GTX 970 G1 Gaming SLI Performance issues.

[mull54]

Hey guys.

I reviced my two GTX 970 G1 Gaming(Really good looking cards, and damn they run quiet att 80% even) and i have a problem with performance.

First of, i ran a custom fan curve and got 67C on the top card on 80% fan speed. And 60C on the bottom card. Is that normal temps?

But to my main question. I ran Firestrike(free with demo) and only got 17000 in score. I know this is not normal. Do you guys have any idea what the problem is? Also i got a bit over 2600 on Heven bench on 1080P Ultra 8xaa and Extreme tess.

 

[Kursah]

What drivers are you using? New drivers just dropped yesterday... that's probably going to be 90% of your issues. The other 10% is the games/benches you run.

 

[mull54]

What drivers are you using? New drivers just dropped yesterday... that's probably going to be 90% of your issues. The other 10% is the games/benches you run.

The 3xx.16. It says they are the newset drivers. Other people get 23k on the same bench and i get 17k. Something is wrong.

 

[cadaveca]

The 3xx.16. It says they are the newset drivers. Other people get 23k on the same bench and i get 17k. Something is wrong.

if they aren't running the same CPU, at the same speed, your comparisons are null and void.

 

[mull54]

if they aren't running the same CPU, at the same speed, your comparisons are null and void.

It's the GPU score i'm talkning about. And i compared to stock.

 

[MxPhenom 216]

It's the GPU score i'm talkning about. And i compared to stock.

His point still stands.

 

[Kursah]

It all ties in together. You have a good build, but if those guys are running OC'd Haswell refresh/Devil's Canyon CPU's, faster RAM, RAID'd SSD's, etc...everything is going to be faster and smoother. Texture loading, processing. Just because an application is for testing graphics doesn't mean the rest of the system isn't included in the scoring. Sure the GPU's make up the bulk of it.

You are running X16 and X8 on your PCI-E correct? You running the most recent BIOS? Are you OC'd (I hope not with a stock cooler!)? What speed is that DDR3?

Does your system match your system specs with just a GPU change or is it totally different? The more details you provide when you're experiencing problems or suspect something will ultimately be more helpful for you.

 

[cadaveca]

It's the GPU score i'm talkning about. And i compared to stock.

The main differences in these chips that are out now, from 2600K and it's family, to 3770k, and 4770K, is differences in cache performance. Since GPU gets it's data from CPUs, the ability of a CPU to pass data to the GPU does in fact affect 3D performance. Do keep in mind that when it comes to graphics, not all work is done by the GPU... usually wireframe data is done by CPU, then that gets SKINNED, lighted, shadowed, and detailed, by the GPU. But CPU still generates the base that the GPU uses.

Otherwise, CPU OC would not affect the results in this benchmark you mention, but it does.

 

[mull54]

The main differences in these chips that are out now, from 2600K and it's family, to 3770k, and 4770K, is differences in cache performance. Since GPU gets it's data from CPUs, the ability of a CPU to pass data to the GPU does in fact affect 3D performance. Do keep in mind that when it comes to graphics, not all work is done by the GPU... usually wireframe data is done by CPU, then that gets SKINNED, lighted, shadowed, and detailed, by the GPU. But CPU still generates the base that the GPU uses.

Otherwise, CPU OC would not affect the results in this benchmark you mention, but it does.

The weird thing is that i get 12k GPU score with each card without SLI. So that should still me a combined GPU score of 22k+-2000. I even overclocked the core on both cards by 125MHz(stock is 1178). And the score was still 179XX.
Parhaps it is my old SLI bridge that i have had laying unprotected in a drawer for 3 years?

 

[Naito]

The weird thing is that i get 12k GPU score with each card without SLI. So that should still me a combined GPU score of 22k+-2000. I even overclocked the core on both cards by 125MHz(stock is 1178). And the score was still 179XX.
Parhaps it is my old SLI bridge that i have had laying unprotected in a drawer for 3 years?

It's still early days for the drivers. While Nvidia is very good with updates, let them mature a bit. A new SLI bridge may help, but unsure.

 

[zsolt_93]

It is the PCIE2.0 vs PCIE 3.0 thing too. I have read something with SLI/CFX on 2.0 motherboards with this generation is not optimal performance. Although bandwidth should be more than enough.

 

[mull54]

It is the PCIE2.0 vs PCIE 3.0 thing too. I have read something with SLI/CFX on 2.0 motherboards with this generation is not optimal performance. Although bandwidth should be more than enough.

Yeah it's PCI-E 2.0. Parhaps it's my motherboard that's killing the performance.

 

[Naito]

Yeah it's PCI-E 2.0. Parhaps it's my motherboard that's killing the performance.

And CPU, if you want do go down the path of PCIe 2.0 not being sufficient.

 

[mull54]

And CPU, if you want do go down the path of PCIe 2.0 not being sufficient.

I have a 2600k clocked at 4.2 GHz so i don't think it would "kill" the performance.

 

[RCoon]

It is the PCIE2.0 vs PCIE 3.0 thing too. I have read something with SLI/CFX on 2.0 motherboards with this generation is not optimal performance. Although bandwidth should be more than enough.

While not bleeding edge optimal, I still believe that PCI-E 2.0 x16 is enough for most GPU's. Granted an x16 PCI-E 3.0 lane has double the bandwidth, but in SLI both a 2.0 1150 mobo and 3.0 1150 mobo will provide the same bandwidth.
3.0 mobos have 16 lanes total, while 2.0 has 32. SLI/Crossfire on the 2.0 mobo will still provide x16/x16 performance (with no additional PCI-E devices), and the 3.0 will provide 3.0 x8/x8 performance, but is the equivalent of 2.0 x16/x16.
It's all a bloody mess basically. I'm sure 2.0 x8 is still enough bandwidth for modern GPUs, the x16 just provides a small leeway that can sometimes provide up to 1-2% performance increase, but not much more.

EDIT: Although his MoBo only provides x8/x8 on 2.0, my mistake. Still shouldn't gimp him too much. Although I imagine he may lose a single percentile of performance on that mobo due to the lanes.

 

[Naito]

I have a 2600k clocked at 4.2 GHz so i don't think it would "kill" the performance.

More that the 2600K doesn't support PCIe 3.0, but RCoon explained it well anyway.

 

[RCoon]

While the 2600K is a damn good i7 chip, it's still likely the baseline cause for lower scores. I think Firestrike is a poor performance measure for multi GPU setups, especially when considering that you're not using a 2011v3 or a Haswell refresh i7 chip like all the other benchmarkers. Trying using a game instead.

 

[FreedomEclipse]

2 970s for 1 1080p screen. Wasteful much?

 

[buildzoid]

The i7 2600k is running the GPUs on PCI-e 2.0 x8. That's 1/2 of what any Z77 or later board provides so that's probably most of your problem. The other issues can come from your OS CPU and RAM performance.

 

[RCoon]

The i7 2600k is running the GPUs on PCI-e 2.0 x8. That's 1/2 of what any Z77 or later board provides so that's probably most of your problem. The other issues can come from your OS CPU and RAM performance.

I don't think PCI is an issue at all...

Some additional information:
http://www.techpowerup.com/reviews/Intel/Ivy_Bridge_PCI-Express_Scaling/

 

[mull54]

I don't think PCI is an issue at all...

Some additional information:
http://www.techpowerup.com/reviews/Intel/Ivy_Bridge_PCI-Express_Scaling/

Parhaps It's just the combination of the gigabyte 970 and firestrike thats the problem?

 

[RCoon]

Parhaps It's just the combination of the gigabyte 970 and firestrike thats the problem?

Could you update your system specs for me? Just so I know what I'm making assumptions about hardware-wise.

 

[zsolt_93]

Those graphs are single card even if multigpu, there is the PLX chip on board of those. In multi card it might matter more with all the bandwidth shared. 2600K is pretty old and certainly not comparable with Haswell even if it is only 20% raw difference. With PCIE 3.0 being out for like 2 years 2.0 might not work perfectly, there were cards that wouldn't run perfectly in 1.1, those were 2.0 cards. But mostly the CPU is the problem, those people might be using Ivy/Haswell/E.

 

[mull54]

Could you update your system specs for me? Just so I know what I'm making assumptions about hardware-wise.

The same as in the spec sheet. But new air cooler and clocked to 4.2ghz. And a Soundblaster Z. Also a 120 evo ssd.

 

[RCoon]

And a Soundblaster Z

BINGOOOOOOOOOOOO
Is that a PCI-E Sound Blaster?

If you have an additional PCI-E card, it will take x4 lanes from your x8/x8 GPU's, making one of them run at x4. Try running the benchmark without the PCI card installed. I could be totally wrong, but these things can happen.