Gold Leader you have some history in your collection of VGA cards.
I have just one old card X800 GTO still hanging around and still in use since new late 2006

kaninja
Vram limit fear is running rampant lol. So far none of the review sites testing at 4k have shown freezing, stuttering, etc. because of lack of vram on the 780ti. In fact, they all show the gpu's inability to push the resolution far before any vram limit is hit, on the Titan, 780ti, and 290x.

^^^^ Absolutely correct.....folks are forgetting something called Memory Leaks with VRAM....there is a huge difference between what a Card utilizes or uses (VRAM) vs what the game actually requires for smooth gameplay.....for example BF4 played on a single 1440p monitor using a Titan surely may show utilization of 3.5 - 4GB VRAM, BUT the same game played on the same monitor and system using a 3GB GPU may only show utilization of 2.5GB VRAM, yet no ill effects vs the 6GB Card.....this has been discussed to death in the past.
Single monitor 1440 / 1600p single 3GB VRAM GPU should be absolutely fine.....multi monitors / 4K another story.....

EVGA_JacobF
As of today it is not possible to make a 6GB 780 Ti unfortunately.
 
Whether or not it will change in the future is not known yet.

Not possible... so I guess that's something on Nvidia's end, as triss was insinuating earlier in the thread. Those of us looking to build new systems will indeed have to wait until at least Feb or March 2014 for a Maxwell release to get started, it seems = ((( Thanks for the refresh/update, Jacob.

 
Due to complaints regarding some of the posts. Trading insults, and other things of that nature will not be tolerated from this point on. Stick to the subject of this thread, and keep the posts civil. Warnings may be issued if this continues. Thank you for everyone's cooperation with this issue.
 
 
                                                               "General Warning"

Are you still saying Titan has ECC vram?

LCRava
 
Thanks Jacob. Makes me feel better on my decision of buying 4 Titans for $700 a piece. Pulled the trigger today and now will keep them until a fully enabled Maxwell comes out.

Is your wallet hiding from you yet?
Should be a badass setup, are you planning to water cool(judging by the rest of your system, YES)?

lampinridcm5

LCRava
 
Thanks Jacob. Makes me feel better on my decision of buying 4 Titans for $700 a piece. Pulled the trigger today and now will keep them until a fully enabled Maxwell comes out.

Is your wallet hiding from you yet?
Should be a badass setup, are you planning to water cool(judging by the rest of your system, YES)?

Lol. I've been working on it for 4+ months. It's not the system in my signature. I will be putting a build log up in the Mods Rigs soon.

Aggressor Prime
Guess nVidia will only be able to push 1920x1080 with games as modern as Battlefield 4 then. Granted, they can run these games at insane refresh rates with less than 4 cards, but it kind of defeats the purpose of an optimized, uber rig. Maxwell it is then!

 
I would take off the advertisement for surround or at least be more specific.
 
EVGA- "GIVE US THE GREEN LIGHT ON MORE VRAM!!!!!!!"
NVIDIA - "NEVER OR YOU WILL SUFFER!!!!!!!!!"
 
Titan wins! lol

I could be wrong but isn't the vram usage shown in precision, afterburner, etc. just the memory allocation and actual usage is way below these numbers?

MilenaM
I could be wrong but isn't the vram usage shown in precision, afterburner, etc. just the memory allocation and actual usage is way below these numbers?

Yes and no. In some games mem is allocated to almost the max you've got just to have a smooth run through the game (C2, C3). But when usage is high you're still not going to go over what you have on your card, because of the drivers, but you will feel some sort of stutter. So you never actually know what is the exact figure that the game needs as actual usage and still play fluid.
 
Example: while i had my 480 (1,5gb) SLI in BF3 on certain maps i would get vram usage shown in Precision of 1300-1400mb and bouncing between those figures. Fine and dandy by those numbers i still had some 100 left but the framerate was awful jumping from 120 to 30 and sometimes even less, frustrating to play to say the least. Now with my 680 (4gb) on those maps i get around 1900mb and smooth play. So the game actually needed more but what was shown as usage/allocation was a different story.
 
So what do people want to be right on the edge with actual usage and if case a new texture gets placed with a patch have a less than stellar gameplay or have a bit more headroom for the game to stretch it's legs (allocation)?

A 4.5GB card that uses a 384-bit memory controller would be a very odd thing indeed. 3GB, 6GB, and 12GB are the only formats that make sense. You have to realize how they come up with these amounts. A memory controller connects to 1-2 chips per 32-bit channel. There has to be equal amounts of memory in each 32-bit channel, otherwise, after using all channels which have at least 3GB of memory, the performance will degrade to half in order to use that 1.5GB in the remaining half channels. An example will be helpful.
 
Currently, Samsung produces 1Gb, 2Gb, and 4Gb GDDR5 chips:
http://www.samsung.com/global/business/semiconductor/product/graphic-dram/catalogue
 
Only the 2Gb and 4Gb chips can run at 7GHz so we will only talk about those here.
 
In the current arrangement, the 780 Ti uses 12 2Gb chips for 3GB of memory, each chip connected to a 32-bit bus.
 
For a 4.5GB arrangement, nVidia will either have to use 6 2Gb chips and 6 4Gb chips, or 18 2Gb chips with 12 of them acting in pairs, each pair on a 32-bit bus via a dual rank memory setup. Either way, the logic is going to act in the same manner (although more stress will be put on the controller if it has to use a hybrid dual rank setup; and I don't even know if mixing dual rank with single rank is possible with the GK110). Now we load the card with 3GB of data. That first 3GB will want to spread itself on as many bus connections as possible to achieve 336GB/s connection with the GPU. Luckily, we have 3GB spread out on all 12 32-bit channels. Yet once we go beyond 3GB usage, only half of those channels have more memory, so the memory bandwidth is cut in half. It can work, but it would be a very misleading product. You would have to advertise the fast, limited memory speed, and the slow, more expansive memory speed. You can produce a much simpler design with an even amount of memory per channel. It isn't like memory is expensive. What is more expensive is the GPU itself and the complexity you have to account for in designing the board. A 4.5GB card may end up being more expensive than a 6GB card because it would require a greater degree of complexity in the board to maintain stability (because you are either using a mix of single rank and dual rank memory, or you are using different memory chips).
 
And don't get me started about all these claims about 4GB or 5GB on a 384-bit memory controller. Such are just as absurd. Remember, memory, whether on the GPU, or on the CPU, works in the same manner. If a LGA2011 board has 8 memory slots, you want to of course fill at least 4 of those slots to have access to the full 256-bit memory controller. You want to use the same exact RAM as well. If you use different speeds, the whole will run on the slowest speed. Even with the same speed, latency, and amount, different brands or variants within the same brand may cause compatibility issues. Now what happens when you install 6 DIMMs of 8GB memory? Only 32GB runs at 256-bit bus speeds. The 16GB extra is only connected in a 128-bit bus fashion so that once you overload 32GB, your bus speed for the remaining 16GB gets cut in half. Now it is better than running those programs on the HDD or SSD, but it is not running at the advertised memory speed (256-bit 1866MHz, or 59.7GB/s) anymore. And if a program has an element in the fast RAM and an element in the slow RAM, both of which depend on each other for progression in the program, the program will be delayed by the slow RAM and overall run according to the speed of the slow RAM. So the effect can be worse than merely that 16GB above the 32GB. The same applies if you install 2 8GB modules and 2 4GB modules for your quad channel setup. After 8GB which is evenly shared across the 256-bit bus, speed drops to half, again, sometimes slowing even elements running in the fast 8GB domain because of dependence to elements in the slow upper 8GB domain.
 
The short of it is this:
nVidia will only release cards that have [bus width]/[32]*[1Gb]OR[2Gb]OR[4Gb]*[1]OR[2] memory. If you come up with a number that does not fit in this equation according to a GPU's memory controller's bus width, such a card is highly unlikely for the reasons given above (can be done, yes, but makes absolutely no sense). (Note, 8Gb per 1GB when doing the calculation.)

Aggressor Prime
A 4.5GB card that uses a 384-bit memory controller would be a very odd thing indeed. 3GB, 6GB, and 12GB are the only formats that make sense. You have to realize how they come up with these amounts. A memory controller connects to 1-2 chips per 32-bit channel. There has to be equal amounts of memory in each 32-bit channel, otherwise, after using all channels which have at least 3GB of memory, the performance will degrade to half in order to use that 1.5GB in the remaining half channels. An example will be helpful.
 
Currently, Samsung produces 1Gb, 2Gb, and 4Gb GDDR5 chips:
http://www.samsung.com/global/business/semiconductor/product/graphic-dram/catalogue
 
Only the 2Gb and 4Gb chips can run at 7GHz so we will only talk about those here.
 
In the current arrangement, the 780 Ti uses 12 2Gb chips for 3GB of memory, each chip connected to a 32-bit bus.
 
For a 4.5GB arrangement, nVidia will either have to use 6 2Gb chips and 6 4Gb chips, or 18 2Gb chips with 12 of them acting in pairs, each pair on a 32-bit bus via a dual rank memory setup. Either way, the logic is going to act in the same manner (although more stress will be put on the controller if it has to use a hybrid dual rank setup; and I don't even know if mixing dual rank with single rank is possible with the GK110). Now we load the card with 3GB of data. That first 3GB will want to spread itself on as many bus connections as possible to achieve 336GB/s connection with the GPU. Luckily, we have 3GB spread out on all 12 32-bit channels. Yet once we go beyond 3GB usage, only half of those channels have more memory, so the memory bandwidth is cut in half. It can work, but it would be a very misleading product. You would have to advertise the fast, limited memory speed, and the slow, more expansive memory speed. You can produce a much simpler design with an even amount of memory per channel. It isn't like memory is expensive. What is more expensive is the GPU itself and the complexity you have to account for in designing the board. A 4.5GB card may end up being more expensive than a 6GB card because it would require a greater degree of complexity in the board to maintain stability (because you are either using a mix of single rank and dual rank memory, or you are using different memory chips).
 
And don't get me started about all these claims about 4GB or 5GB on a 384-bit memory controller. Such are just as absurd. Remember, memory, whether on the GPU, or on the CPU, works in the same manner. If a LGA2011 board has 8 memory slots, you want to of course fill at least 4 of those slots to have access to the full 256-bit memory controller. You want to use the same exact RAM as well. If you use different speeds, the whole will run on the slowest speed. Even with the same speed, latency, and amount, different brands or variants within the same brand may cause compatibility issues. Now what happens when you install 6 DIMMs of 8GB memory? Only 32GB runs at 256-bit bus speeds. The 16GB extra is only connected in a 128-bit bus fashion so that once you overload 32GB, your bus speed for the remaining 16GB gets cut in half. Now it is better than running those programs on the HDD or SSD, but it is not running at the advertised memory speed (256-bit 1866MHz, or 59.7GB/s) anymore. And if a program has an element in the fast RAM and an element in the slow RAM, both of which depend on each other for progression in the program, the program will be delayed by the slow RAM and overall run according to the speed of the slow RAM. So the effect can be worse than merely that 16GB above the 32GB. The same applies if you install 2 8GB modules and 2 4GB modules for your quad channel setup. After 8GB which is evenly shared across the 256-bit bus, speed drops to half, again, sometimes slowing even elements running in the fast 8GB domain because of dependence to elements in the slow upper 8GB domain.
 
The short of it is this:
nVidia will only release cards that have [bus width]/[32]*[1Gb]OR[2Gb]OR[4Gb]*[1]OR[2] memory. If you come up with a number that does not fit in this equation according to a GPU's memory controller's bus width, such a card is highly unlikely for the reasons given above (can be done, yes, but makes absolutely no sense). (Note, 8Gb per 1GB when doing the calculation.)

Well...
 

My head almost exploded trying to read through this thread - hahaha
 
So basically no 6gb 780.
 
That definitely sucks, but if you are gaming on a higher res or multi-monitor my guess is you went with a Titan anyway.
 
The way I see it since the new Consoles are out in a couple of weeks and they basically allow 5gb-6gb of ram for developers to use, PC gamers better get used to that level or ram usage QUICK because this will be the norm with all the ports. This is even more so since both MS and Sony are using x86 platforms.  

I'm due for an upgrade, currently using 3x gtx580 3gb.. Since I play at 5760x1080, vram is a concern. I don't want to buy 2-3 780's and be able to max out vram... So, if they don't come out a card higher than 3gb, I'll end up buying titans or 290's....

gtxjackbauer,
 
Definitely agree . Its about time new consoles are hitting, but a lot of folks will get caught in the upgrade path over the next two years because it will also make a decent push in PC gaming, kinda like raising minimum wage!