Strange Combined Test Issue - Please Help

[QueueCumber]

I'm getting strange combined scores. The results are listing my system as low compared to similar systems. I can understand this from the perspective that some people with 4-way SLI are overclocking their CPUs to 5GHz or more while mine is at 4.6 to 4.7 GHz. Here is what is strange. With 3-way SLI it is saying my system is not low in comparison, but it only changes when I put the 4th card into my system (GTX 680s). Likewise, my FPS on the combined score goes down when putting in the 4th card even if I have the 3-way CPU set to 4.6 while the 4-way CPU is set to 4.7 (both in the same exact system but with one extra card added). I don't understand why the score would go down instead of either staying the same or going up?!

Here is a comparison of one of my 3-ways with a 4-way (3-way @ 4.6 GHz and 4-way @ 4.7 GHz):

GraphicsTest1
171.11
142.51
FPS

---

GraphicsTest2
196.36
148.17
FPS

---

GraphicsTest3
247.33
179.77
FPS

---

GraphicsTest4
113.14
86.9
FPS

---

PhysicsTest
48.49
47.35
FPS

---

CombinedTest
54.76
57.13
FPS


Here are links to the scores:

Comparison:
http://3dmark.com/compare/3dm11/3187896/3dm11/3182308

4-way:
http://3dmark.com/3dm11/3187896

3-way:
http://3dmark.com/3dm11/3182308

Can anyone explain to me what is going on with this?

Thanks.

[Jarnis]

The PC Health check is not very accurate with 3/4 card configs, far too little data. It also has some other issues... we're trying to fix it in the near future but at the moment I wouldn't put too much weight on what it says.

Beyond that, Combined Test is a complicated one on multi-GPU setups and a lot depends on the video drivers. On Multi-GPU systems the GPU simulation is duplicated across the cards and on some configurations the system sometimes seems to fail. This indicates an issue with the drivers not doing exactly what the benchmark asks for.

[QueueCumber]

The PC Health check is not very accurate with 3/4 card configs, far too little data. It also has some other issues... we're trying to fix it in the near future but at the moment I wouldn't put too much weight on what it says.

Beyond that, Combined Test is a complicated one on multi-GPU setups and a lot depends on the video drivers. On Multi-GPU systems the GPU simulation is duplicated across the cards and on some configurations the system sometimes seems to fail. This indicates an issue with the drivers not doing exactly what the benchmark asks for.

Can you think of any reason why the combined test would be getting lower when using a 4-way vs. a 3-way setups on my system?

Thanks.

[Jarnis]

Can you think of any reason why the combined test would be getting lower when using a 4-way vs. a 3-way setups on my system?

Thanks.

"video drivers are Doing it Wrong, causing a performance hit" would be the prime guess.

Edit: Another possibility is that you are hitting a CPU bottleneck - more GPUs equals more physics simulations in combined test and there is always some driver overhead from all that. Usually it shows as less-than-100% scaling with GPU count, but I guess in extreme cases it could be a lot worse. The test is very complicated and somewhat of an "ultimate hardcore test" for multi-GPU setups with complex DX11 computation features.

It is also the test that pushes the system most overall (peak power consumption) and I guess it is possible that something is throttling under the load?