Poor Quad Fire Scaling

[Lepton87]

I built myself a brand new rig and I'm not entirely satisfied with it. The rig is as follows:

i5 2500K
Asus P8p67pro(gone bad) then Asus Maximus IV ROG
2x2GB DDR3 1600MHz (leftover part from previous computer)
Radeon 6990+ 2x6950s unlocked reference
Adata SSD 64GB SF 1200

I first run 3Dmark11 at what I thought was stock clocks
(actually my mobo auto-overclocked my processor to 42x103)
here are the scores:
6990+6950

11555

6990+2x6950

11800

I partly blamed the piss poor quad fire scaling on PCI-E lanes. (8x/8x/4x) clearly seemed insufficient. I wanted to run some games to see if it still holds true in games but then my mobo just stopped working. I couldn't wait a couple of weeks for a replacement so I decided to fork out a lot of money for ASUS ROG IV MAXIMUS. Once again my intention to test everything at stock failed because the mobo pegged the processor at constant 3.7GHz and I realized that after testing, but it's a pretty mild OC. Here are my results:


Single 6990 8900

6990+6950 (8x/8x) 11394

2x6950+6990 (8x/16x/16x) via NF200

11800

So, I'm still not seeing any real gain from adding a 4th gpu.

Is that perfectly normal or is something wrong with my system?

 

[toyota]

are you running 3Dmark11 at 1280? if so then of course its not going to scale worth a crap. and yes having just a 4x pci-e lane will limit it too. really I cant imagine why you bought such a gpu setup in the first place if your mobo can only handle 4x. not to mention you better be running at least 2560x1600 or more for that much gpu power to make any sense.

 

[Grooveriding]

Get a better motherboard, ditch the two 6950s, get a second 6990.

 

[Lepton87]

My new mobo supports 8x/16x/16x via NF200 bridge, tomshardware tested such a configuration and it outperformed Nehalem X58 native solution, that's why I chose to go with SB despite its pathetic 16 PCI-E lanes connectivity. Anyway 8x/16x/16x didn't really improve much over 8x/8x/4x, I'm gonna have to retest my ROG mobo at the same clocks as my previous mobo to see how much it really helped. As far as I remember 3dmark11 performance preset uses higher resolution then 1280x1024.

 

[Lepton87]

Get a better motherboard, ditch the two 6950s, get a second 6990.

Which mobo is better? ASUS ROG MAXIMUS IV is top of the line. And how is packing the same amount of power into one slot instead of two supposed to help? Remember those two 6950s run at the same clocks and have the same amount of SP as 6990.

 

[sandorski]

Neither Quadfire nor QuadSLI make much sense. For Gaming anyway. 3-way is about the max for either setup.

 

[Lepton87]

http://www.guru3d.com/article/his-radeon-hd-6990-crossfire-review/21


After reading this article it seems that 3dmark11 just benefits very much from HT, that's why he got a better score on a single 6990 on a paltry 3750MHz nehalem system than I got on my SB system. Thankfully it's not gonna hold true in actual games.

 

[toyota]

My new mobo supports 8x/16x/16x via NF200 bridge, tomshardware tested such a configuration and it outperformed Nehalem X58 native solution, that's why I chose to go with SB despite its pathetic 16 PCI-E lanes connectivity. Anyway 8x/16x/16x didn't really improve much over 8x/8x/4x, I'm gonna have to retest my ROG mobo at the same clocks as my previous mobo to see how much it really helped. As far as I remember 3dmark11 performance preset uses higher resolution then 1280x1024.

what do you mean as far as you can remember? the results are right in front of you and list the resolution. not to mention you can clearly see its pretty low res when running.

and what is your native res?

 

[Lepton87]

what do you mean as far as you can remember? the results are right in front of you and list the resolution. not to mention you can clearly see its pretty low res when running.

and what is your native res?

1920x1200


The culprit for my low scores in benchmarks is my processor. I also did ran 3Dmark Vantage and my graphics score(45500) is higher than that in 6990CF review.

http://3dmark.com/3dmv/3310981


so as it turns out HT makes a huge difference in synthetics. It still doesn't make me feel bad about skimping on a processor.
As far as I remember means I ran those benchmarks yesterday and I'm currently typing on a laptop from the comfort of my bed.

 

[toyota]

1920x1200


The culprit for my low scores in benchmarks is my processor. I also did ran 3Dmark Vantage and my graphics score(45500) is higher than that in 6990CF review.

http://3dmark.com/3dmv/3310981


so as it turns out HT makes a huge difference in synthetics. It still doesn't make me feel bad about skimping on a processor.
As far as I remember means I ran those benchmarks yesterday and I'm currently typing on a laptop from the comfort of my bed.

first off it is silly as hell to be running 4 high end gpus for 1920x1200. that fourth gpu and in some cases third gpu will be severally wasted.

next of course your cpu is going to be a bottleneck at just 1280 with that much gpu power.

 

[Lepton87]

first off it is silly as hell to be running 4 high end gpus for 1920x1200. that fourth gpu and in some cases third gpu will be severally wasted.

next of course your cpu is going to be a bottleneck at just 1280 with that much gpu power.

Well, the problem is the i5 2500k is currently the fastest processor for games money can buy, barring only 2600K which is at most 2-5% faster due to having more cache, HT doesn't help in games.

 

[toyota]

Well, the problem is the i5 2500k is currently the fastest processor for games money can buy, barring only 2600K which is at most 2-5% faster due to having more cache, HT doesn't help in games.

its not a problem because nobody in their right mind is going to use FOUR high end gpus at just 1280. again even at 1920, its a ridiculously inefficient gpu setup. get rid of either the 6990 or both 6950s and go play games at 1920 instead of running 3dmark at 1280.

 

[Lepton87]

I was just trying to find out if there's something wrong with my rig because it posted low scores on popular benchmarks, why being so rude? I haven't even got the time yet to run a game. chill out.

 

[toyota]

I was just trying to find out if there's something wrong with my rig because it posted low scores on popular benchmarks, why being so rude? I haven't even got the time yet to run a game. chill out.

lol, I am being my usual blunt self. you just do not see how silly what you are doing is. you bought way too much gpu power for 1920 never mind 1280.

 

[Smartazz]

I would sell either the 6990 or both 6950s. Crossfire HD6950s or an HD6990 is more than enough for 1920x1200.

 

[Smartazz]

Yeah, using four GPUs is pretty ridiculous for 1600p let alone a res nearly half that.

 

[MagnusTheBrewer]

The silliest thing is using benchmarks to gauge performance in games if your interest is games rather than benchmarks.

 

[Lepton87]

Don't be so narrow sighted, maybe I just like SSAA? Yes, SSAA really does need that much raw graphics power. I would have packed 3x6990 if the drivers allowed that

"The silliest thing is using benchmarks to gauge performance in games if your interest is games rather than benchmarks." I never attempted that, in fact I said as much about HT not making a difference in games as opposed to synthetics.

 

[Bman123]

Run some game benchmarks

 

[toyota]

Don't be so narrow sighted, maybe I just like SSAA? Yes, SSAA really does need that much raw graphics power. I would have packed 3x6990 if the drivers allowed that

"The silliest thing is using benchmarks to gauge performance in games if your interest is games rather than benchmarks." I never attempted that, in fact I said as much about HT not making a difference in games as opposed to synthetics.

you would have used 6 gpus for 1920 if you could have?

 

[Lepton87]

you would have used 6 gpus for 1920 if you could have?

Nah, that would warrant upgrading the monitor


I measured how much NF200 chip helped.

103x42 P8P67pro 8x8x4x scored 12022 17855/6875/5169 GF/PH/C respectively
100x44 ASUS MAXIMUS IV EXTREME(price tag) 8x/16x/16x scored P13023 Graphics Score 18329/Physics Score 8108/Combined Score 5759

Physics score discrepancy just doesn't make any sense. NF200 is just a band-aid for that paltry 16 lanes but it does make a difference. Now I'm gonna test how the rig actually play games, one of the reasons why I chose quad GPUs is that it actually should have less micro-stutter than dual GPUs.

 

[Dark Shroud]

That NF200 is just a bridge/Plex chip. What I mean by that is the chip does not make any more bandwidth while adding more lanes. It just manages the exsisting bandwidth in a smart way. The down sides is this adds more latency to the set-up. And if something goes wrong it can cause more problems than one slot just not being usable.

 

[notty22]

The latency issue is more internet myth than reality. Its a high speed digital switch.
It multiplexes communication which in the end, allows for more bandwidth.

 

[Jacky60]

Neither Quadfire nor QuadSLI make much sense. For Gaming anyway. 3-way is about the max for either setup.

Depends which games you're playing, I'd guess he, like me, is playing VERY demanding games not console ports and need the xtra horsepower. Arma 2, Metro 2033, and Shogun 2 (I think) could all benefit from the extra power and crossfire scaling is exceptional these days so your statement is incorrect.

 

[Jacky60]

Don't be so narrow sighted, maybe I just like SSAA? Yes, SSAA really does need that much raw graphics power. I would have packed 3x6990 if the drivers allowed that

"The silliest thing is using benchmarks to gauge performance in games if your interest is games rather than benchmarks." I never attempted that, in fact I said as much about HT not making a difference in games as opposed to synthetics.

I agree with you Lepton, loads of people seem to object to gamers whacking up the detail at 1920x1200 and follow it with a lecture that its either stupid, inappropriate or just 'wrong'. I can't even run SSAA on my 2 6990's without the frame rate going a bit wobbly and thats at 1920/1200 in ARMA 2. People with crap cards can't imagine that some games need more power. You really need that 2500K at 4.6 Ghz for those cards to stretch their legs though.