybe learn something new about the cards. That is never pointless.

Why would I feel anything is attacked here? All I'm saying is that it's pointless to try to conclude anything else from these numbers. We can't "have a meaningful discussion about this".

It's just that these results are intruiging, because I don't see why there would be any difference at all.
Maybe it's not even the nvidia card we should be looking at, but the ATI card. Why does it loose performance in 3DMark05 and not in the other game benchmarks?

Like I said before, really tiny differences become visible in these extreme conditions. It could be truely anything. Air humidity influencing connection quality.

And if you still think it is pointless, than why not just ignore this thread?

It just strikes me that people waste their time with something this useless, thinking it really means something. By all means, feel free to continue discussing it, but I haven't seen much solid facts come out of it and I don't expect any...

 

[Fodder]

Not only in games actually. You can see in specviewperf that nvidia loses more performance than ATI when reducing PCI-E lanes.

Might this have been a factor in NVIDIA choosing an external bridge for their multi-card solution? In the same vein (and I apologise for the thread hijack), would a bridged solution present difficulties in simultaneous inter-card communication and regular PCIE traffic?

Would you run an AGP 8x card in PCI (not PCI-E, just ordinary PCI) mode too?

Well, yes.

 

[Nick]

Would you run an AGP 8x card in PCI (not PCI-E, just ordinary PCI) mode too?

Well, yes.

But would you?

It's clearly not the way it's meant to be played to get in the game.

 

[2senile]

@ Nick.

I hope the members of B3D continue to discuss aspects of 3D graphics cards that go beyond the simple frames per second that are covered by a multitude of review sites.

There should indeed not be any enthousiasm for this. Not even interest, it's just artifically created circumstances with extremely little correlation to real-world situations.

I'm interested to see opinions on this even though it will never effect "real world performance".

I've always considered B3D to be the site to visit in order to get something more than the standard view.

 

[Kaotik]

Not only in games actually. You can see in specviewperf that nvidia loses more performance than ATI when reducing PCI-E lanes.

Might this have been a factor in NVIDIA choosing an external bridge for their multi-card solution? In the same vein (and I apologise for the thread hijack), would a bridged solution present difficulties in simultaneous inter-card communication and regular PCIE traffic?

Would you run an AGP 8x card in PCI (not PCI-E, just ordinary PCI) mode too?

Well, yes.

You have seen the "AGP Express" benches haven't you? IIRC worst case scenarios were about ~70% performance drop compared to AGP 8x

 

[Fodder]

You have seen the "AGP Express" benches haven't you?

Indeed, it's horrible. I'm just doing my thing and making a point though.

 

[Nick]

I hope the members of B3D continue to discuss aspects of 3D graphics cards that go beyond the simple frames per second that are covered by a multitude of review sites.

I've always considered B3D to be the site to visit in order to get something more than the standard view.

Please don't understand me wrong. I am more than interested in 3D graphics theory and implementation myself. And I too wish that at Beyond3D we can continue to discuss every interesting subject that goes beyond what other sites cover.

But in my eyes, everything that can be said about these numbers, has already been said. I even think I gave the biggest theory about it myself. Like I've said before, please do start a discussion about it if you can deduce more information out of it, but frankly, I'm not expecting much...

 

[davepermen]

what i like is the fact that even at 1x, the cards perform very well.. this could lead to an ultra-sff-mainboard with just 1x on it

seriously, it could lead to a final design where gpu's get "just another piece of cards", and lower end cards won't require 16x pcie anymore. i mean, if they balance it well, about all non-highest-end cards could work out with 1x.. now they just need a small form factor for the card, and passive cooling on it..

reminds me of the 1gb network cards for pci.. at work, we had, at the release dates of 1gb lan huge cards.. now, they are just big enough to have the pci connection, and the lan connection.. the rest is just.. to connect the two (and one chip for the hw of course).

small cards like the 9600 where could run on 1x today.. hihi..