**SLI FAQ**

[Gothgar]

I have seen about five hundred threads pop up about SLI, so I created this one to try to help with that.

Here are some article you should read before asking us questions about SLI.

nVidia FAQ

Anand's Article

Basically, SLI will work on any two similar video cards provided they have the same BIOS, if you keep the brands the same, it will increase your success rate quite a bit.

You have to buy an SLI capable motherboard to utilize SLI, and then two of the same video cards (see above.) [Exception: Gigabyte's 2x SLI bridged card]

You will need a lot of power for SLI. Don't skimp on power if you are willing to spend 500 or more on video cards, get a trusted brand, Antec, Enermax, PCP&C, etc. etc.

Asus SLI Power Recommendations that shows what the power requirements are

MSI asks for at least a 450 watt powersupply for SLI

When using the two video cards in SLI, you cannot use more than two monitors, as in, the video capibility on the second video card is disabled, you can disable SLI however and run 4 monitors, if you so chose.


Please bring any other questions here, and I will try to help out with them, I am not an expert, but I have read a lot about it, others feel free to chime in about anything.

If you feel the need to ask whether or not SLI is worth it, no one can answer that for you. Research and learn about it, then decide if it is for you or not.

 

[Blain]

But if I read your SLI thread, I won't need to start one of my very own. :shocked:

 

[Vegito]

your forgetting the gigabyte single pcie sli card.. no need for sli motherboard

sli bridge on 1 card

 

[Gothgar]

Originally posted by: forcesho
your forgetting the gigabyte single pcie sli card.. no need for sli motherboard

sli bridge on 1 card

thanks, updated with that info, and some power reccomendations

 

[Omizzle]

OMG j00 won a free bump!

 

[BenSkywalker]

Please bring questions here!

One thing I've been wondering(ignoring AFR)-

If you have a shader on the lower half of the scene that requires light interaction from multiple light sources on the top half of the screen, how is the SLI configuration dealing with that at the driver level? Is it requiring duplication of all data- or are any requisite inputs simply handed over to the dependant GPU in terms of interactions by using some sort of lookup based on vector data? This leads in to my second question- What happens if you have light sources on one side of the screen that are then impacted by a shader which then interacts with another shader and two of these are being rendered by one GPU and the other by the other GPU? For instance, say you were to have a light shining through a stain glass window and that colored light source was interacting with water below, how does the modified data of the light make it to the second GPU? Is it by 'request' or is the data pool shared/duplicated at the board level? Hope you can clear this up as I've been trying to figure out what would be the best approach to dealing with these issues. Seems like you would be dealing with a bit of latency with back and forth communication, although I suppose in shader limited situations the payoff would more then cover any minor hit it would incur.

 

[fkloster]

Originally posted by: Gothgar

Originally posted by: forcesho
your forgetting the gigabyte single pcie sli card.. no need for sli motherboard

sli bridge on 1 card

thanks, updated with that info, and some power reccomendations

This is incorrect... I was told that you still need an SLI motherboard to utilize a single-card DUAL GPU solution....

 

[fkloster]

Randabis
Junior Member

Posts: 16
Joined: 12/26/2004
quote:

--------------------------------------------------------------------------------
Originally posted by: Zebo
I believe gigabyte has a dual 6800 GPU card...watch for it..

Most of us will opt for dual ultra card with single PCIe.
--------------------------------------------------------------------------------



Except for the fact that you cannot run these dual cards without an SLI motherboard. The Gigabyte 6600GT dual will cause systems without SLI to not POST at boot up.

-------------------------
I'm mommy's little nightmare driving daddy's car around.

 

[Matthias99]

Originally posted by: BenSkywalker

Please bring questions here!

One thing I've been wondering(ignoring AFR)-

If you have a shader on the lower half of the scene that requires light interaction from multiple light sources on the top half of the screen, how is the SLI configuration dealing with that at the driver level? Is it requiring duplication of all data- or are any requisite inputs simply handed over to the dependant GPU in terms of interactions by using some sort of lookup based on vector data? This leads in to my second question- What happens if you have light sources on one side of the screen that are then impacted by a shader which then interacts with another shader and two of these are being rendered by one GPU and the other by the other GPU? For instance, say you were to have a light shining through a stain glass window and that colored light source was interacting with water below, how does the modified data of the light make it to the second GPU? Is it by 'request' or is the data pool shared/duplicated at the board level? Hope you can clear this up as I've been trying to figure out what would be the best approach to dealing with these issues. Seems like you would be dealing with a bit of latency with back and forth communication, although I suppose in shader limited situations the payoff would more then cover any minor hit it would incur.

AFAIK, all memory accesses are done by one GPU, and any data requested is sent across the SLI link to the other one (and any data produced by the other GPU is sent back across the link to be written to the first card's memory). This maintains coherency in the memory, and handles the 'what if you need data from the part of the scene being drawn by the other GPU' question. This is also why you do not get an increase in effective bandwidth from using SLI.

WRT multiple shaders: they interact the same way they would on a single card. You can only run one shader at a time; the effects you describe with multiple shaders would have to be done in multiple passes, and then there's no problem, because the SLI system enforces memory coherency. Also, from the descriptions I've seen, the load is split up on a per-pixel (or at least a per-line level), not such that one GPU renders one shader and the other renders another one.

 

[KoolDrew]

How exactly would overclocking SLI work? Will you have to have both cards at the same speed when overclocking to work properly? How would overclocking work?

 

[Gothgar]

Originally posted by: BenSkywalker

Please bring questions here!

One thing I've been wondering(ignoring AFR)-

If you have a shader on the lower half of the scene that requires light interaction from multiple light sources on the top half of the screen, how is the SLI configuration dealing with that at the driver level? Is it requiring duplication of all data- or are any requisite inputs simply handed over to the dependant GPU in terms of interactions by using some sort of lookup based on vector data? This leads in to my second question- What happens if you have light sources on one side of the screen that are then impacted by a shader which then interacts with another shader and two of these are being rendered by one GPU and the other by the other GPU? For instance, say you were to have a light shining through a stain glass window and that colored light source was interacting with water below, how does the modified data of the light make it to the second GPU? Is it by 'request' or is the data pool shared/duplicated at the board level? Hope you can clear this up as I've been trying to figure out what would be the best approach to dealing with these issues. Seems like you would be dealing with a bit of latency with back and forth communication, although I suppose in shader limited situations the payoff would more then cover any minor hit it would incur.

very good question, didn't expect ones like this, I was just trying to reduce the amount of threads in GH on SLI, but here you go.

SLI or Scalable Link Interface works in a few different ways. rendering is handled a few different ways.

First, Split-frame rendering, simply one graphics card renders all graphics on half the screen, and the other graphics cards renders all graphics on the other half of the screen. This is the mode that the card should perform better in games. What the driver does is split upthe pixel fragmentations that need to be rendered will only be sent to the GPU that will be rendering them, therefore the first video card won't even know what is getting written onto the other half of the screen and vise versa. Plus, using SFR the driver will theroretically load balance the rending process, so say you are running through water, the lower part of the screen will have more to render than the top half, so one GPU could be processing the top 60% of the screen, while the other one, working just as hard is processing the bottom 40%. The problem with SFR is that with these load balancing calculations and sycronization calculation (so the right image from each GPU hits the screen in sync) take away from the processing power, so it really doesn't perform as well as two full cards.

Secondly, alternate frame rendering does its rendering a very different way, as it sounds, it alternates frames back and forth between the two video cards, this puts less overhead into the equation because half the calculations aren't needed in this mode, however, this is a more CPU intensive mode, as both GPU's with be processing frames (in the same way a single video card would)

Seeing how these work, should have answered your question(s)

edit: Cliff Notes
SFR = more overhead for the video cards
AFR = More latency

 

[Gothgar]

Originally posted by: KoolDrew
How exactly would overclocking SLI work? Will you have to have both cards at the same speed when overclocking to work properly? How would overclocking work?

both cards are overclocked simultaneously, meaning, if you overclock one x, you have to voerclock the second one x