SLI / CrossFire FAQs

Page 49 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Status
Not open for further replies.

JaxAxRho

Honorable
Aug 7, 2015
249
0
10,710
So I read through this as I am interested in possibly using crossfire. My mobo supports 1x16, 2x8, 1x8x2x4. I was hoping that the OP might have some specs of these against each other. I did not read through all 28 pages, but is there an updated place that we can see the 1x16 vs. 1x16x8 vs. 1x16x4 vs. 2x8 vs. 1x8x4 vs. 2x4? If not I'd be glad to do some more digging, but perhaps OP could be updated to 2014/2015 mobo and specs?
 
The author of this thread is currently way, but I will try to help you.

First off if you have 3 slots and only plan to use to cards 2 x 8 slots will be best, but the last slot will most likely have to be left empty.

The reason for these combinations is because a system only has so many pci-e lanes to use.

So if you have 16 lanes(some boards have more or even less to work with. Boards with less should not have SLI or crossfire as an option clearly), you normally break them up just like your board.

Single card, use all 16
2 cards, use 8 per slot.
3 cards use 8 for the top card and split the remaining 8 lanes for last 2 cards. PCI-e does not work on other non standard configs. They can not go 5 per card for instance.

Now how much performance does a lower number of links cause? well for that I recommend you look at these article from techpowerup. The first one is a bit older, but the cards are very fast and it gives you an idea of what to expect. These are single cards tests as well. If you would not loose much in a single you should not loose much in a multi card setup(again all but the fastest cards on the market).

Ivy Bridge PCI-Express Scaling with HD 7970 and GTX 680

and

GeForce GTX 980 PCI-Express Scaling

You will notice these tests also have different revisions of PCI-e. This was important when people moved from 2.0 to 3.0 because many worried that the new 3.0 cards would be too slow on 2.0 systems. It takes a pretty fast card to require more bandwidth.

I hope this helps clear some things up.
 

JaxAxRho

Honorable
Aug 7, 2015
249
0
10,710


It does indeed. I actually found the GTX 980 article and scanned through it. Makes a lot of sense. I think the most interesting was the Tomb Raider comparison. Obviously small overall percentages, but PCI3 x4 3.0 took the highest frame rate at the highest test resolution. I'd still like to see a side by side of 1x16 vs. 2x8 vs. 2x4, but this was very informative for each individual slot itself. Pretty sure I found some benchmark testing of 1x16 vs 2x8 but I cannot find it again. I'd love to try some testing out myself, but I don't have extra mobo's or the cash for said mobo's lying around to make benchmarks.
 

JaxAxRho

Honorable
Aug 7, 2015
249
0
10,710


I think 5 years would be recent enough to compare, though I would like to see some numbers with the 3rd gen i Intel processers. Sandy bridge is not too new, but still well used.
 

11skip11

Honorable
Aug 6, 2015
12
0
10,510
Hey, i'm a little confused. Is it possible to use a sli bridge on only two exactly the same gpu's like GeForce GTX 760 and another 760 gpu, or can you use it only on members of the same family like a 760 and a 780 Ti?

Another question... I have a motherboard that only supports crossfire (asrock z77 pro4) but the original post mentioned something about z77 being a newer version and supporting sli so i'm quite confused about that too?

And one last thing (this might sound stupid) but is it really necessary to have a motherboard that supports sli because it's only a little cable, not even connected directly to the motherboard. Only attached to the gpu's.

Hope someone can provide me with an answer
 

mmaatt747

Distinguished
Sep 26, 2011
563
0
19,060
SLI is meant for using the same gpu. The bridge itself is not the limiting factor, it's the SLI technology that wont' allow different GPU's. The motherboard DOES have to support SLI. It's based on the PCIe slots on the motherboard and how they communicate with the CPU, not the SLI bridge you use to connect the cards.
 
That is correct.

In some circumstances you can SLI video cards with different names, but that is because they are re brands(so the same core, but named something different).

On the crossfire front, you have some interesting pairings too.
 

Tehseus

Reputable
Sep 29, 2015
3
0
4,510


Hi there! I have a question.
Assuming there is a LGA1151 Z170 motherboard
3 PCIe 3.0 slots are directly connect to CPU
They can work at x16/0/0 or x8/x8/0 or x8/x4/x4 mode
SLI does not support 3-way SLI on x8/x4/x4 mode
So the only config is 2-way SLI on x8/x8 and leaving the third slot empty
Now, the Z170 PCH has a maxium of 16 PCIe 3.0 lanes can be enabled at any time
Although the PCH's Datasheet says that they can config as x1 or x2 or x4 lanes only
I'm still waiting for reply from Intel that if these PCIe lanes could config as a x8 or even x16 slot
But if it is possible to config as one or two x8 or one x16 slot
does SLI work on this/these PCIe 3.0 slot come from PCH?
like:
2-way SLI on x16/0/0/x16(PCH)
3-way SLI on x16/0/0/x8(PCH)/x8(PCH)
4-way SLI on x8/x8/0/x8(PCH)/x8(PCH)
is that possible or not?
Thanks!
 

Vogner16

Honorable
Jan 27, 2014
598
1
11,160
just wanted to say this for you guys to think about.

I had cfx 7870's but I found the power of the gpu was greater than that of what the frame buffer was capable of handeling.

for example when maxing out textures on games like acu and bf4 I would get 55 to 60 fps, but I would get constant stutters as the memory pool was at 100% and couldn't handle the amount of textures I was using regardless of the gpu's ability to render the frames.

I switched to a 7970 and now I get less fps and 0 stutters.

this point is really relavent when you consider cards like the 970. sli of them could power a full 6 gb of textures at 40 fps easy but the 3.5 gb frame buffer will restrict you to lowering the settings and getting stupid high fps (resulting in screen trearing and or frame caps like vsinc therefor making the second card pointless.

the only gpu I know of that has the ram to handle the power they have in cfx/ sli are either the titan x or r9 390x as they both have what seems to be stupid amount of vram, but when you double the power of the gpu it is a perfect amount of vram. the 980ti could be considered for this list too but none is as perfect for crossfire as the 390x. this card is the perfect crossfire card. xdma has also proven to have better scaling than the outdated SLI bridge to greatly improve valve to perf ratio of amd cards in cfx.

 

I think this may be upto the board designer(licensing issues may apply as well, but I am not sure what they are doing now days). I am also not sure if the chipset lanes have any extra latency or other concerns. those extra slots would also pay off for PCIe SSDs.


Game to game you will certainly find this to be a true issue. It is also one that may become more of an issue in the future as games want to use more vram.
 

Tehseus

Reputable
Sep 29, 2015
3
0
4,510

Thanks for your reply!
I did some search then I found this article:
http://www.hardocp.com/article/2010/08/25/gtx_480_sli_pcie_bandwidth_perf_x16x16_vs_x4x4
Two GTX480 cards were forced working on x4 mode by taping the PCIE pins, and they were SLI Enabled!
How did that working?
Is it possible that, geforce drivers detect the PCIE slots whether they meet the requirements or not, like:
at least x8 links
all are 2.0 or 3.0
all come from CPU
but drivers doesn't care what link speed the GPU actually on?

 
The drivers do not care about link speed.

They only care that it is an SLI ready board. This is so much so to the point that modified drivers have been used in the past to get SLI running on a non SLI board.

The performance loss for lets say 8x + 8x vs 16x + 16x is very minimal. Multiple tests have been done to show that single gpu cards are not bandwidth starved yet.

X4 on pci-e 3.0 is as fast as the first generation x16 slots, but I am sure it is not quite optimal(3 card setups do not have the same scaling as 2 cards in most cases either.). This may be the reason boards do not tend to use 8 + 4 + 4.

I can not be sure about using pci-e lanes from the chipset and cpu together. It may or may not add some latency. Only a board maker or engineer can answer this question properly(I am neither).

From time board makers have AMA events here on Tom's. These ask me anything events would be the perfect time to ask the manufacturers these kinds of questions for clarification.
 

SoapPL

Reputable
Dec 2, 2015
2
0
4,510
I don't quite understand. It is possible to SLI Gtx 970 and Gtx 980?
How many can I have screens when I build 2-way 970 ?
 

Vogner16

Honorable
Jan 27, 2014
598
1
11,160


NVidia has always and more than likely always will require you to have as near to identical gpu for an SLI configuration as possible.

for example. I don't think you can do a 2gb 960 with a 4 gb 960 unless the drivers disable the extra 2gb on the larger storage card.

essentially, NO you cant put a 970 with a 980.

you also will NOT get a perfect 100% performance boost.
if you get 30 fps on a 970 expect 40-50 on dual 970's and only on games that support it.

having 2 cards does not double the Vram. SLI requires that both cards have the same data stored on each one meaning that even if you have 3 970's (with 3.5gb each) you will still only have a total 3.5gb of vram for the game to use. (this will be a bottleneck)

end of the day dual GPU setups are only usefull to those who want a machine faster than the best single gpu setup available.
Every dual gpu system WILL have problems. games will crash due to SLI. some will even have worse fps than if you disabled the other card, but if you want faster than a TITAN X then this is what you have to deal with.

I do NOT recommend a dual 970 setup to anyone.

EDIT: dual gpu setups support the same amount of monitors as a single gpu setup. ususally through a DP daisy chain with single gpu and with dual gpu there are plenty of on PCB ports.
 

SoapPL

Reputable
Dec 2, 2015
2
0
4,510
Thank you for answer, this is what I want to hear.
By directx 12 you will not get 3,5 GB + 3,5 GB VRAM?
 

Vogner16

Honorable
Jan 27, 2014
598
1
11,160


directx is a API used to decode video game code and turn it into something the graphics card can read.

SLI is a technology implemented into the driver by NVidia to make 2 graphics cards work on one compute task.

memory buffer on a graphics card is to provide instant data to the GPU so that it can render the 360 environment within a certain distance of the user. the more memory you have the more distance around the user can be loaded into the ram and experience less texture pop and enable more tessellation etc...

when you add another graphics card to the system it doubles the gpu cores to work with but separates the memory.

for example: a graphics card has 2000 cores and 4 gb memory on that card. it can access that memory fast enough to provide the performance needed for playable framerates.
If you had 2 graphics cards with 1000 cores each and 2 gb on each one, now you only have 2 gb with a fast connection to the cores. if you were to compile some data that gpu 1 needs on gpu2's memory then it would need to use the PCI lanes speed to transfer memory data as fast as if it had been simply soldered to gpu1's board. the PCI lane is not fast enough for this.

It is fast enough to do frame alternation where gpu1 renders frame 1 then gpu2 does frame 2 and back and forth etc... but that is given both gpu's have the same data on each set of vram. try to split it up and performance dumps even worse than frame alternation does (AFR).

there are many other methods to use dual gpu's but sharing ram is found to be one of the slowest and therefor NVidia and AMD do not use it.

directx 12 like all other API's does not mess with how NVidia or AMD want to implement dual gpu config and therefor will use the method that the gpu manufacture has found to be fastest for their specific graphics card.

simple answer: by adding a second graphics card you will double the core count of graphics processing but will NOT double the storage amount of the VRAM regardless of the API used. 970 will have 3.5gb ram and so will a 970 SLI config.

EDIT: I stress again. 3.5 gb is NOT enough vram for a dual 970 setup. these setups will run out of ram before they run out of processing power and will be wasting gpu potential and experiencing frame stuttering as well as texture pop when the memory buffer is maxed. the new batman is brought to mind...
 

apolloe11even

Reputable
Feb 15, 2015
30
0
4,530
Hello Fellas.
I took a look at the Posts..
so.. i just wanted to know how about SLI a 980ti XG With a 970 XG (Xtreme Gaming) ? luckily i already have a 980ti XG 6 GB on my rig and double luckily i've won the lottery of Gigabyte competition , the award was one GTX 970XG 4GB graphics card. so as i'm new with SLI totally :) i just wanted to know if i can SLI my 980 ti XG 6 GB with this new one 970XG which is scheduled to come.. any help would be great!
Thanks.
 
While the cards are based on the same gpu, The 970 has quite a bit of it disabled.

I do not think you can SLI those cards and if you could the memory performance would tank because the 980ti has much more bandwidth and would be waiting on the 970.
 

apolloe11even

Reputable
Feb 15, 2015
30
0
4,530


Thanks for respond.
I was thinking maybe i have this opportunity to use this Technology,SLI
But it seems there's no way :)
What about giving it a try? Is there any chance to check this stuff out?should be any problem with if i do?is there any thing that i must worry about if i try to SLI both i mentioned?
 
Status
Not open for further replies.