***NVIDIA 3-Way SLI and AMD Tri-Fire Redux***

[Final8ty]

Today we are bringing you a unique performance evaluation that has evolved out of a recent article we published.

On April 28th, 2011 we published an evaluation that compared NVIDIA GeForce GTX 580 3-Way SLI and AMD Radeon HD 6990+6970 Tri-Fire performance. In that evaluation all of the evidence we gathered pointed toward the Tri-Fire solution being the better value and delivering the absolute best performance, outclassing the more expensive GTX 580 3-Way SLI configuration. After publication, we received feedback that perhaps 3-Way SLI was not getting its fair shake at gaming performance due to our then current system configuration limiting its ability. While we did not think that scaling the CPU clock would actually flip-flop our real world gaming results, we thought our readers had made some really good points with us and we wanted to retest to see what validity we could find in their questions. Over the last 8 years or so, real world gameplay testing has taught us a lot things and it was about to teach us a few more things about multi-GPU setups, which honestly, we do not spend a lot of time with unless we just happen to have a lot of time open in our schedules. But let it be said, and we are eating some crow here, you knew more about it than we did, and we are glad we listened to you.

We have been using an Intel X58 chipset motherboard with an Intel Core i7-920 overclocked to 3.6GHz. This system has worked great for us for a long while now, but more than dual-GPU performance may benefit from a faster system. The time has come for us to upgrade our video card testing rig for super high-end video card reviews. We are now using the absolute latest motherboard and most powerful CPU in order to find out if a faster CPU really does affect 3-Way SLI and Tri-Fire performance. It is time for use to "upgrade" from our X58 "flagship" system.

http://www.hardocp.com/article/2011/..._trifire_redux

[Vinas]

This is starting to seem like a publicity stunt or something...

[hurleybird]

Well, now I don't know what to think anymore. Lower performance on trifire on higher performing system? Seems like a bug or something not being set up right, and I'd also like to see them test true tri-fire, not quasi "6990 + 6970" trifire.

Not to mention, at eyfinity/surround resolutions like 5760x1200 with 3 GPUs, both AMD and Nvidia are doing it wrong. Each GPU should drive a single monitor. The ones on the side don't even need to be perfectly in sync with the one in the middle (it's not like you're going to get tearing with bezels seperating the picture!), and you wouldn't need to worry about input lag or micro stutter issues either.

Heck, even in a 2 GPU system you could have one GPU for both side monitors (each being less demanding) and one GPU for the middle screen (more demanding), or possibly in such a configuration run the side monitors at half frame rate with some driver magic.

triple head gaming is the best reason to use multiple graphics cards, but the implementation still sucks and is designed around using a single monitor.

[eXa]

They are doing Quad sli/fire next and will come back to 3 580s/3 6970s

[Final8ty]

Well, now I don't know what to think anymore. Lower performance on trifire on higher performing system? Seems like a bug or something not being set up right, and I'd also like to see them test true tri-fire, not quasi "6990 + 6970" trifire.

Not to mention, at eyfinity/surround resolutions like 5760x1200 with 3 GPUs, both AMD and Nvidia are doing it wrong. Each GPU should drive a single monitor. The ones on the side don't even need to be perfectly in sync with the one in the middle (it's not like you're going to get tearing with bezels seperating the picture!), and you wouldn't need to worry about input lag or micro stutter issues either.

Heck, even in a 2 GPU system you could have one GPU for both side monitors (each being less demanding) and one GPU for the middle screen (more demanding), or possibly in such a configuration run the side monitors at half frame rate with some driver magic.

triple head gaming is the best reason to use multiple graphics cards, but the implementation still sucks and is designed around using a single monitor.

That is how Nvidia is doing it & maybe that's why its needs more CPU power.

[ryboto]

I was just going to observe this whole thing, but how does the AMD setup lose performance with a much faster CPU???

[generics_user]

I was just going to observe this whole thing, but how does the AMD setup lose performance with a much faster CPU???

i guess that it's suffering from the weaker I/O system on P67

remember that X58 offers 2 native fullsize x16 slots and P67 only has 2 x8 slots

maybe amd is more pcie bandwidth starved while nvidia requires more cpu power?

[bill_d]

I was just going to observe this whole thing, but how does the AMD setup lose performance with a much faster CPU???

dosn't the nf200 chip hurt performance with less cards used, something about shifting lanes to unused slots
i think i remember something about the ASUS P6T7 WS Supercomputer running 2 way cf slower than a x58 mb without a nf200

[[XC] hipno650]

Well, now I don't know what to think anymore. Lower performance on trifire on higher performing system? Seems like a bug or something not being set up right, and I'd also like to see them test true tri-fire, not quasi "6990 + 6970" trifire.

Not to mention, at eyfinity/surround resolutions like 5760x1200 with 3 GPUs, both AMD and Nvidia are doing it wrong. Each GPU should drive a single monitor. The ones on the side don't even need to be perfectly in sync with the one in the middle (it's not like you're going to get tearing with bezels seperating the picture!), and you wouldn't need to worry about input lag or micro stutter issues either.

Heck, even in a 2 GPU system you could have one GPU for both side monitors (each being less demanding) and one GPU for the middle screen (more demanding), or possibly in such a configuration run the side monitors at half frame rate with some driver magic.

triple head gaming is the best reason to use multiple graphics cards, but the implementation still sucks and is designed around using a single monitor.

well if you look at the top part of your browser you will see [H] that should explain everything for you.

not event he most hardened Nvidiot could believe that the AMD system performs CONSIDERABLY worse on a much faster base system...

[Eastcoasthandle]

I believe the problem stems from the fact that neither set of motherboards offer AMD and Nvidia full x16/x16 or x16/x16/x16. Either one board will do x16/x16/x4 or the other will do x8/x8 which will hamper performance on those games sensitive to bandwidth constraints.

All we are seeing in the redux review is the performance allowed to the 580s in 3-way SLI when the NF200 chip is used vs 6990/6970 capped at x8/x8 which is what the P8P67 WS Revolution defaults to when 2 cards are used w/o the nf200 chip. IMO, they either need to go back to the x58 board and use 6990/6970 vs 590/580 or use the P67 board with a similar video cards.

[[XC] hipno650]

I believe the problem stems from the fact that neither set of motherboards offer AMD and Nvidia full x16/x16 or x16/x16/x16. Either one board will do x16/x16/x4 or the other will do x8/x8 which will hamper performance on those games sensitive to bandwidth constraints.

All we are seeing in the redux review is the performance allowed to the 580s in 3-way SLI when the NF200 chip is used vs 6990/6970 capped at x8/x8 which is what the P8P67 WS Revolution defaults to when 2 cards are used w/o the nf200 chip. IMO, they either need to go back to the x58 board and use 6990/6970 vs 590/580 or use the P67 board with a similar video cards.

it would be interesting to see and triple 6970 setup but I think the per would be the same. all of the benches I have seen point to there being a negligible performance hit when using 16x or 8x on P67 with even a 6990. not to mention that the NF200 chip only gets 8x PCI E bandwidth anyways...

I would like to see this done on a site like HWC rather than [H] to make a final decision. I can say however that the scaling on Tri SLI on my setup is quite good. most games offer about 2.7 times the perf compared to a single GPU. some even more than that.

[kaktus1907]

Hardocp fails once again ?

^ lol [H]ardFail

[SKYMTL]

I won't say much other than the fact everyone already knows: this should have all been worked out BEFORE publishing the FIRST article. I hate seeing sites that drag their readers merrily along while they slowly discover some of the original and even current benchmarking methodology is FUBAR'd. Publish hard facts from the outset and there won't be any issues. Publish assumptions like "AMD and NVIDIA are both doing it wrong" and someone's going to call the bluff.

One thing's for certain; it's good for traffic.

This approach is also good for relations with potential advertisers though. AMD is happy with the first article while NVIDIA will be happy with the follow-up. Meanwhile, readers are left wondering which results they should believe.

[cold2010]

Hardocp fails once again ?

yes This time with ,16x/16x/16x vs 16x16x/16x/4x

[hurleybird]

That is how Nvidia is doing it & maybe that's why its needs more CPU power.

I believe that even though you need to hook up monitors to different cards, the end result is still AFR. They should be splitting the 5760x1200 "screen" into three vertical parts, one for each monitor.

[Final8ty]

I believe that even though you need to hook up monitors to different cards, the end result is still AFR. They should be splitting the 5760x1200 "screen" into three vertical parts, one for each monitor.

But the syncing between monitors is important & that's why they don't do that.
Maybe you don't care about that but most do.

I find it very distracting if they are not & others too on Widescreengamingforums.

[hurleybird]

That would need to be a pretty significant lag to notice the difference with the image seperated by bezels.

I understand why AMD and Nvidia do AFR though -- higher frame rate number. Unfortunately frame rate between and AFR setup and a single card setup is not really comparable thanks to ms.

[CedricFP]

I believe the problem stems from the fact that neither set of motherboards offer AMD and Nvidia full x16/x16 or x16/x16/x16. Either one board will do x16/x16/x4 or the other will do x8/x8 which will hamper performance on those games sensitive to bandwidth constraints.

All we are seeing in the redux review is the performance allowed to the 580s in 3-way SLI when the NF200 chip is used vs 6990/6970 capped at x8/x8 which is what the P8P67 WS Revolution defaults to when 2 cards are used w/o the nf200 chip. IMO, they either need to go back to the x58 board and use 6990/6970 vs 590/580 or use the P67 board with a similar video cards.

IIRC, you can't do 590 + 580.

Also, you're probably spot on with the x8/x8. 6990 alone probably suffers from x8.

[sin0822]

I won't say much other than the fact everyone already knows: this should have all been worked out BEFORE publishing the FIRST article. I hate seeing sites that drag their readers merrily along while they slowly discover some of the original and even current benchmarking methodology is FUBAR'd. Publish hard facts from the outset and there won't be any issues. Publish assumptions like "AMD and NVIDIA are both doing it wrong" and someone's going to call the bluff.

One thing's for certain; it's good for traffic.

This approach is also good for relations with potential advertisers though. AMD is happy with the first article while NVIDIA will be happy with the follow-up. Meanwhile, readers are left wondering which results they should believe.

that is very true, of course [h]ocp needs to get rid of their over current protection lol. i think only i found that funny.

[Final8ty]

That would need to be a pretty significant lag to notice the difference with the image seperated by bezels.

I understand why AMD and Nvidia do AFR though -- higher frame rate number. Unfortunately frame rate between and AFR setup and a single card setup is not really comparable thanks to ms.

Not really as you should know that everyone is different to how sensitive they are to frame rate, tearing & lag.

[PatRaceTin]

we still need fast cpu for fast gpu

[zanzabar]

whats up with bc2 that is a huge drop in frames in 2 diffrent sports, and why are they taking AA out to make it not hinder vram but leaving AO (its on by default on bc2 and metro) thats not really fair for the bench since amd is better at AA and AO is crap that NV likes but it dose nothing for graphics.

why did they do this in the 1st place its just going to confuse people and were is the 2x sli to compare, and why did they test with an nf200 based board that will give amd a nice 5-10% penalty.

[BababooeyHTJ]

Something is clearly wrong with this comparison.

[cold2010]

Something is clearly wrong with this comparison.

Now only

[tajoh111]

whats up with bc2 that is a huge drop in frames in 2 diffrent sports, and why are they taking AA out to make it not hinder vram but leaving AO (its on by default on bc2 and metro) thats not really fair for the bench since amd is better at AA and AO is crap that NV likes but it dose nothing for graphics.

why did they do this in the 1st place its just going to confuse people and were is the 2x sli to compare, and why did they test with an nf200 based board that will give amd a nice 5-10% penalty.

Doesn't the nf100 give NV the penalty as well?