I knew that the GTX590 would share the same fate as the HD5970, for GPU power bottlenecked by not having affluent VRAM. Thus I avoided quad-SLI of GTX590s, and avoided SLI of GTX580 1.5GB. Today I grabbed a GTX590 from my girlfriend and did a comparison in my machine with a pair of GTX580 3GB in SLI (downclocked to match the GTX590 clocks). That is, the same clock for core and memory, the same rig, driver, OS, game versions etc. I can't say such comparison is perfect, as there's still a difference of the PCI-E lanes used by both GF110 GPUs in each group, but this is so far the best I can do.
 
I thought it would take some time to actually meet games that can exceed 1.5GB VRAM usage at 1920x1200, but surprisingly this just happened so fast. There have been several candidates already:
 
1) Total War Shogun 2, with DX11 patch and the "Rise of the Samuri" DLC installed. The graphics of the main game menu would eat more than 1.5GB VRAM already, according to MSI Afterburner readings. To exercise how this affects the real game play, simply enter the "Advanced Land Battle" tutorial. When playing the game at max settings with a setup of only 1.5GB VRAM per GPU, there will be a log file generated in the folder %appdata%\The Creative Assembly\Shogun2\logs:
 

 
For more details read this thread. To enforce the game engine to use 8AA instead of 4AA, an override of the parameter in the script file is required: change the value of gfx_video_memory to 2147483648, and use NTFS privilege to forbid the game engine to revert the modification.
 
2) Crysis 2, with the official DX11 patch and high res texture patch installed. When the player wears the nano suit for the first time and walks outdoors, the VRAM usage can exceed 1.5GB.
 
3) RAGE, with the 9th Oct patch from Steam, and rageconfig.cfg to enable 16k texture and 16AF, the VRAM usage can exceed 1.5GB.
 
4) World of Warcraft Cataclysm, in DX11 mode, in Orgrimmar during peak time (at least 40 players in sight), start two instances concurrently (e.g. use a character to do enchanting for another character or simply mule items), the VRAM usage can exceed 1.5GB.
 
5) BF3 beta, as mentioned by many players already, the VRAM usage can exceed 1.5GB.
 
I could not find a precise way to test WOW (with repeated scenes), but I can tell that the lag spikes / stuttering that occurs especially when the player turns around quickly (rotating the camera quickly) are pretty obvious on the GTX590, but not happening at all on the GTX580 3GB SLI. I haven't downloaded BF3 beta but I've heard about many players complaining about lag spikes / stuttering already, though I don't know if such subjective claims can be trusted.
 
For the first three games I've carried out some tests myself:
 

 
This is the Frametimes measure. When the average fps and the adhoc min fps cannot tell a difference between two groups, it is necessary to look into each second carefully to identify the lag spikes / stuttering. Techreport used the same method to identify microstuttering, though microstuttering  caused by CrossfireXis not the same as the old skool stuttering caused by VRAM shortage I'm talking about here today.
 
As my observations the GTX590 are more vulnerable to lag spikes / stuttering in Crysis 2 and Shogun 2.
 
It is another story in RAGE with 16k texture and 16AF. The GTX590 simply cannot afford such settings. The game freezes at single-digit fps and fails to load all the textures:
 

 
While such config is a piece of cake for the GTX580 3GB:
 

 
Conclusions: if you trust that I haven't manipulated the data, then I have already proved the following (in)formally:
 
a) If you can see a VRAM usage reading approaching or above a certain capacity of dedicated VRAM, then the card is more vulnerable to lag spikes / stuttering;
 
b) 1.5GB VRAM cannot satisfy those who don't care about the actual gains of image quality, but only like the feeling of being able to max out every possible graphics setting in every game.
 
Don't take me wrong - I do know that there is minimal difference between 4AA and 8AA under native resolution, and almost zero difference between 8k texture and 16k texture. If you are a practical person, just turn down the AA level and 1.5GB should be fine for 1920x1200 resolution.
 
The question of "Is 1.5GB VRAM enough for 1920x1200?" is just like a question of "Is 12 Mpixels enough for full-frame digital cameras?". Only you know the right answer for yourself, because you will be the one to use it.
 

I got sli 580 HC2 1.5 GB.
 
I play all games maxed @ 1920x1200.
 
Not only do I not get those problems with the exact games you are testing, but I get PERFECTLY smooth game play. My Vram does get full 99-100% on each card but I do not get any issues with textures not loading, or lag spikes. 

@thinkfly
It amazes me how many times you feel it necessary to re-iterate the same information.

yup

^^
 
@wdflyer: you clearly did not read what I wrote carefully before you started to attack  As expected you have less than or equal to 1.5GB vram per GPU lol

thinkfly
^^

@wdflyer: you clearly did not read what I wrote carefully before you started to attack As expected you have less than or equal to 1.5GB vram per GPU lol

Oh, I read what you wrote, an unlike you I actually read the article that you used as a reference.   The simple fact of the matter is this isn't formal proof of anything but preconceived notions and twising of facts in an attempt to prove you're point. 
 
I'm not saying 1.5GB is enough or that 3GB is required, but I actually took the time to read what you wrote as well as what you used as to support your theory, er, uh "proof".

1.5 gb is good for 1080p or 1920*1200 no need more, see reviews etc..
3 gb for surround or resolution>1920*1200

VRAM is a problem right now. My next card purchases will be ones with considerably more VRAM. I'm tired of maxing out my VRAM.

dukenuke88

Sumazi

VRAM is a problem right now. My next card purchases will be ones with considerably more VRAM. I'm tired of maxing out my VRAM.

agreed...i hope the next generations of cards come with 3gb as standard on their high end cards....and 2gb as standard in their mid level cards

I am betting it will be 2gb standard and 4gb instead of three...

wdflyer

thinkfly
^^

@wdflyer: you clearly did not read what I wrote carefully before you started to attack As expected you have less than or equal to 1.5GB vram per GPU lol

Oh, I read what you wrote, an unlike you I actually read the article that you used as a reference.   The simple fact of the matter is this isn't formal proof of anything but preconceived notions and twising of facts in an attempt to prove you're point. 

I'm not saying 1.5GB is enough or that 3GB is required, but I actually took the time to read what you wrote as well as what you used as to support your theory, er, uh "proof".

First of all, go back and read carefully, I already stated that the old skool stuttering / lag spikes I'm talking about here, is different from the microstuttering caused by CrossfireX. I have read that article from techreport a long time ago and I probably know more about AFR than you think. I even had email communications with the author of the article. The author agreed with my opinion, i.e. vram shortage can increase the probability of lag spikes. He even mentioned that in his article there was a figure (560 Ti 1GB SLI fails Bulletstorm) showing the vram issue. However, in order to proof vram impacts, you'll need to compare between the same core (in my case, GF110) with different vram size.
 
I honestly have no idea why you would think I have mistaken the lag spikes/stuttering with the microstuttering. These are two different things. Microstuttering comes with a period, i.e., alternative frames does not synchronize well, making the frametimes not evenly distributed. Meanwhile, the lag spikes I'm talking about here, due to vram shortage, is a random event, which happens when the system is doing thrashing. Both cases need "Frametime Analysis" to investigate, because average fps or adhoc min fps fails to show these phenomena.
 
Is it not clear?

wdflyer
No mention of VRAM causing stutter in the entire article, but it is apparent now why the 590 has microstutter and the 580 doesn't. It is due to the AFR employed between the two GPUs of the 590, which wouldn't happen in the 580 as it is a single GPU card, again, not a VRAM issue.  

If you actually take the time to read the entire article, it's clear that the article attributes the microstutter issues to something other than VRAM, as in AFR.
 

Sorry, I just take offense when I feel someone is insulting my intelligence by not having read (or understood) their own referenced source material and then try to pawn it off as proof of something it's clearly not.

You did notice that I was comparing a GTX590 agains GTX580 SLI? The fact that you thought I was comparing a GTX590 against a single GTX580 already shows that you have no idea what is called "Frametime Analysis". Tell me, what plot of Frametimes would you expect, if I was comparing a GTX590 against a single GTX580?
 
Usually if you don't read someone's work carefully before you make comments (or even attack), you show little respect to him. I don't mind, as what I'm doing may hurt the pride of those with SLI of 1.5GB (or less) vram per GPU. Believe me, I've been through a lot more vram wars than you think (in different forums, in different languages). This is why I use GF110 1.5GB SLI vs GF110 3GB SLI to do the comparison in the same rig, to minimise the number of variables. Ask better questions / make better points if you want a good discussion here. For example, question the accuracy of my data, question why there is no error bar, question why there is no advanced statistical view. Even if you want to troll, make sure you understand the discussion.

Yeah this vram issue sucks, I will definitely be getting atleast 3GB gpu's next time around.

I agree with the next generation of cards from Nvidia, we will most likely see the upper echelon of cards with a minimum of 2GB. I just wish I was smart enough to think about this when I purchased my 3 1.5's at release;)
Thanks Thinkfly for posting your info. 
 
 
 

Reddawne

I'm running 1920 x 1080, and with Crysis 2 on max settings (with the DirectX 11, etc) my vram exceeds 1700MB. I will also always wait and purchase the next gen gpu with the highest vram.

 
Great point.  DX11 games with AA cranked up eat VRAM for breakfast, lunch and dinner.

I dont have the source but it was a legit article, not somethin off the forums.  But I read that the VRAM for the GTX680s is planned to be 2GB.  However 2GB is already becoming inadequate for powerhouse video cards and I am going to have to wait a bit longer after the GTX680 release for 4GB when 3GB would have been perfect.  But wth, always good to have too much VRAM than too little IMO.