Why did video cards stop having extra ram?

Morphes

Supreme [H]ardness
Joined
Jul 16, 2001
Messages
4,336
I remember back in the day, the stock cards would have like 512mb then asus or msi would make a 1gb version. Why did companies stop doing that? Or maybe they are and I'm just oblivious. Thoughts?
 
There are 1GB, 2GB, 3GB, 4GB, 6GB, 8GB and 12GB cards.
You missed some, they are not all the same size.

The most recent to double up a stock size were the 290/290x cards that had 8GB.
Then the 3xx versions came out,
 
I added a bit :)
Its not so easy to double up a standard design due to space constraints.
And there has to be a market which is often only for SLI/XFire, standard cards arent powerful enough to be able to make use of more ram (in general).
 
AMD launched the 290 / 290X with 4GB of VRAM. Later on AMD allowed their partners to create 8GB revisions of those cards. Those would essentially become the 300 series with the 390 / 390X all sporting 8GB of VRAM.

So it does still happen. Just not a lot. There are 2GB and 4GB variants of cards like the GTX 960, but since they were available at launch, I assume that Nvidia intended it.
 
They do.

But in general, it's expensive and very rarely does it result in better frame rates.
 
For current generation GPUs, TechArp has the Nvidia GTX 950 and GTX 960, with 2GB and 4GB RAM. They have the AMD Radeon R7 370 and R9 380, with 2GB and 4GB RAM. They have the R9 390 with 4GB and 8GB RAM.
 
Last edited:
^ you are correct, ram became too fast for socketed use in video card applications.
 
Cards like the Fury and Fury X simply cannot get extra VRAM due to the limitations of the HDR they are using. Doubling the VRAM on the 980 ti is called a Titan X. 8GB is more than enough for the 390/390x. Really, most cards have the VRAM they can handle. All the different VRAM configurations we used to see were neat, but in most cases they didn't really provide any meaningful benefit.
 
They do.

But in general, it's expensive and very rarely does it result in better frame rates.

It never did. But stupid shit like FX-5200's with 256MB of RAM fucking sold because people equate RAM with speed.
 
Right but I'm saying for example a 6gb version of the 970

That's nVidia asking you to pony up for the 980Ti for the 6GB privileged. No reason they couldn't pump out a 7GB 970 (with 1GB side port crap).
 
That's nVidia asking you to pony up for the 980Ti for the 6GB privileged. No reason they couldn't pump out a 7GB 970 (with 1GB side port crap).

What's the point of a 970 with that much VRAM?
 
What's the point of a 970 with that much VRAM?

Well the guys trying to run GTX 970 SLi @4K resolutions say that their cards hitch and stutter due to the weird ram config that Nvidia did. To add more VRAM to a video card you need to double it. Would be nice if there was a way to do 4, 5, 6, 7, and 8GB configs.
 
They still do, it just isn't that big of a deal anymore. You're not seeing huge enough gains to justify spending another $100, even in the cases where it's only $50 more expensive. It's simply a bragging right for the most part since beyond 1080p is still extremely niche.
 
Well the guys trying to run GTX 970 SLi @4K resolutions say that their cards hitch and stutter due to the weird ram config that Nvidia did. To add more VRAM to a video card you need to double it. Would be nice if there was a way to do 4, 5, 6, 7, and 8GB configs.

Yet even two of them do not have the power to use up 8GB of VRAM, Two 970s are on average, what, right around a 980 ti?
 
And a 980ti has 6GB. Two 970s in SLI effectively have <2GB.

"Each" card has 3.5GB of ram at full speed, 4GB if you consider its built in cache and the overall performance drop of the last 1GB.
The 3 to 3.5GB (at full speed) doesnt reduce under SLI, where did you get less than 2GB?
 
Yet even two of them do not have the power to use up 8GB of VRAM, Two 970s are on average, what, right around a 980 ti?

Well the point I'm trying to make is that for the GTX 970 they need more than 3.5GB of VRAM to not hitch and stutter @4K resolutions in SLi as a 980ti doesn't stutter with 6GB. The problem is that the manufacturers seem to have to double the VRAM to upgrade the VRAM. Otherwise we would see a 9GB 980ti. You don't because the only other configuration possible as far as I know of is a 12GB 980ti. Which would be faster than a Titan X.

Same for AMD. I think they wanted to do a refresh of the R9 285 with a larger VRAM count, but never did due to the fact that doubling the ram would increase the manufacturing cost.
 
Adding extra RAM probably wouldn't make sense in most cases.
Take the GTX 970 that many are talking about here. Adding extra RAM would bump up the price probably making it close to the price of the GTX 980. What would be the point? People would buy the 980 if the price is the same.

Also, the GTX 970 has 3.5GB of ram that runs at full speed. The last .5GB runs slower.
Nvidia did the same type of thing with the GTX 660Ti. This isn't the first time.
The 660Ti had 1.5GB at full speed and .5GB was slower.
There were 3GB versions of the 660TI, and since it was a doubling of the fast 1.5GB, all 3GB was all full speed. Great card, but it was probably expensive. If Nvidia wanted to make a high ram version of the 970, and if they wanted it all at full speed, it would need to be a 7GB card. It would be so expensive nobody would buy it vs a 980 4GB card.

The doubling thing is still done with professional cards though. The Quadro K2200 for example is a professional version of the GTX 750TI, but it has 4GB RAM. There are other examples as well, but the pro cards need the extra RAM for compute or high res displays.
 
Well the point I'm trying to make is that for the GTX 970 they need more than 3.5GB of VRAM to not hitch and stutter @4K resolutions in SLi as a 980ti doesn't stutter with 6GB. The problem is that the manufacturers seem to have to double the VRAM to upgrade the VRAM. Otherwise we would see a 9GB 980ti. You don't because the only other configuration possible as far as I know of is a 12GB 980ti. Which would be faster than a Titan X.

Same for AMD. I think they wanted to do a refresh of the R9 285 with a larger VRAM count, but never did due to the fact that doubling the ram would increase the manufacturing cost.

I could be way off but I believe it has something to do with the bus speeds. Someone with a lot more knowledge than me will have to let me know if I'm even remotely right.
 
Personally... I think they stopped doing it on highend cards to force users to have to upgrade to the latest and greatest instead of getting a second card for SLI on the cheap.

Take this generation (and the last few for that matter) for example. You can buy a 980ti with 6gb right now and have excellent performance... after some time, games advance a bit and you start running out of GPU horsepower. Your options are to either pick up a second 980ti and run SLI or pick up the latest and greatest GPU (let's say a mythological 1080ti in this case). The problem with picking up a second 980ti is that over the amount of time it took for you to start running out of GPU power, the VRAM usage in games also increased and 6gb may not be enough.

For the best performance possible (which is what I expect when I spend $600+ on a GPU), this basically forces you to upgrade to the GTX 1080ti (now with 12gb!) instead of going SLI with your 980ti... OR shelling out an extra $350 to begin with so you can get a 12gb Titan X and SLI it later down the road.

Just a case of planned obsolescence as far as I'm concerned (shut up, I'm not paranoid!).
 
Last edited:
Go mobile GPUs, they can go batshit on the vram. GTX 980m with 8GB or a GTX 970m with 6GB?
Should be noted they have different versions with 4GB and 3GB respectively as well.


I suppose the issue is that the majority of targets for each card don't need double the vram currently available, and it's normally not a bottleneck anyway.
Hence the fraction that need more vram, will naturally buy a better GPU anyways, because it will benefit from the vram.
 
I think the VRAM is pretty well matched for the GPU power.

The other problem with throwing more VRAM at something is the bandwidth usually doesn't change. I once did rough calcs and if you actually did fill (and use all of) a 12GB Titan X your FPS would be pretty low, IIRC around 30FPS, due to bandwidth limitations. They did a really good job balancing cost, GPU power and memory config IMO.
 
Apparently there are skyrim mods that use a lot? I can't say for sure though.

Knocking on using all 6GB in skyrim myself, the 970 would die on my install of it. Even if there was a 6GB version of that card, it would be useless at that level anyway i would think, unless one didnt use an ENB.

I remember being 15 and looking at the 8600GT, and the salesman justifying to me why the 1gb model was worth the massive premium over the 512, how much faster it was supposed to be. Unfortunately for him i was on as a trainee in a PC repair store, and was taught much the wiser. I bought the 512 to his dismay, after letting him gve me his bullshit story. That was when i was playing Oblivion just after release, or there abouts. I ate those words a year later when mods became bigger, end up with a 4870 anyway.
 
Well, the ram capacity is kinda dictated by the bus width. So, each GDDR5 chip is 32-bits wide. So for a 256-bit bus you need 8 chips. The chips come in specific sizes. I think 2gbit (256MB) 4gbit (512MB) and 8gbit(1GB) are common right now, with 8gbit being very new and less common.
So, you can also run the chips in clamshell mode, where you can run two chips in each channel, (kinda like having 2 sticks of ram on one channel on a dual channel CPU) so you could use 16 chips and still have a 256 bit bus.
So that means you have have:
8x256MB
8x512MB
8x1GB
or 16x any of those a well.

so, when manufacturers make 'Double RAM" cards they are typically taking a stock design and running it in clamshell mode (unless bigger chips are available/more cost effective)
So, if you had a 8x512MB setup (like on a GTX 970/980) you could move to a 16x512MB setup for 8GB. Or if you could get 1GB chips you could go with 8x1GB.

Now, there are the unbalanced cards like the 970, the 660, and several cards in the past have done it as well. These have, for example 12 chips on a 256-bit card, where you have 8 chips running in clamshell(4x2) mode (on the first 128-bits) and 4 chips running on the second 128-bits. That means that first bank of memory is twice as big but sharing the same amount of BW as the second half. For example, if it was a 3GB card on 128-bit bus, you would have 4x512MB(2GB), thats 2x2, in clamshell mode, on the first 64-bits plus 2x512MB(1GB) chips on the second 64-bits.

Now the GTX 970 is a bit different.

It was an odd card because nVidia did something a bit different with the memory.
For accessing its memory, the GTX 970 stripes data across 7 of its 8 32-bit physical memory lanes, at 196 GB/s. The last 1/8 of its memory (0.5 GiB on a 4 GiB card) is accessed on a non-interleaved solitary 32-bit connection at 28 GB/s, one seventh the speed of the rest of the memory space. Because this smaller memory pool uses the same connection as the 7th lane to the larger main pool, it contends with accesses to the larger block reducing the effective memory bandwidth not adding to it as an independent connection could.
So basically they have 8 chips, on 7 channels, and the last channel has 2 chips on it, while the first 6 only have 1. Channel 1-6 have 1 chip, channel 7 has 2.
 
Last edited:
I remember the good 'ol days when graphics cards had sockets for extra ram to be added. The S3 Virage and the ATi Rage II's....

It's a shame their performance still sucked.
 
Well, the ram capacity is kinda dictated by the bus width. So, each GDDR5 chip is 32-bits wide. So for a 256-bit bus you need 8 chips. The chips come in specific sizes. I think 2gbit (256MB) 4gbit (512MB) and 8gbit(1GB) are common right now, with 8gbit being very new and less common.
So, you can also run the chips in clamshell mode, where you can run two chips in each channel, (kinda like having 2 sticks of ram on one channel on a dual channel CPU) so you could use 16 chips and still have a 256 bit bus.
So that means you have have:
8x256MB
8x512MB
8x1GB
or 16x any of those a well.

so, when manufacturers make 'Double RAM" cards they are typically taking a stock design and running it in clamshell mode (unless bigger chips are available/more cost effective)
So, if you had a 8x512MB setup (like on a GTX 970/980) you could move to a 16x512MB setup for 8GB. Or if you could get 1GB chips you could go with 8x1GB.
.
Assuming OP's premise is even correct, I think this is getting at the heart of it. Bigger chips could be too expensive, adding packages to run in clamshell mode might be limited by PCB space or trace routing issues. Either that, or in the case of the 970 it could be that nvidia's memory config just doesn't work right in other configurations. That said, there HAVE been plenty of "extra ram" cards in the last few years, as has been noted already in the thread
 
It never did. But stupid shit like FX-5200's with 256MB of RAM fucking sold because people equate RAM with speed.

Dat BFG Tech FX5200 OC w/ 256MB DDR and a 10% factory OC, though! :D
 
Right but I'm saying for example a 6gb version of the 970

Well in the 970's case the memory situation is so fucked up anyway that I'd imagine increasing it would be a pain in the ass for all of the AIB partners.

What's the point of a 970 with that much VRAM?

My 4K monitor would gladly demonstrate that to you.... especially now that I've got two 970's in SLI.
 
Dat BFG Tech FX5200 OC w/ 256MB DDR and a 10% factory OC, though! :D

Lol.

My girl back in the day bought that very card and was so proud.

I gave her my old 9800 pro just to show her what time it was....

Good times
 
I think higher-end RAM configs on mid-range cards would eat into and cannibalize higher-end product lines.

Imagine a 970 with 6GB VRAM. Majority would have bought that over the 980 4GB variety.
 
If they'd put 6gb on the 970 (or even full 4gb or so), it would have been doing better in vram intensive 1440p sli or similar situations.

In this day and age it can be a little beneficial. 6 months ago many were decrying the 390s for having a 'useless' 8 gb.. that 8gb turned out to be a good decision in the end.
 
What's the point of a 970 with that much VRAM?
My VRAM usage is about 3950-4060MB playing Black Ops 3. I have a 4k monitor, but have graphics scaled down to 1080p. I don't have too many games installed right now, but it was the same issue with Witcher 3... It's extremely easy to max out 4GB. I should be using a 980Ti, but it's not worth the expense right now with new cards within a year that promise way more VRAM to handle 4k properly.

Also SLI configs do not benefit from the VRAM of additional cards. If you have 970s in SLI - which can handle a lot of games just fine, even at 4k - you're going to be limited on VRAM.
 
Back
Top