32 bit vs 64 bit: How it relates to video cards

[xBruce88x]

wow i should have read this much sooner thanks for the info! so which is my real bottleneck then, 32bit XP with my newer games (apps fighting for my 2gb ram), or my Pent. D? there's an "unofficial" bios for my motherboard that lets it use 4gb instead of only 2... thinking about doing that and installing win7 64bit. of course i'd really like a core2duo/quad

 

[nt300]

you'd be capped to 2GB per app on a 32 bit OS, as for 64 bit (with 32 bit games) you're raised to a 4GB cap - including only 1GB of the video ram (second cards ram isnt used in crossfire, therefore no address space used)

Sounds to me both ATI and Nvidia need to address this issue as soon as possible so they can make sure 2nd cards ram is well in use. Sounds like the 2nd cards ram is a waist which means the dual gpu is the better way forward if you plan on crossfire or sli.

 

[Mussels]

Sounds to me both ATI and Nvidia need to address this issue as soon as possible so they can make sure 2nd cards ram is well in use. Sounds like the 2nd cards ram is a waist which means the dual gpu is the better way forward if you plan on crossfire or sli.

if it was easy to do, it would have been done long ago.

 

[MatTheCat]

second cards ram isnt used in crossfire, therefore no address space used)

So by this rationale, a 1GB 4870 + a 512MB 4870 Crossfired would give you 1GB of vRAM, just as 2x 1GB 4870's?

 

[Mussels]

So by this rationale, a 1GB 4870 + a 512MB 4870 Crossfired would give you 1GB of vRAM, just as 2x 1GB 4870's?

no. they have to match - you end up with 512MB of ram.


card 1's ram is duplicated into the second card, to make sure their images match. think RAID 1

 

[hellrazor]

So DirectX copies VRAM to normal RAM? Maybe that's why I naturally hate it so much..... OpenGL forever, baby.

 

[Mussels]

So DirectX copies VRAM to normal RAM? Maybe that's why I naturally hate it so much..... OpenGL forever, baby.

openGL likely does it too, since it uses the same hardware instructions. If the video cards could be accessed directly in ram, DirectX would have done it... and since they didnt, its logical to assume openGL cant either

 

[Mussels]

cadaveca - All you're doing is taking my thread off topic.

This will result in mass deletion if it continues.


edit: well, it was even taken to PM's telling me how wrong i am about this topic, so its cleanup time.

 

[hellrazor]

openGL likely does it too, since it uses the same hardware instructions. If the video cards could be accessed directly in ram, DirectX would have done it... and since they didnt, its logical to assume openGL cant either

OpenGL doesn't care what's in VRAM (or RAM), it merrily goes along completely oblivious to it, and I'm fine with that because that means there's more RAM for me.

 

[Mussels]

OpenGL doesn't care what's in VRAM (or RAM), it merrily goes along completely oblivious to it, and I'm fine with that because that means there's more RAM for me.

i have no proof either way with openGL. Until we can find some, i'll stay neutral. You got any info on that anywhere?

My info is from Richard Huddy, Mussels, an ATI official. I'm not wrong, if anything, he is. I could provide a link, but you are being stubborn. Ofcourse, I was going to provide sources, but you deleted the relevant info, so I'll post no source. Good day.

Your information proves absolutely nothing, since none of it has anything to do with the address space limitations of a 32 bit OS. please stop posting crap in my thread - i've already had to give you one infraction over it. You've even resorted to insulting me in PM's, and i'm serious now that you are on your last straw.

 

[hellrazor]

I find it hard to believe that OpenGL will keep a copy of all resources in RAM - especially since the Nintendo DS uses a very similar library, which would have a horrid performance penalty if it did (seeing how it has 656k of VRAM and only 4 MB of RAM - it gets more complex than that since it's controlled by two different processors, and it isn't TOTALLY opengl, but it's close enough that a huge difference like that wouldn't magically appear out of nowhere).

I admit it is iffy.

 

[Mussels]

I find it hard to believe that OpenGL will keep a copy of all resources in RAM - especially since the Nintendo DS uses a very similar library, which would have a horrid performance penalty if it did (seeing how it has 656k of VRAM and only 4 MB of RAM - it gets more complex than that since it's controlled by two different processors, and it isn't TOTALLY opengl, but it's close enough that a huge difference like that wouldn't magically appear out of nowhere).

I admit it is iffy.

without proof either way, we really cant say. I fully beleive that openGL using DX10 hardware can bypass the limit - but if we're talking DX9 hardware in a DX9 OS, it wont be able to access the hardware features to access memory directly.

 

[Mussels]

WDDM 1.1 removes the need for data duplication. So unless on Win7, the data MUST be duplicated. HD5xxx will never be able to do so though.

WDDM 1.1 uses directX 10 to remove the duplication - this only applies to AERO, the windows interface. any DX10 (or above) application (be it a game, or aero) does not require duplication.

ati 5k cards have no problems with DX10, unless you have yet another unrelated theory as to why it wont work.

 

[95Viper]

Here are a couple of good reads:
Graphics Guide for Windows 7 ages 11 through 17; the whole paper, really. Well, I enjoyed it.

Graphics Memory Reporting through WDDM

Good place for info: Device Fundamentals

Just for grins:Calculating Graphics Memory

Just info...,'cause I ain't getting in the middle of this.

 

[Mussels]

documentation is always welcome - so long as people arent making weird assumptions about what it really means.

 

[Super XP]

I can't help it, I really like this thread. Very informative

 

[kaneda]

I understand how and why the Graphics ram is duplicated in the system ram, but what im curious about is if only whats currently being used in the graphics ram is duplicated or if the OS will completely take up the ram.

say you had a gpu with 1gb of ram, would 1gb of system ram be used or would only whats being used of that 1gb of graphics ram be duplicated?

 

[cadaveca]

I understand how and why the Graphics ram is duplicated in the system ram, but what im curious about is if only whats currently being used in the graphics ram is duplicated or if the OS will completely take up the ram.

say you had a gpu with 1gb of ram, would 1gb of system ram be used or would only whats being used of that 1gb of graphics ram be duplicated?

The how and why dictates what's duplicated.

For Crossfire and SLi, because the gpus do not share framebuffers, there must be a repository of the info that they can both access...

 

[kaneda]

The how and why dictates what's duplicated.

For Crossfire and SLi, because the gpus do not share framebuffers, there must be a repository of the info that they can both access...

This was not helpful XD, sorry.

 

[cadaveca]

I know, but it depends on the situation.


The OS reserves potentially more than what is in the graphics ram, as it must cache desktop data as well, when in 3D, or alt-tabbing wouldn't be possible, for example. It's also not nessecary to duplicate anything other than what may be accessed by other processing elements, whether it's another vga, or say the cpu, with Phys-X.

So this space is typcially greater than vga ram, but it's actual size and contents will vary.

 

[kaneda]

Thanks, thats very helpful

 

[Mussels]

I've seen some games (sins of a solar empire, for one) where it crashes at 1.75GB everytime - leading us to conclude that it reserves 256MB for video ram.

So it seems that programmers can hardcode a reserved amount, or it can happen dynamically as to how much gets duplicated/cached.

 

[cadaveca]

I've seen some games (sins of a solar empire, for one) where it crashes at 1.75GB everytime - leading us to conclude that it reserves 256MB for video ram.

So it seems that programmers can hardcode a reserved amount, or it can happen dynamically as to how much gets duplicated/cached.

There's both. 256MB is a hardware requirement(which bios memory remapping steps around, and not just vgas will take up this space, every device reserves a bit), the other is driver-level/software.

The 256MB is like a buffer for communication over the pci-e bus, which is one-way communication(cpu writes to, vga reads), and the driver-level is for the OS/apps, which is shared between rendering devices.

In XP's driver model, you couldn't run concurrent 3D apps, but the shared render cache offered by WDDM 1.0 and 1.1 allows you to run as many as you wish, although, of course, performance will be hampered as both apps fight for vga ram space/communication...

WDDM 1.1 takes this further, and allows direct access to vga ram, using things like CUDA and CTM, without any data duplication... With WDDM 1.0, there is still a need for the data to be duplicated, although I don't understand why exactly.

So, when an app can use 2.0gb, and the driver itself can use just as much, ideal performance for Vista and Win7's driver models requires a 64-bit OS, as 4GB can easily be consumed just in 3D, nevermind whatever the OS and other stuff uses.

 

[Mussels]

WDDM1.1 takes it further by using DX10 instead of DX9.C - remember that that only concerns the OS and its usage of the graphics, basically for aero and media playback.

 

[Hockster]

It's 2010 and a half. Anyone still using a 32 bit OS should be flogged.

In public.