SLI/Xfire effective Vram

[PCBudgetSolutions]

Editing this because it's confusing.

here is 2 ways SLI/Xfire work. Depending on the application however I believe option 2 is more common.

ption 1: Both GPU's render 1 frame, however each GPU renders HALF of the frame effectively getting it done up to twice as fast. Each GPU renders part of the frame independently of each other.

Option 2: Each GPU renders alternate frames. Master renders off frames and secondary renders even frames. Again each GPU renders independently of each other.

In scenarios of option 1 people get confused thinking the GPU's render the same frame. How would you see an fps increase there? Spoiler there wouldn't be one. The GPU's cannot share Vram adding it to 6GB effective, but because how SLI/Xfire works, they can render more than the max Vram of 1 card at a time. Each card is independent of each other when rendering.

Master card is rendering frame 1. Total Vram usage of frame 1 is 2.1GB.

Secondary card is rendering frame 2. Total Vram usage of frame 2 is 1.8GB

Total amount of Vram being used for 2 frames is 3.9GB. Now as frame 1 is half way done frame 2 starts and etc. Because in events where you get 60 FPS, each card is rendering 30 FPS. The Vram will not get bottle-necked or bogged down because if 1 HD 7970 can do 45 fps, and 2 are doing 60 fps, you are not stressing it as much. Where the advantage is, is if 1 card gets 30 fps and in a perfect world 2 cards get 60 fps. If each frame has exactly 3GB of textures and such to render, then 1 card will do 90GB of textures per second. If 2 cards do 60 FPS with 3GB of textures and such per frame, thats 180GB of textures rendered per second. Now it doesn't really work this way but I am using this model to prove my theory. So do you get 6GB of Vram? No, but you are rendering twice the textures in the same period effectively doubling the performance. So to say Vram is only that of 1 card seems very off to me. 3GB of Vram x2 cards x 60 fps =180GB of textures vs 90GB of textures from 1 card.

[Godlygamer23]

You're still getting 3GB worth of VRAM. Effective is effective no matter how it's used. Both cards consume the same amount of VRAM at the same time. It never differs. 

[ltguy]

This post hurts my head.

[PCBudgetSolutions]

You're still getting 3GB worth of VRAM. Effective is effective no matter how it's used. Both cards consume the same amount of VRAM at the same time. It never differs. 

But you can compute 6GB at 1 time.  A be it 2 separate frames.

[PCBudgetSolutions]

It's very misleading to say that tho.

[Godlygamer23]

But you can compute 6GB at 1 time.  A be it 2 separate frames.

They both store the same textures and information AFAIK, unless you have a source to prove me wrong.

[ltguy]

But you can compute 6GB at 1 time.  A be it 2 separate frames.

You're missing the point, whats on the first 3gb of memory, is a mirror image of whats on the 2nd 3gb. There is no additional memory gain. 

[PCBudgetSolutions]

They both store the same textures and information AFAIK, unless you have a source to prove me wrong.

SLI/Xfire render separate frames in sequence.

[ltguy]

From amd's website

Q: Does ATI CrossFire™ technology double the memory? 

A: No, ATI CrossFire™ technology does not double the memory,  If you have a 512MB card, adding another 512MBcard will not increase your total memory to 1GB; it still will be 512MB.

 

http://support.amd.com/us/kbarticles/Pages/ATICrossFire_comp_req.aspx

[PCBudgetSolutions]

You're missing the point, whats on the first 3gb of memory, is a mirror image of whats on the 2nd 3gb. There is no additional memory gain. 

They don't render the same frame???

[PCBudgetSolutions]

From amd's website

Q: Does ATI CrossFire™ technology double the memory? 

A: No, ATI CrossFire™ technology does not double the memory,  If you have a 512MB card, adding another 512MBcard will not increase your total memory to 1GB; it still will be 512MB.

 

http://support.amd.com/us/kbarticles/Pages/ATICrossFire_comp_req.aspx

I never said the memory doubles, but it's also misleading.  Since you have 2 cards rendering separate frames at once, each card is using their memory independently of each other.

[ltguy]

I waving the white flag, I posted the answer and the source, i can do no more.

[PCBudgetSolutions]

I waving the white flag, I posted the answer and the source, i can do no more.

It's why micro shutter/studder exists.  I learned this concept from PCPer.  Unless he is wrong.

[PCBudgetSolutions]

I waving the white flag, I posted the answer and the source, i can do no more.

http://www.overclock.net/t/1210604/how-does-sli-work

[IGJoe2192]

I waving the white flag, I posted the answer and the source, i can do no

 

It's why micro shutter/studder exists.  I learned this concept from PCPer.  Unless he is wrong.

 

 

What you are missing is. both cards are using the same amount of VRAM with the same textures loaded, therefore it does not increase the total capacity to 6GB, it stays at 3. It doesnt increase the amount of textures that can be loaded, it just allows the GPU's to share the work load of computing the images hints the performance gain. my last 3 systems have been Crossfire based. 

[Godlygamer23]

I run folding@home on my one GPU. Even though my one GPU is only doing the work, they both get the same exact textures loaded into them, because they are in SLI.

[PCBudgetSolutions]

What you are missing is. both cards are using the same amount of VRAM with the same textures loaded, therefore it does not increase the total capacity to 6GB, it stays at 3. It doesnt increase the amount of textures that can be loaded, it just allows the GPU's to share the work load of computing the images hints the performance gain. my last 3 systems have been Crossfire based.

They don't load the same textures though. Generally speaking most games are coded for alternating frames for SLI/Xfire rather than split frames. Each card in these cases can independently use up to 3GB of Vram per frame. Yes in situations where you split frames, you are capped at 3GB of Vram but benefit from having each frame rendered up to twice as fast.

EDIT: To add however, In cases in alternate frame rendering, if 1 HD 7970 loads 3GB's of textures @ 45 FPS, 2 HD 7970's will render roughly 1.8-2GB of textures per card @ 60 fps. However, 2 cards are not fast enough to process 3GB of textures per card w/o being bottle-necked. Where the advantage is, since each card only renders 1/2 the frames, the build of textures and such is much lower and can render them faster. Effective reducing the amount of Vram eat up per card.

[PCBudgetSolutions]

I run folding@home on my one GPU. Even though my one GPU is only doing the work, they both get the same exact textures loaded into them, because they are in SLI.

I don't believe current programs/games use split rendering anymore.  Even so in 1 frame each GPU renders HALF of the frame and HALF of the textures.  Reduces the load to HALF IE HALF the Vram used per frame.

[Godlygamer23]

From NVIDIA's website: 

When I configure two graphics cards in SLI mode, do the graphics cards work together to create double the memory size?

No. In SLI mode, each graphics card uses its own frame buffer memory to render a 3D application. The operating system will report a graphics card frame buffer memory size that is found on a single graphics board.

 

http://www.geforce.com/hardware/technology/sli/faq#c26

 

[PCBudgetSolutions]

 

From NVIDIA's website: 

When I configure two graphics cards in SLI mode, do the graphics cards work together to create double the memory size?

No. In SLI mode, each graphics card uses its own frame buffer memory to render a 3D application. The operating system will report a graphics card frame buffer memory size that is found on a single graphics board.

 

http://www.geforce.com/hardware/technology/sli/faq#c26

 

 

I am not saying 2 3GB cards means 6GB of textures.  But what I am saying is that saying you only have 3GB of effective memory is misleading.  Because the way SLI/Xfire work, cards work independently of each other and render their own textures/frames.

[IGJoe2192]

okay, so the answer you are looking for is, you have 6GB of effective memory, BUT, your frame buffer is still only 3GB. /end

[laxer]

I apologize as I don't really have time to answer this properly but, my understanding is there is memory coherency going on.

 

Additional information I pulled up quickly on google:

 

http://web.mit.edu/6.173/www/currentsemester/handouts/L10-cache-coherence.pdf

http://en.wikipedia.org/wiki/Cache_coherence

 

This is my understanding of it: Each GPU is loaded with the same data, that is the vRAM on each gpu contain the same instruction/data set. Since the processors on the cards are each independent they can theoretically get through this work twice as fast. It however is not that simple, since there are two entities working on the same data set there has to be some overhead for them to communicate between one another.

 

I could get into models for parallel computing but I don't think it is necessary.

 

The bottom line to my understanding is that yes all 6GBs(for example) are in use however, they contain duplicate information so AT BEST you have 3GBs of useful memory. (It's not uncommon for multiple sets/orphan instructions to occur reducing this number)

 

Let's see what others think, the above is based off my knowledge of parallel computing and shared memory systems. It is therefore an assumption.

[PCBudgetSolutions]

okay, so the answer you are looking for is, you have 6GB of effective memory, BUT, your frame buffer is still only 3GB. /end

Depends on the program/game but kinda sorta.  If you are rendering 30 frames per sec per video with 3GB worth of textures I don't know how well that will work.

[PCBudgetSolutions]

I apologize as I don't really have time to answer this properly but, my understanding is there is memory coherency going on.

 

Additional information I pulled up quickly on google:

 

http://web.mit.edu/6.173/www/currentsemester/handouts/L10-cache-coherence.pdf

http://en.wikipedia.org/wiki/Cache_coherence

 

This is my understanding of it: Each GPU is loaded with the same data, that is the vRAM on each gpu contain the same instruction/data set. Since the processors on the cards are each independent they can theoretically get through this work twice as fast. It however is not that simple, since there are two entities working on the same data set there has to be some overhead for them to communicate between one another.

 

I could get into models for parallel computing but I don't think it is necessary.

 

The bottom line to my understanding is that yes all 6GBs(for example) are in use however, they contain duplicate information so AT BEST you have 3GBs of useful memory. (It's not uncommon for multiple sets/orphan instructions to occur reducing this number)

 

Let's see what others think, the above is based off my knowledge of parallel computing and shared memory systems. It is therefore an assumption.

There is 2 way SLI/Xfire work:

 

Option 1: They render 2 halves of the same frame, similar to how to hyperthreading but much more efficient

Option 2: Each GPU renders every other frame independently, however there is some performance loss putting the frames together (so not a 100% gain)

 

I think most newer games utilize option 2 rather than option 1.  Each GPU does NOT render each COMPLETE frame because that would yield ZERO performance gain...

[laxer]

There is 2 way SLI/Xfire work:

 

Option 1: They render 2 halves of the same frame, similar to how to hyperthreading but much more efficient

Option 2: Each GPU renders every other frame independently, however there is some performance loss putting the frames together (so not a 100% gain)

 

I think most newer games utilize option 2 rather than option 1.  Each GPU does NOT render each COMPLETE frame because that would yield ZERO performance gain...

 

I see what your saying and am intrigued but, I do not have a factual answer.

 

I can dig through some textbooks/online and see if I can find anything later if no one is successful here. The only problem is the most recent text books I have only go up to 2011 making the knowledge limited to 6000 series and 500 series.