Do dual-chip VGAs really have half of the advertised video memory?

[Guest]

I keep hearing that if you get, let's say 590 GTX 3GB, then you will really have 1.5GB of RAM, instead of advertised 3GB. Is it true or just a misleading information? My guess right now, is that it's BS... Why would both Nvidia and AMD advertised it that way? 6990. 590. 690 etc. must have what they got written on their boxes.

Also, there are same rumors about SLI/CFX configurations, so if you get 2x580 GTX 3GB each, then in the end you will still have 3GB, not 6GB. Really?

 

[bystander]

The technically have 3GB's of v-ram on the dual GPU cards, however, given that both GPU's on each card must have the same information loaded in vram for their own rendering, they behave like they have half the advertised amount.

 

[ilysaml]

To simplify, in SLI/CF the amount of VRAM is not doubled but it's mirrored between the GPUs like you set up a Raid 1 array.

 

[Guest]

To simplify, in SLI/CF the amount of VRAM is not doubled but it's mirrored between the GPUs like you set up a Raid 1 array.

So let's say I have a game that uses up to 3GB of VRAM, just for instance, and have 2 VGAs in SLI with 1.5GB on each GPU. Will the game bottleneck on them or not?

 

[aicom]

If there were such a game, then yes it would be severely limited. But as far as I know, no such game exists.

 

[whoareyou1982]

It will be bottlenecked. The data is mirrored between the two cards because they are each spitting out alternate frames therefore they need identical data in each set of VRAM feeding each GPU. I didn't know that either when I bought my old 9800 GX2. 500mb of RAM was a hard reality.

 

[Guest]

What is advantage of mirroring data that way in SLI/CFX?

So, really all the dual-chip VGA are half of their stated VRAM?

 

[whoareyou1982]

Not sure it's an advantage so much as a limitation.

And yes. Would be the same if you SLI'd or Crossfired two cards. You would not double your EFFECTIVE ram.

 

[aicom]

Mirroring was a conscious design decision when developing the technology. It allows the GPUs to render alternating frames, which greatly simplifies what the driver vendors need to do to support applications. If it were not done that way, the application or driver would have to dictate which particular objects were rendered on which GPU. In addition, techniques like Z-culling would be difficult if one GPU didn't know what the other was rendering, especially if they rendered overlapping objects (which nearly all objects in games are).

 

[Guest]

Is everything else like engine clock, memory clock, meory interface and memory bandwidth doubled in SLI/CFX/Dual-GPU config?

 

[whoareyou1982]

Well, not really. You have two GPU's (and sets of RAM) running at a set clock rate, so it's not really doubled. You have two GPU's running at say, 900 mhz, not one card running at 1800 mhz. Close, but not exactly the same thing in my eyes.

 

[aicom]

The GPU performance would theoretically be doubled (usually it works out slightly less than double the performance). The memory interface and bandwidth would not be doubled since each one is only connected to one GPU.

 

[bystander]

A dual GPU card is the same as two single GPU cards in almost all respects. Because it does have the same vram on both cards, they can put the total value of vram. You just have to realize it's like two cards with half the vram on each card.

 

[Guest]

Ok, so for example 690 GTX actually has 2GB of VRAM, instead of 4GB, and 256-bit memory interface, instead of 512-bit stated?

 

[whoareyou1982]

If you look at the GTX 690 info they totally call it a 512 bit memory interface. Really it's two 256 bit memory interfaces. Really misleading. I don't like misleading marketing.

I was typing that up when you posted that question and you're exactly right. 2GB of EFFECTIVE VRAM.

 

[Guest]

I see, yes, that is pretty misleading...
Manufacturers should state at least something like "690 GTX - 4GB Total Memory (2GB Effective Memory)" on their boxes...

 

[whoareyou1982]

I agree. Like I said, I bought a 9800 GX2 thinking I was getting a 1GB card. In a roundabout way your are.. And that's how they can get away with it. But oh well, I learned my lesson. At least you asked beforehand.

 

[Guest]

So, in the end, what exactly makes dual-gpus and SLI/CFX configs perform 70-100% faster than a single GPU of the same model? Engine/Core clock?

 

[whoareyou1982]

Sheer brute force. You have two GPU's firing off alternate frames, so if you're running at 60 fps say, each GPU is running at 30 fps. If you had one GPU, it would have to fire all 60 fps off by itself to give the same performance.

 

[Guest]

Ok, but then shouldn't SLI/CFX mode exactly double FPS you were getting on a single GPU? I know some SLI/CFX configs perform about 70% faster than a single setup, why it is not doubling performance 2x then?

 

[whoareyou1982]

Good question. Probably in the software implementation. Some games can, or at least come really close to, doubling frame rates.

 

[aicom]

Scaling isn't perfect. It degrades based upon the number of GPUs involved for similar reasons that a quad-core CPU doesn't perform 4 times as well as a single-core CPU.

 

[Guest]

I see.

Thanks a lot for the answers, now I have a little more idea about SLI/CFX/Dual-GPU setups

 

[Guest]

Best answer selected by Veirtimid.