Graphics cards on Asus motherboard running at half speed problem, need help

Associate
Joined
24 Nov 2011
Posts
26
Hi,

I have a problem that my 2 x SLI NVIDIA GTX-590 graphics cards appear to be running at PCI-Express speed of x8 lanes rather than x16.

Indeed the NVIDIA driver control panel says PCI-E x8 Gen 2, for both graphics cards, rather than PCI-E x16 Gen 2.

I have an Overclockers Tyrannosaur PC computer Intel Core i7 2600K Sandybridge System with Asus Maximus IV Extreme-Z Intel Z68 Motherboard plus 2 x SLI NVIDIA GTX-590. Windows 7 x64 with 16GB RAM. AMI BIOS 1.5.1.1041

As instructed by Maximus IV Extreme-Z User's Manual (English) here:
http://www.asus.com/Motherboards/Intel_Socket_1155/Maximus_IV_ExtremeZ/#download

I've put one graphics card in slot 3 (PCIE_X16_2) and the other in slot 5 (PCIE_X16_4) to use the NVIDIA NF200 controller to give x16 lanes of speed. In slot 1 & 2 I have a OCZ RevoDrive3 and a firewire IEEE 1394 card.

I've also benchmarked how fast the CPU can read off a graphics cards. Comes to 2.4GB/s, I think for x16 performance I'm expecting somewhere near 5GB/s. This I did by writing some code in CUDA and using Parallel Nsight to at the profile reports.

Does anyone know why my graphics cards are running at PCI-E x8 speed instead of PCI-E x16?

Thanks for any help,
Jules
 
Hi,

Thanks for the reply but I am not sure I understand. The detailed specs for the graphics card here
http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-590/specifications

Quote "Designed for PCI Express 2.0 ×16 for a peak bandwidth (counting both directions) of up to 20 gigabytes (GB) per second (PCIe 2.0 devices are backwards compatible with PCI Express 1.x devices)"

Admittedly I think the specs are optimistic. Maybe 8GB/s theoretical maximum one-way and I'd expect 5GB/s in practice.

Unfortunately, as mentioned earlier, my tests show I am only getting 2.4GB/s maximum.

Is there something I am not understanding?
Jules
 
790c8cb1_mivepcie.jpeg


Someone can make sense of that, lol.
 
It's all about the motherboard.

With one graphics card it runs at x16 (slot 1).

With 2 graphics cards it runs at x8/x8 directly from the CPU (slots 1 & 3).

With 3 graphics cards it runs at x8/16/16 by using the NF200 chip (slots 1, 2 & 4).

I believe that Asus uses the x8/x8 configuration directly from the CPU as they consider it offers the same or better performance than having more PCI-E lanes but with the added latency caused by the NF 200 chip.

You could try putting the second card in the second or fourth x16 slot instead of the third.

This may not work at all and it may not make the 2 slots work at x16/x16.

You'd have to try it and see.
 
Last edited:
On a x58 R3E mobo I have GTX 590s in SLI in slots 1 and 3 which would normally give full x16/16 support. I then added a Revodrive to slot 4 which droped my GTX 590s down to x16/8. The difference in performance was barely messurable less than 1% in the unigene heaven benchmark.

My point is even for cards as potent as GTX 590s pci-e 2 @x8/8 you lose very little if anything.

I believe that Asus uses the x8/x8 configuration directly from the CPU as they consider it offers the same or better performance than having more PCI-E lanes but with the added latency caused by the NF 200 chip.

I have also read reviews where the above setup works better than using the NF 200 chip.
 
Hi Kaapstad

Yes, for games or similar, it doesn't make much of a difference because they use the built-in hardware rasterizers to draw. So things can mostly stay on the graphics card.

In my particular work, I'm combining ray-tracing with film effects, there is a lot of large traffic going to and from host CPU to the graphics cards' GPUs and back again, many times a second.

Your comment
I then added a Revodrive to slot 4 which droped my GTX 590s down to x16/8

Maybe that's the clue. I have my Revodrive in slot 1 and that could be causing the down-clocking to x8 for my graphics cards in slots PCIE_X16_2 and PCIE_X16_4?

I wish Asus were more flexible here and provided a way to force NF200 to engage. Or at least provide more documentation.


Doesn't look like there is an easy answer?

Jules
 
Unfortunately, as mentioned earlier, my tests show I am only getting 2.4GB/s maximum.

I remember something like this when I installed a Revodrive in a R3E board. When I first installed it I put it in a PCIe 1 slot and when booting it reported 2.4Gb/s speed. Then I moved it to one of the GPU slots which are PCIe 2. Then when it booted it reported 5.0Gb/s.

When your Revodrive boots up in slot 1 does it report 5.0Gb/s which it should. It will not matter if the slot is running at x8 or x16 your revodrive should report 5.0Gb/s. If it only reports 2.4Gb/s the slot is running in PCIe 1 mode. This maybe the problem with your GTX 590s as well. I am not familiar with your mobo but on some mobos you can select in the bios what PCIe standard they run at.

If I have got it right the speed you get 2.4Gb/s 5.0Gb/s or 8.0Gb/s refers to wether you are running PCIe 1 PCIe 2 or PCIe 3. Not if it is x8 or x16

I do not know which type of revodrive you use but the above refers to a revodrive3 x2. A revodrive3 I think uses PCIe 1 and would report 2.4Gb/s when booting.
 
Last edited:
790c8cb1_mivepcie.jpeg


Someone can make sense of that, lol.

Oh wow, is that how the PCI-E lanes are given on the board? What a bloody mess!

Ok, from reading this, with a wild guess, if you want x16/x16 SLi, both cards needs to be in slot 1 and slot 2, or slot 1 and slot 4. Using slot 1 + 3 will cause it to go x8/x8, however this may be better as the signals goes through less chips and in turn means less latency.

And if you have a Revodrive, connect it to the PCI-E x4 (bottom block slot), since this slot is controlled by the Z68 chipset and won't steal any lanes from the CPU.
 
Quote "Designed for PCI Express 2.0 ×16 for a peak bandwidth (counting both directions) of up to 20 gigabytes (GB) per second (PCIe 2.0 devices are backwards compatible with PCI Express 1.x devices)"

PCIe 2 = 5Gb/s x 16 lanes x 2 (both directions) = 160Gbits/s or 20 gigabytes
 
Hi Kaapstad

Yes, for games or similar, it doesn't make much of a difference because they use the built-in hardware rasterizers to draw. So things can mostly stay on the graphics card.

In my particular work, I'm combining ray-tracing with film effects, there is a lot of large traffic going to and from host CPU to the graphics cards' GPUs and back again, many times a second.

Your comment


Maybe that's the clue. I have my Revodrive in slot 1 and that could be causing the down-clocking to x8 for my graphics cards in slots PCIE_X16_2 and PCIE_X16_4?

I wish Asus were more flexible here and provided a way to force NF200 to engage. Or at least provide more documentation.


Doesn't look like there is an easy answer?

Jules

You're not going to be able to get SLI at x16/x16, it's a limitation of Sandybridge and the Z68 chipset and exactly the same limitation on the number of lanes on Ivybridge and Z77 also. Natively you only have 16 PCIE lanes available for use.

Sure the NF200 muliplexes and offers more lanes but it's another hop thats added to the PCIE bus which impacts the performance again and you're back to where you started. There are newer PCIE multiplex solutions around on other boards and NF200 does seem somewhat gimped. The broad consensus is that you're just as well to live with x8/x8. Also, looking at the slot configuration on the Extreme-Z they all have single slot spacing between 2, 3 and 4, which means using double width cards limit which slots and how many slots you can and can't use. I think this is done deliberately so that you have to use slot 4 in certain configurations which then forces NF200, for all the good it does - but I could be wrong.

If you want x16/x16 then ideally you should be looking at a different platform such as Sandybridge-E and X79 chipset. You really don't have the right solution for what you seem to require and I'm faced with the same conumdrum myself at the moment.

Also, I notied some folk still seem to be banging on about Ivybridge-E but what they neglect to tell you (probably because they haven't done their homework) is that there won't be an Ivybridge-E. Maybe an EP for a Xeon shrink but thats all. Haswell is supposed to launch sometime after Q2 in 2013 and the delayed IB launch has basically meant that Intel forced to drop any IB-E plans in order to stay in tune with it's 'Tic Toc' production cadence further down the line.
 
x16 , x16 on 1155 needs added chip as CPU does not have enough lanes. For full X16 X16 support X79 is the best route.

As a side point Andre Yangs Unigen world record is set with the ASUS Max V formula Z77 with 3 x 7970 at 8x 4x 4x so does bandwidth even limit the 7970 in tri fire it seems not. Especailly as Unigen is a GPU limited bench not like Vantage and earlier 3D marks.
 
Here is my bios screen

bios_screen.jpg


It says x16, x16... yes! Unfortunately, in Windows from the NVIDIA control panel

NVIDIA_sys_info.jpg


x8, x8. Something is not telling the truth. Strange? My tests show that the maximum throughput is 2.4GB/s, indicating the latter x8 speed.

Note the revodrive3 x2 in slot 1 does report booting up 5Gb/s.

I don't know :( Thanks everyone for the help.

Jules
 
Clutching at straws what happens to your GTX 590s if you remove your firewire IEEE 1394 card

The reason I mention this is the slot it is in would normally be covered by a card in a 3 way sli setup.
 
Last edited:
If you're serious about wanting max PCIE x16/x16 why are you persisting with this setup? It's not the ideal solution - your requirements seems to dictate X79 and not this.

For what it's worth, I suspect that the nvidia app is not "NF200 aware" and is purely reporting the lanes it is using directly off the CPU, which will only ever be x16 in total. Lets not forget you're also using dual GPU cards too not to mention a storage device also. I'm not sure why you seem surprised by the result.

NF200 is a cludge at best and at worst you could even describe it as snake-oil. What 8 Pack says is correct, you really need an X79 platform if it matters that much to you.
 
Back
Top Bottom