R700 to Come in 2048MB Flavour?

[Rash-Un-Al]

Marchitecture

2GB is likely to be a simple and pure function of marketing.

The 48xx series cards have proven to not be memory buffer- or bandwidth-limited. This is especially so – at nearly any conceivable resolution, with or without filtration – with the 4870, with it's monstrous GDDR5 bandwidth capabilities. Add to this the fact – with 512 MB or 1024 MB of memory – the interface will remain 256-bit. Thus, whether it has 512 MB or 4096 MB it can access, write to, and read from the memory only to the extent to which the (256-bit) interface allows. This is why 1GB variants of cards of the same model and technology do not perform any faster. On the contrary, in some instances, they perform even more slowly because 3rd-party, larger-memory renditions of a card often bring about cheaper and less capable memory, which run at similar frequencies to the original memory, but that also introduce higher latencies to achieve said frequency.

For your (GDDR5) reading enjoyment:
http://forums.techpowerup.com/showthread.php?p=866224#post866224

 

[Bluesman]

Not Viable for a 32-bit OS?

IMHO a 2 gig video card is only viable with a 64-bit operating system. Why? Because you will eat up 2 gig of onboard RAM before you account for other hardware memory reservations in the upper 32-bit registers.

In a 64-bit OS you don't have to deal with the bs 4 gig limit we see in a M$ 32-bit OS. Here is a better explanation of my point = EXPLANATION

 

[Siman0]

i can see like DAMMIT has a monolithic GPU trump card that takes the 55nm but the same die size as the 280 fill it with transistors and increase the bit rate on the gddr5 could you even comprehend the power it would radiate from your computer o ok wait lets be real could you even comprehend the heat it would radiate from your computer <- is it jest me or does that smiley look ok nvm

 

[largon]

2GB is likely to be a simple and pure function of marketing.

There are tons of games that easily exceed 512MB in pure texture data. Without any filterings.

The 48xx series cards have proven to not be memory buffer- or bandwidth-limited. This is especially so – at nearly any conceivable resolution, with or without filtration – with the 4870, with it's monstrous GDDR5 bandwidth capabilities. Add to this the fact – with 512 MB or 1024 MB of memory – the interface will remain 256-bit. Thus, whether it has 512 MB or 4096 MB it can access, write to, and read from the memory only to the extent to which the (256-bit) interface allows.

Are you saying the additional memory "wont be used" because bus width is "only" 256bits? Cause that is not the case.

IMHO a 2 gig video card is only viable with a 64-bit operating system. Why? Because you will eat up 2 gig of onboard RAM before you account for other hardware memory reservations in the upper 32-bit registers.

In a 64-bit OS you don't have to deal with the bs 4 gig limit we see in a M$ 32-bit OS. Here is a better explanation of my point = EXPLANATION

Except the fact that video card memory has absolutely nothing to do with the limitations of 32bit OS'. Video RAM is only local to the GPU. OS/CPU doesn't address it.

 

[flashstar]

Largon, you are correct but you can only have 4gb of memory in your computer TOTAL without PAE on a 32 bit system. The OS does count the onboard video memory in this total. So if you have a 2gb graphics card, you will only be able to use a little less than 2gb of system memory since some of that will be reserved for devices.

 

[Rash-Un-Al]

Are you saying the additional memory "wont be used" because bus width is "only" 256bits? Cause that is not the case.

Thanks for the opportunity to clarify. No, I was not stating that it "won't be used" and I do understand that the additional memory could be utilized, if necessary.

There are tons of games that easily exceed 512MB in pure texture data. Without any filterings.

I am not in disagreement. However, those instances rarely represent a window of opportunity that would allow for merely doubling of the memory quantity to contribute to significant increases in performance, on the same architecture (with the same bandwidth limitations). How effective is double the memory, if the path through which the memory's data traverses serves as the restriction?

Examples concerning both camps:
http://techgage.com/article/palit_geforce_8800gt_super1gb/3
http://techreport.com/articles.x/14654

It is no mystery why, when 1GB 8800GT cards bump into a memory (bandwidth) limitation, cards like the 8800 GTX and Ultra continue to shine in the same conditions – with less memory. In these situations, the G80s are not only able to take advantage of a larger buffer, but can also write to and read from all of 768MB of memory at the same rate its G92 counterparts can with 512MB, thanks to a memory interface which was widened proportionate to the memory buffer (thereby, increasing memory bandwidth, and not solely the buffer size).

One can counter this argument with how significant differences in performance are observed when comparing 256MB versions of the 8800GT and HD 3850 with their 512MB counterparts. However, one may want to also consider the fact both the aforementioned architectures were developed and optimized for 512MB through a 256-bit interface. Therefore an increase from 256MB to 512MB would not represent a situation in which an increase in buffer size was limited by the memory interface.

 

[Bluesman]

Except the fact that video card memory has absolutely nothing to do with the limitations of 32bit OS'. Video RAM is only local to the GPU. OS/CPU doesn't address it.

Windows reserves memory for hardware devices like video cards. The higher hex addresses, such as A0000 to BFFFF, are there for that purpose. If you install a 512 meg video card, the system will reserve 512 Meg whether you need it or not. Same with SLI - 2 cards at 512 meg each and the system will reserve 1 gig. You can confirm this for your own video card by opening Device Manager, using the View -> Resources By Connection option.

Couple the above with other devices, like sound cards, and you will substantially limit availabe memory for programs. See HERE

 

[Mussels]

Windows reserves memory for hardware devices like video cards. The higher hex addresses, such as A0000 to BFFFF, are there for that purpose. If you install a 512 meg video card, the system will reserve 512 Meg whether you need it or not. Same with SLI - 2 cards at 512 meg each and the system will reserve 1 gig. You can confirm this for your own video card by opening Device Manager, using the View -> Resources By Connection option.

Couple the above with other devices, like sound cards, and you will substantially limit availabe memory for programs. See HERE

bluesman is correct. video ram is included in the 4GB limitation of 32 bit OS's.

 

[eidairaman1]

You can run 4 GB total, just thing is its with Inobvious OS tweaks to mem management, i would say modifying 64bit mem management would do the trick.

 

[magibeg]

Well i guess another take on things would be that ati/amd wouldn't put 2gb on that card unless they felt it was really needed. Its also really hard to say at this point if the 4870 would benefit from having all the extra memory or not but a lot of games now do use up more than 512mb or video memory. Now when people talk about it being held back by its slow 256bit bus they also have to take into account they're comparing it to the alternative which is having the extra textures just stored in ram, which would undoubtedly be much slower.

 

[Rash-Un-Al]

Well i guess another take on things would be that ati/amd wouldn't put 2gb on that card unless they felt it was really needed.

Agreed. The reports are a bit conflicting, as to whether the reference design will be 512MB x2 or 1024MB x2. My original understanding was that the absolute reference design would involve 512MB x2 and that variants (as a result of partner request, demand, or custom order) would sport 1024MB x2. I imagine these details remain to be confirmed by AMD/ATI.

As you may have already seen, "leaked" images suggest 512MB x2:
http://www.engadget.com/2008/07/03/amd-radeon-hd-4870-x2-images-leaked-rumored-for-august-release/

 

[Mussels]

Well i guess another take on things would be that ati/amd wouldn't put 2gb on that card unless they felt it was really needed

Please, please keep in mind that its only 1GB real mem. dual GPU cards dont share memory, they have 1GB 'each' and its not shared in crossfire.

when comparing it to single cards, they're 1GB models... and people are seeing good boosts from 1GB 8800GT's, so a monster like this would definately need the 1GB ram.

 

[btarunr]

^Yes, he makes a valid point. It's 1GB of mirrored memory.

Say you're playing a game, the textures are loaded into the memory arrays of both GPUs, each GPU renders in either AFR or tiles. For that, each GPU should have the same texture and mesh information as the other one does.

 

[AsRock]

Kinda cool, i just hope it actually needs it this time. Should help ATI to get the prices closer to the 280 price tag.

 

[Rash-Un-Al]

HD 4850 with 2GB of Memory

There are tons of games that easily exceed 512MB in pure texture data. Without any filterings...

Here is an interesting review of an HD 4850 with not 1GB, but 2GB of memory…

… for those of you who hoped more than 512MB would increase the 48xx series' performance.

http://www.guru3d.com/article/radeon-hd-4850-2-gb-gddr3-review/1

No Surprise.

 

[eidairaman1]

looks like Crysis likes the ram, but not by much, id be surprised if the drivers are only detecting 512 of the ram, because usually cards that have more ram also get a slight clock speed increase.

 

[mab1376]

you would never utilize it unless you run 1600p (2560x1600) and even that might be a stretch.