would 2 8800320 meg in SLI be equal too 1 8800 640 meg card?

J

jordan12

[H]F Junkie
Joined
Dec 29, 2000
Messages
10,225
I was just curious. Please ignore any cost difference. But would they be equal in performance?
 
No. The ram doesn't add like that.
At lower res, where the memory doesn't hurt, SLI 320mb 8800s would probably be faster than a single 640mb 8800GTS. But that's not a good buy IMO. Get 1 320 or 1 640, but not SLI 320s.
 
So the newer games would not use the all of the ram on each card together?
 
The content of each card's memory is identical. Even though there is 640 MB of memory between the two cards, there is only 320 MB addressable under SLI. This is a property of SLI, and games can not change it.
 
The memory buffer remains at 320mb using them in SLi. This is because both cards have to have the exact same data in their memory. So memory bandwidth is doubled but buffer isn't.
 
It almost sounds like the cards get turned into a Raid-1 array. If Nvidia started doing physics on their boards wouldn't there be a performance increase in (2) 320's over (1) 640 in that realm though?
 
Bottom line is: as long as you don't run out of memory on the 320 sli setup it will be faster than single 640 card. But when you do run out of memory, the 640 catches up quickly.
 
It would certainly be nice if we could simply add memory modules to our cards, rather than buying a whole new card. Sigh!
 
jordan12 said:
I was just curious. Please ignore any cost difference. But would they be equal in performance?
Click to expand...

As previously said, the video RAM doesn't add like that with SLI, both cards hold the exact same information in the RAM and so it's still effectivly only 320Mb.

The cards in SLI are going to be faster in other ways, bare in mind that the speed isn't always directly propotional to the amount of video RAM, however it might have some negative side effects such as stopping you from using very high texture resolutions and such like.
 
dumb question... but why does sli help in higher resolutions so much then when talking about cranking up AA and AF? I thought the amount of AA and AF was largely dependent on the ram on the card so then sli would give no benefit at all in this
 
Jared701 said:
dumb question... but why does sli help in higher resolutions so much then when talking about cranking up AA and AF? I thought the amount of AA and AF was largely dependent on the ram on the card so then sli would give no benefit at all in this
Click to expand...

It doubles the memory bandwidth. And you need memory bandwidth for AA and AF and high resolutions.
 
if SLi doesnt use both sets of memory per card, than wouldnt it be better just for video card makers to just make a card that has no memory on it to be used exclusively for SLi only, thereby making the cost of the second card used for SLi cheaper, which could entice more people to actually use SLi since the price is lower??
 
Tetrahedron said:
if SLi doesnt use both sets of memory per card, than wouldnt it be better just for video card makers to just make a card that has no memory on it to be used exclusively for SLi only, thereby making the cost of the second card used for SLi cheaper, which could entice more people to actually use SLi since the price is lower??
Click to expand...

From what I just read (read the post above yours) that wouldn't be wise b/c it doubles the bandwidth :) But, I could see them doing that and simply giving 2x the gpu power with the same amount of memory bandwith. I think that would need a slightly altered sli setting though (could be wrong about that)
 
So then just buy the cards with dual GPU's on them in a single card and no use for SLI

having memory on both helps, period.
 
And in SLI mode, you still only use one connector on one video card and leave the others not connected to anything, right?
 
MrGuvernment said:
So then just buy the cards with dual GPU's on them in a single card and no use for SLI

having memory on both helps, period.
Click to expand...

well there isnt a dx10 card with dual GPUs out yet, so I can not do this yet..

I am sure having memory on both cards 'helps' but if it is not utilizing all of the memory then its kinda of like buying a car with two engines, where yes it makes the car faster but fails to use all of the second engine, its just wasted space to a point.
 
remove the memory you remove the double memory bandwidth which would then lower performance.
 
MrGuvernment said:
remove the memory you remove the double memory bandwidth which would then lower performance.
Click to expand...

i am not arguing that it isnt happening, i am saying it sounds like it is wasting available memory size, there is difference. another example.. lets say you have 2GB of system mem at x mem bandwidth and if you add 2GB more sys mem you will get double the mem bandwidth but still have only 2GB of available ram (i know that is NOT true, it is an example of how SLi is wasting available mem size). I plan on going SLi on my next build though, I am just trying to understand why nvidia is doing with SLi.
 
Tetrahedron said:
i am not arguing that it isnt happening, i am saying it sounds like it is wasting available memory size, there is difference. another example.. lets say you have 2GB of system mem at x mem bandwidth and if you add 2GB more sys mem you will get double the mem bandwidth but still have only 2GB of available ram (i know that is NOT true, it is an example of how SLi is wasting available mem size). I plan on going SLi on my next build though, I am just trying to understand why nvidia is doing with SLi.
Click to expand...

Well how do you suppose nvidia does it? When both video cards are used to render the same scenario, both cards have to have the same data in their memory. Isn't the point of sli to get better performance? I don't know would it even be possible for nvidia to somehow get the memory buffer doubled and memory bandwidth kept the same. Even if that were possible, I don't see any point in it. If you are running out of memory then just get a card that has enough of it.
 
You guys aren't understanding how SLI works.
The guy who said "RAID 1" Was most correct.
Having 2 cards does NOT double memory bandwidth.
Each core has to use it's own, original bandwidth, with identical data on each set of ram.
The reason AA and AF are faster on SLI is because they have half as much stuff to AA and AF per card. The bandwidth isn't doubled, just those particular memory intensive jobs are halved per card.

/edit: on another note, a nice memory SLI bridge would make sense, and make the memory amount double. there isn't much of a reason that the memory could not be SLI'd. It's done on multi processor motherboards all the time.
 
eno-on said:
You guys aren't understanding how SLI works.
The guy who said "RAID 1" Was most correct.
Having 2 cards does NOT double memory bandwidth.
Each core has to use it's own, original bandwidth, with identical data on each set of ram.
The reason AA and AF are faster on SLI is because they have half as much stuff to AA and AF per card. The bandwidth isn't doubled, just those particular memory intensive jobs are halved per card.
Click to expand...

see that makes more sense, thanks.
 
eno-on said:
You guys aren't understanding how SLI works.
The guy who said "RAID 1" Was most correct.
Having 2 cards does NOT double memory bandwidth.
Each core has to use it's own, original bandwidth, with identical data on each set of ram.
The reason AA and AF are faster on SLI is because they have half as much stuff to AA and AF per card. The bandwidth isn't doubled, just those particular memory intensive jobs are halved per card.

/edit: on another note, a nice memory SLI bridge would make sense, and make the memory amount double. there isn't much of a reason that the memory could not be SLI'd. It's done on multi processor motherboards all the time.
Click to expand...

Well this just depends on how you look at it. But you do have 2 cards. Both have, say, 64gb/s memory bandwidth. Effectively this adds up to 128gb/s.
 
SuperKeijo said:
Well this just depends on how you look at it. But you do have 2 cards. Both have, say, 64gb/s memory bandwidth. Effectively this adds up to 128gb/s.
Click to expand...
No it doesn't, since both cards have mirrored data on the RAM.
does a raid 1 array double your throughput?
short answer:
no.
 
eno-on said:
No it doesn't, since both cards have mirrored data on the RAM.
does a raid 1 array double your throughput?
short answer:
no.
Click to expand...


err yes it dose, but overhead, error correction, and platter sync timing prevents you from seeing the true dubble thoughput, (i have played with fresh raid0 builds and seen better than dubble throughtput vs a single disk, but it dissapperes as soon as you reboot the pc. but lets not get into this)

what the cards need is either a local bus link, rather than just a timeing sync, if you ask me this is where dual gpu cards have the potential to shine, the ability to put a custom memory controller on the card to make each vid core/ram set work on totally different data this should dubble throughput/framerates and all effect generation hands down. but what im haveing trouble with understanding is how it being on a single pci-e slot will affect memory push to the card itself

i know how the links (lanes) work (pci-e 16x is litterly equal to 16 pci-e 1x slots in terms of bandwidth) but how dose that compair to agp? as i understand it the orriginal agp (1x) was litterly 2x pci slots (not really there was some other times channels ect, but just for arguments sake) it was supposed to be dubble a pci slot in every way (and it pretty much was). a 4x agp slot was dubble that, and an 8x agp was dubble that again. (or 16x a pci slot...) fine and dandy now, how exactly dose pci-e stack up when compaired to pci

AGP 8x
A 32-bit channel operating at 66 MHz, strobing eight times per clock, delivering an effective 533 MHz resulting in a maximum data rate of 2133 MB/s (2 GB/s); 0.8 V signaling

First-generation PCIe is often quoted to support a data rate of 250 MB/s in each direction, per lane. This figure is a calculation from the physical signalling rate (2.5 Gbaud) divided by the encoding overhead (10bits/byte.) This means a 16 lane (x16) PCIe card would then be theoretically capable of 250 * 16 = 4 GB/s in each direction.

basicly pci-e16x is twice what a agp 8x was, only it has its own bus(but not tied into the pci bus)

if i remeber correctly, the fastest cards released for agp 8x where no where close to maxing out the bus. but this may be different today.

(ok thore so what are you geting at), well what im thinking is that the most any one pci-e slot supports is 16x@ 4gbs, if you have 2 8800gtx class chips sharing that bus, this would be equivilent to sticking an 8800gtx on an agp bus, while we know a 6800 didnt max out the agp 8x bus, i bet an 8800 dose. let alone 2 of em.

what the other guys are getting at, is buy putting a seconed gpu on a board without its own ram, you could dubble your hourse power, without sacrificing system throught put and cutting cost somewhat. what would have to happen is the 'master' chip would end up with more ram (there buy eliminateing the ram controller i mention next),

on the otherhand, if they included a localbus ram connector with a memory controller, they could eliminate this bottle neck and truly dubble all bandwidth, but the system will be extrealy complex and REALLY expensive..

---------------------------------------------------

on a side note, if a 8800gtx is capable of outputting 86.4GB in its on board ram, and the fastest system ram out there is capable of outputting 9136mb (9.1gb), that is shared acrossed the whole system, and limited to only 4gb on the pci16x bus, how is it that it is keeping the 8800's fed??? i realise that alot of this bandwidth is used up doing subsequent opperations on given textures/pixels, but 768mb is lot of ram to fill, and considering the 4gb limit, its gotta be getting tight to buss limits as it is
 
thore said:
err yes it dose, but overhead, error correction, andwidth is used up doing subsequent opperations on given ....snip.......textures/pixels, but 768mb is lot of ram to fill, and considering the 4gb limit, its gotta be getting tight to buss limits as it is
Click to expand...


No.

I did NOT say RAID 0.

I said RAID 1.

SLI is NOT like RAID 0.

It is like RAID 1.

Raid 1 does NOT double usable throughput.

Since both disks are holding the exact same data, usable throughput stays the same, it does not double (though twice the amount of data is being written in the same time, it's identical).

The RAM on both cards in SLI must hold the same data exactly. Therefore, bandwidth is the same, not doubled.

Please read before writing a book about something.
 
Tetrahedron said:
if SLi doesnt use both sets of memory per card, than wouldnt it be better just for video card makers to just make a card that has no memory on it to be used exclusively for SLi only, thereby making the cost of the second card used for SLi cheaper, which could entice more people to actually use SLi since the price is lower??
Click to expand...

It doesn't really "double" the memory bandwith, but each card GPU only have half the resolution to render which in turn use less memory bandwith.
 
Cyant said:
It doesn't really "double" the memory bandwith, but each card GPU only have half the resolution to render which in turn use less memory bandwith.
Click to expand...


It only uses less bandwidth in terms what card memory is used for outside of texture buffering. Most of it is used for just that, textures.
 
eno-on said:
No.

I did NOT say RAID 0.

I said RAID 1.

SLI is NOT like RAID 0.

It is like RAID 1.

Raid 1 does NOT double usable throughput.

Since both disks are holding the exact same data, usable throughput stays the same, it does not double (though twice the amount of data is being written in the same time, it's identical).

The RAM on both cards in SLI must hold the same data exactly. Therefore, bandwidth is the same, not doubled.
Click to expand...

DoH.. yeah.. my bad.. i misread, sorry about that.. raid 1 is redundency, not speed. i hear so many people dis raid 0 i assume.. when i know i shouldnt.. i know the bad side of raid0, and i know the consequences of it, but so many people are misinformed about what raid 0 is capable of (it makes for some damned fast swap drives(mostly why i use it)), and they dont use the correct settings.. any way the above would be correct, the way the data lies in the ram on sli'ed cards is very much the same as the way it lies on raid 1 drives
eno-on said:
Please read before writing a book about something.
Click to expand...

uhh, in case you didnt notice, i really only wrote.. oh.. one sentence about that.. the rest of my post was refering to the way the bandwidth works, and what could be done to increse performance vs what they do now, and its possible draw backs,
 
RAID 1 (not 0) can allow for increased READ performance if the controller is capable of it.

so YES it is kind of like RAID 1 with reading data faster allow for greater throughput.

So instead of having 2 cores reading from the same memory, each has it's own.
 
MrGuvernment said:
RAID 1 (not 0) can allow for increased READ performance if the controller is capable of it.

so YES it is kind of like RAID 1 with reading data faster allow for greater throughput.
Click to expand...

Raid 1 can marginally increase random read performance.
I don't see how this translates to SLI, or increased throughput.

SLi does not read data faster. Each GPU is accessing ONLY the memory physically attached to it's own card.
Therefore, no increase. Period.
 
What about SFR? I thought Split Frame Rendering meant each card rendered about 1/2 of the image. Do you use both video cards' RAM then? (in this case, 640 MB)?
 
MrToilet said:
What about SFR? I thought Split Frame Rendering meant each card rendered about 1/2 of the image. Do you use both video cards' RAM then? (in this case, 640 MB)?
Click to expand...

One card does half the frame. It uses only the memory on it's card.
The Second card does the other half of the frame. It uses only the memory on it's card.

Both cards are carrying the exact same data in the RAM. They are mirrored. It is still 320mb. It is always 320mb. The throughput number is still the same.
There will be better AA performance because each card must only AA half a screen.
 
Can I just say that SLI is a waste unless you are using 2 of the highest end cards available, and nothing less. I was on SLI'ed 7800gt's for awhile. SLI is great if your game supports it. And making custom profiles doesn't guarnatee that a game will use both cards. SO you end up wishing you had just one higher performance card that all games would benefit from. I quickly realized I would have been better off with one card having 512mB instead of 512mB split between 2 cards.
 
Ok, regardless. 1x 8800GTX 768 will be faster then the sli 320 at most res. If you are running low res. WHY bother with SLi? AND a GTX will be about the same price as two 320 GTS.

And in 4 - 6 months once they do a refresh and you can afford a second GTX. Woot FASTER. The extra memory will help with games that want it. Some of the newer games will auto downclock your setting if it detects you have low memory.
 
Sli is a waste, it does indeed help make games look pretty and still keep a decent frame rate, but as has been said a million times, the next new card is better than 2 old cards and indeed sli only uses the memory for each card, it doesnt double the memory, so if your vid card has 256mb of ram then the game you are playing will only be able to address 256mb of video memory albeit its using 256mb for each card.

Because each card is only rendering half of what you see ( when Sli works properly that is ) then their is some gains to be had from it, but just because you have 2 video cards with say 256mb of ram on each, it doesnt mean you can add those 2 amounts together and get 512mb of Vram which the game can use similar to a single video card with 512mb of ram, example I would take a 7900gtx 512 over 2 7800gtx 256mb, I would probably even take a 7800gtx 512 over 2 256mb versions of the 7800gtx, but for the fact with Sli your epenis USED to get bigger, not anymore tho as everyone knows its a waste and 6 months down the line you will be pissed off that you spent $/£1000 on 2 video cards when the next new versions come out and cost $/£400-500 and kick the shit out of the sli setup you once had.

With the advent of the 8800gtx their is NO need for sli, I sure as f00k regretted going sli as 1, its a pain in the ass to configure every fucking game to work properly with it and 2, I was sick as a dog that less than 5 months my sli setup of 78s got stomped on for half the price.
 
You must log in or register to reply here.
Back
Top