How does the Nvidia GTX 690 work?

JPNpower

Honorable
Jun 7, 2013
1,072
0
11,360
So the GTX 690 has 2 gpus in internal SLI. what the heck is that? Does that mean it has the exact same performance as 2 GTX 680 gpus in "external" SLI?

Are 4 cards in SLI in effect octo-SLI or something? What is the advantage of having 1 card over 2? I'd assume that power efficiency would be similar with 2 GPUs having to be fed on one card over only a single GPU on all the others.
 
Solution


The evidence is in how SLi works, both GPU's have to have access to the same data. If you are still confused then research how SLi works for yourself instead of asking others to spoonfeed you.

RaisingTheBarHD

Honorable
May 8, 2013
822
0
11,160


the 690 is basically two 680 in one. the 690 is considered an sli card and if you wanted a 4way sli it could only be two 690 or you could do four 680s, basically pairing up two cards will give you better performance in fps and graphics. You will only need a 780 max for a single monitor usually sli is meant for triple monitors
 

casparolesen

Honorable
Dec 14, 2012
92
0
10,640


The 690 is, as you mentioned, just two underclocked 680s on a single PCB. A 690 will be a little bit slower than two 680s, because the ones on the 690 are underclocked. 4 cards in SLI is called quad-SLI, so two 690s in SLI are quad-SLI. The advantage in having 1 card over 2 is, as you mentioned, power efficiency and heat output. Both will be lower on the 690.
 

Two GPU in one PCB.



No.



No. You would only have one SLI, which is technically a Quad-SLI since the GTX 690 by itself is a SLI.



Less space; less heat; no need for a motherboard that supports the required bandwidth for two. So like a Micro ATX or Mini ITX can work with it.



No.
 

farrengottu

Distinguished
Aug 28, 2011
912
0
19,160
its not the exact same performance but its with in 5%.... maybe more like 3% but its close. internal sli is basically the same as the external but instead of running a cable its done in the circuit board. making for shorter wires and microseconds less delay when sending signals. but really just the same thing. whats bad about a card like a 690 is you have twice as much that can go wrong. and its a pain when one gpu dies after your warranty is up. cuz your out both.

4 690s isnt supported. i wish it was. but ive looked and i cant fine one person who has made it work and the drivers only support up to 4 gpus. though ive seen a guy get up to 7 680s working. he couldnt get the 8th to work. but he didnt state if it was just working like for expanded displays for physx or if it was part of the gaming graphics.


 

JPNpower

Honorable
Jun 7, 2013
1,072
0
11,360
So now I have more questions.

Is the 690 sort of split into two parts for the two GPUs?
Does one GPU have access to only half of the vRAM?
Since the two GPUs are sharing a PCIe x16 bus, wouldn't bandwidth be the same as 2 x8 as normal SLI is?
How is bandwidth, in general, handled? Is it dynamic? Or does one GPU have access to only half at maximum?
 

I don't understand that question.



No; it has access to all its VRAM.



Correct.



I don't understand this question either. Define "handled" and what do you mean by "only half at maximum". Maximum what?
 

JPNpower

Honorable
Jun 7, 2013
1,072
0
11,360
I mean, if one GPU were to tacke a task while to other sat idle, would that active GPU have 100% access to the entire x16 bus, and all of the vRAM? Or is there some SLI limitation in which the GPU can only access half of the PCIe bus.
 


IIRC the VRAM is split between the two GPU's so with a 4GB card one GPU would only ever have access to 2GB of VRAM.
 

JPNpower

Honorable
Jun 7, 2013
1,072
0
11,360


Now you're completely contradicting the other guy. Do you have evidence? Now I'm seriously confused...
 


The evidence is in how SLi works, both GPU's have to have access to the same data. If you are still confused then research how SLi works for yourself instead of asking others to spoonfeed you.
 
Solution

Actually, there is no contradiction here. I agree with Mousemonkey that you should do your own research on the subject. Each GPU has access to all of its VRAM. However, since each GPU retains a copy and SLI / CrossFireX does not increase VRAM, it makes more sense to say that only half of it is useable. Otherwise, if the GPU itself does not retain a copy of the buffer, it would be incredibly slow and even worse to manage software-wise.