Bottleneck? Comment on built please.

[teddyjai0614]

My question is this, I recently built a new system and I would like to know if my CPU will bottleneck my GPUs or not since I've read about various things about intel socket 1156 isnt fully equipped to handle quadfirex 5970.

This is my build.
Case: Ikonik ra x10 - LCS case (with dual quad 80mm rads)
CPU: 1156 i7 875k unlocked multiplier cpu
Mobo: Asus Maximus 3 Extreme with koolance full chipset liquid cooling
Ram: 4x2gb Corsair Dominator-gt pc1600 cl7
Video (1): 2 Xfx 5970 in quadfirex (2x2gpu) with koolance waterblock
Video (2): eVGA 9800gt 512mb video for dedicated phsyx
PSU: Sparkle 1250w 80plus gold certified psu
Rad (1): 2 single 120mm Black Ice gtx rad with 93cfm 120mm fan
Rad (2): single 140mm Black Ice gtx rad with high cfm 140mm fan
Rad (3): 2 rows of quad 80mm radiator stocked by Ikonik case
Pump (1): Ikonik stock pump
Pump (2): Swiftech 355 pump
HD (1): 2 Intel 80gb SSD x25-m in raid0 (os boot)
HD (2): 1 TB WD Black 64mb 6gbps in SATA3 for storage
OD (1): LG DVDRW
OD (2): LG Blu-ray Writer
Mon: 3 Acer 23'in 1080, 5ms, 1:50k (eyefinity)

System is running 2 LCS loops.
1: 2 rows of quad 80mm radiator on chipset/mosfet and cpu.
2. 2 single 120mm rad and a single 140mm rad on dual ati 5970 video

I'm able to OC cpu to ~4.2ghz stable - havent tried pushing it further yet but fully intend to since I recently changed PSU to 1250w (was using Corsair 950w), added the 140mm single rad, added second 5970 card, and added the 9800gt card as dedicated phsyx.

The temps are around ~30'c idle and ~40-42'c full load with cpu
Videocards is ~30'c idle and ~40 full load.

My questions are:
1. Will my cpu bottleneck my dual 5970 if I ran my cpu in stock speed? 2.93ghz quad core.
2. Will overclocking my cpu help the bottle necking issue? If so to what speed?
3. Will the 1250w PSU handle my rig effectively and stable?
4. Will the board Asus Maximus III Extreme able to handle dual 5970 both in 16x and my 9800gt in 8x (this board does not have a nf200 chipset)?

The board specs says it can handle dual video cards at 16/16, 3 at 8/16/8, 4 at 8/8/8/8 and 5 at 8/8/8/8/8. I notice the lanes runs by pair, lets say slot 1 and 2 is connected, slot 3 and 4 is connected (example if only slot 1 is occupied and slot 2 is not, then it runs on 16x with slot 1, but if slot 1 and 2 is both occupied then both shares the 16x speed so both slot 1 and 2 will both run on 8/8 instead). But the question is theres 5 lanes, 1 shares with 2, 3 shares with 4, and the 5th slot is alone? So am I safe to assume that if I run the video cards on slot 1 and 3, both will run 16x each (as specified by the manufacture?), while I can plug the 9800gt dedicated phsyxs on slot 5 as a stand alone 8x not sharing any bandwidth with any other slots? I wanted to make sure that I'm not throwing in a dedicated phsyxs to gimp my 5970's speed, which leads me to my 5th question...

5. If lets say I run 1 5970 on 16x, 1 5970 and a 9800gt dedicated physx both on 8/8, is the speed noticable than compared to dual 5970 both at 16x lane per without phsyx?

 

[Proximon]

Should you have bought a socket 1366 system? Probably.

Yes that's the way your particular board works and you will have the bandwidth needed. A PhysX card does not need x8, x4 would be fine if that's all you end up with.

I have no eyefinity numbers, but given that the CPU load should not change much between a single monitor setup and three overclocking the CPU around 3.8-4.0Ghz will gain you about 10FPS in demanding games at max settings. These would be mostly CPU intensive strategy games. If you need the 10 FPS then do it.

http://benchmarkextreme.com/Articles//GTX%20480%20%20Radeon%205970%20CPU%20Bottleneck/P3.html

 

[banthracis]

Your mobo is a bottleneck to the GPU's. That mobo does not magically give 1156 32 lanes of PCIe. The NF2000 chipset merely allows both GPU's to share the 16 lanes simultaneously, instead of being divided to 8 dedicated to each like it normally does.

The NF200 hub WILL improve performance over 2x8 lanes, but it does NOT help in your case, as in the end, you're trying to push 4 gpus over 16 lanes, which is a big no no.

You're falling for a marketing ploy. Yes each slot has access to 16 lanes, but end of the day, the TOTAL lanes available to ALL GPU's is still 16.

Don't need physx, it's dead. Also, it only makes the PCIe bottleneck even worse.

Besides that power wise it's fine. You can run this rig fine on a good 850w, or 950w if you really wanna play it safe.

 

[teddyjai0614]

Your mobo is a bottleneck to the GPU's. That mobo does not magically give 1156 32 lanes of PCIe. The NF2000 chipset merely allows both GPU's to share the 16 lanes simultaneously, instead of being divided to 8 dedicated to each like it normally does.

The NF200 hub WILL improve performance over 2x8 lanes, but it does NOT help in your case, as in the end, you're trying to push 4 gpus over 16 lanes, which is a big no no.

You're falling for a marketing ploy. Yes each slot has access to 16 lanes, but end of the day, the TOTAL lanes available to ALL GPU's is still 16.

Don't need physx, it's dead. Also, it only makes the PCIe bottleneck even worse.

Besides that power wise it's fine. You can run this rig fine on a good 850w, or 950w if you really wanna play it safe.

I was wrong in my first post (dam newegg didnt list it -_- ), the Asus MIIIE does have a NF200 chipset and provides 40 pci-e lanes. While it can run 16x/16x/8x, the question will then be whether the 2.93ghz cpu bottleneck the video cards or not and if so, will overclocking the cpu help in preventing the bottleneck or at least soften the blow?

 

[banthracis]

Did you read anything I just said about NF200?

You DO NOT have 40 PCIE lanes. You still ONLY have 16. This is a CPU limitation. There is NO way around it. The NF200 chip merely allows all the lanes access to the full 16 lanes.

It's like the midtown tunnel here in Manhattan.
There are 8 lanes of traffic leading to a 2 lane tunnel. You can either divide it so that the 4 lanes of traffic entering from the north, and the 4 lanes of traffic entering from the south, lead to their own exclusive tunnel lane. Alternatively, you can simply have all 8 traffic lanes merge into the 2 tunnel lanes.

Yes, in the second case each traffic lane has access to 2x the number of tunnel lanes as the first case, however, there aren't actually more tunnel lanes, you're just giving each car lane access to more tunnel lanes.

In the specific case where you have 4 lanes of traffic leading to 2 separate tunnel lanes, and where 1 of the tunnel lanes is being under used (Ie there are no cars trying to enter from the north, but tons of cars entering from the south, it makes sense to allow the south cars access to the second tunnel to speed up traffic).

However, if it's like rush hour where it's bumper to bumper traffic trying to get into the 2 tunnel lanes, it doesn't matter if each tunnel lane can be accessed by 4 traffic lanes, or all 8 traffic lanes.

Now, replace traffic lanes with GPU's, tunnel lanes with PCIe lanes, and rush hour with the data flow from 4 GPU's.

 

[teddyjai0614]

With all that said Banthracis, which leads to my next question.

In percentage wise, how much of a performance drop am I losing here in opposed to going for 1366 platform? Give me a guestimate? I'm guessing around 30-40% minimum?

I've read around and people said 4 gpu is the max people can use for a gaming computer, why is that? Please explain the 4 gpu limitation in correlation with the cpu.

Also is there a cpu speed to gpu speed ratio? People have said if you overclock a cpu to a certain level, you will gain 10 more fps, is there a certain ratio to prove that theory or statement?

 

[Proximon]

In percentage wise, how much of a performance drop am I losing here in opposed to going for 1366 platform? Give me a guestimate? I'm guessing around 30-40% minimum?

No far less. Here:
http://www.tomshardware.com/reviews/p55-crossfire-nf200,2537.html

The NF200 chipset may not magically add 24 lanes, but it does allow better bandwidth in the third slot.
At least with CF 5870s the difference between LGA1156 and LGA X58 is not going to be 30%, and likely never more than 10%.
Here's some partial proof of that:
http://www.overclockers.com/forums/showthread.php?t=619246#post6241419

I've read around and people said 4 gpu is the max people can use for a gaming computer, why is that? Please explain the 4 gpu limitation in correlation with the cpu.

Adding GPUs is a matter of diminishing returns, just like spending too much on a PC. The more you spend the less you get for the buck. Here is a complete comparison done over a year ago, but still very relevant:
http://www.anandtech.com/show/2728

Also is there a cpu speed to gpu speed ratio? People have said if you overclock a cpu to a certain level, you will gain 10 more fps, is there a certain ratio to prove that theory or statement?

No. It would vary game to game and driver set to driver set. It's done with benchmarks.

 

[banthracis]

Proximon is correct, however there is 1 issue. The Tom's article was done with 3 GPU's and demonstrated a tiny loss. However, the 3 GPU - 4 GPU change makes it a big issue. With 3 GPU's, each GPU essentially get 5 lanes with another being used by w/e GPu is in high load. With 4, you're down to essentially 4 lanes each.

The problems is I haven't seen anyone test 4 gpu's over 16 lanes. We have seen 1 gpu being pushed over 4 lanes, and the results aren't pretty. That one IS a 30% loss over 16lanes.
http://www.tomshardware.com/reviews/p55-pci-express-scaling,2517-6.html

Now, how much the NF200 alleviates this, no idea, but you're gonna have a significant performance loss.

What makes this worse is that by adding the 4th GPU, you're actually Lowering the performance of the other 3, by limiting their bandwidth. In fact, I'd say you would get BETTER performance if you only used 3 GPU's.

No articles I've found testing this though.

As to what OC you would need to max out 4 GPU's.
This depends on:
1. Resolution
2. Settings ( like AA)
3. Game type. Strategy games benefit more from CPU OC due to their extensive AI calculations. FPS's are more eye candy oriented and CPU may never reach a bottleneck.

 

[teddyjai0614]

Proximon is correct, however there is 1 issue. The Tom's article was done with 3 GPU's and demonstrated a tiny loss. However, the 3 GPU - 4 GPU change makes it a big issue. With 3 GPU's, each GPU essentially get 5 lanes with another being used by w/e GPu is in high load. With 4, you're down to essentially 4 lanes each.

The problems is I haven't seen anyone test 4 gpu's over 16 lanes. We have seen 1 gpu being pushed over 4 lanes, and the results aren't pretty. That one IS a 30% loss over 16lanes.
http://www.tomshardware.com/reviews/p55-pci-express-scaling,2517-6.html

Now, how much the NF200 alleviates this, no idea, but you're gonna have a significant performance loss.

What makes this worse is that by adding the 4th GPU, you're actually Lowering the performance of the other 3, by limiting their bandwidth. In fact, I'd say you would get BETTER performance if you only used 3 GPU's.

No articles I've found testing this though.

As to what OC you would need to max out 4 GPU's.
This depends on:
1. Resolution
2. Settings ( like AA)
3. Game type. Strategy games benefit more from CPU OC due to their extensive AI calculations. FPS's are more eye candy oriented and CPU may never reach a bottleneck.

Games I play right now is
1. World of warcraft
2. Left 4 dead 1 and 2
3. Starcraft 2

may be more in the future but that is it about now.

I will most likely upgrade to 1366 platform in a month or 2 to take full advantage of my hardware.

My initial build was going to be a single 5970, but somehow I went dual - not noticing about the limits of future proof with the current setup... I messed up. Nonetheless I will probably have to grab something along the lines of a i7 950 or something around there as well as the Asus RIIIE board along with the liquid cooling for a cool oc build.

Any idea if running a dedicated physx is worthwhile or not?

 

[banthracis]

Dedicated physx is pointless. None of the games you listed even use physx.

WOW is mostly GPU dependent IF you have shadows enabled. Even then, you're gonna be at a much higher FPS than your monitor can display. A 60hz monitor can only display 60FPS.

Left 4 dead can be maxed out a last generation hardware at 60fps, so that's also not close to an issue.

SC2, haven't seen release version bench's yet, but if AA is forced, then there's a heavy GPU load, so CPU won't likely hit max.

You can see the 1 5970 + i7 bench's here. Very little benefit in most games going from stock speeds to even a 4.0ghz OC. This is what you'll likely see given that you're running 3 monitors on 2 5970's.

http://www.legionhardware.com/articles_pages/cpu_scaling_with_the_radeon_hd_5970,5.html

 

[dndhatcher]

PhysX only works on a few specific games that are programmed for it. Batman Arkham Asylum is probably the most popular. Last I checked there were only 11 games that incorporated Physx. You can do a quick google search, there is a site that lists them all. Nvidia dropped PhysX when they released the 400 series GPUs so its a dead end at this point.

None of those games you listed need more than a single 5970 to max out at 2560x1600. If thats what you are playing, you should pull the other two graphics cards and see if you can tell any difference. You will likely save alot of money in power use and system wear by leaving them out.

 

[banthracis]

He's running 3 1080p monitors in eyefinity dnd^^

 

[dndhatcher]

I dont know about Starcraft 2, but even at 5670x3240 I am pretty sure a single 5970 can max out WoW and likely L4D at over 60FPS. Its easy to play some with both cards, then with only one and see if you can really tell the difference.