Nvidia Releases World's first Quad GPU - 9800GTX (QG98) [Tri SLI Already Obsolete]

[Yojo]

Nvidia Releases World's first Quad GPU - 9800GTX (QG98) [Tri SLI Already Obsolete]

Anandtech recently did a review of Nvidia's 780i chipset and the chipset's feature TRI-SLI.

One remark which I though was particularly interesting was this:

Then there's the bigger issue of SLI and CrossFire technologies in general, scaling is a little too dependent on software. You're increasing the execution resources of a standard 2-card SLI setup by 50%, but the performance impact is no where near that. Whereas if you added 50% more SPs to those two 8800 Ultras you'd see a much more tangible outcome. It's an extreme version of the way Intel makes quad-core CPUs, but instead of sticking two die on a single package, you have two die spread over two cards - that's hardly efficient. GPU architectures have changed dramatically over the past few years, yet we're still left with the same old multi-GPU technology. It's time for a change.

Instead of making us, the consumer, play 3xfold for f***ing graphics cards, why not make a multi-core GPU! This way the communications between the different GPUs would take place on the same *die* and not through a card, a PCIe lane, and a chipset. Remember earlier when AMD made motherboards that would let you plug in two AMD duo cores to realize a terrible 'quad core' implementation? And then they followed intel and did 4 cores on the same die? Why are we going this stupid route of "tri-SLI" when instead a more efficient implementation would be multiple GPU's on the same die on the same graphics card?

You could have say an 8800GTX (single GPU), 8800GTX (dual GPU), 8800GTX (quad GPU), etc.

Please, please, please tell me I'm not the only one to think of this.
I pray they have hired scientists working at Nvidia that are miles ahead of me on this one...
</end rant>

 

[Sederien]

The title on this thread is grossly misleading.

 

[ChinStrap]

The title on this thread is grossly misleading.

this

x900

 

[reclaimer122]

...Oh...

Lol, I can agree with you, but wouldn't nVidia and ATI make a lot more money selling you three cards?

 

[Theocnoob]

Im pretty sure there's an issue of throughput there which necessitates SLI in the firstplace...

The whole point of having a multi core cpu at this point is that we've reached a sufficient system bus width to permit that without excessive bottlenecking but we're talking about communication via a PCIx16 bus- once you hit the bandwidth limitation of that bus you can have 89 GPUs on your card and it won't matter. You have to GET to the data to process it.

They're working on a new PCI specification because of this.

 

[Yojo]

The title on this thread is grossly misleading.

Sorry, I think I was a little too dramatic in soliciting for comments about my rant. I will be a little more careful and not 'cry wolf' (misleading thread titles) in the future.

You have to GET to the data to process it.
They're working on a new PCI specification because of this.

Agreed and then maybe, hopefully we'll see multi-core GPU's in the future after this issue is addressed?

I'm all for Nvidia making money, but they would make *MORE* money if we got more bang for our buck. How many people will serious consider TRI-SLI, this seems very exclusive to your extreme computer enthusiast.

 

[Yojo]

this
x900

Yeah, but were you not excited?! Come on, was there not a part of you hoping that this was true? That you wouldn't have to spend 3x$700 (exclusively on video hardware) for a top of the line system and instead could spend less for a more efficient multi-core GPU??

 

[ALiEN2953]

notice how large a single G80 is and how much heat it puts out...do this times four...a 1920mm Chip that puts out >500w yeah that would work...

/rant
dont be an idiot in the future
/end rant

 

[ALiEN2953]

read up on R700

http://gprime.net/flash.php/postingandyou

i think there is something about you topic being supposed to relate to your post

 

[White_Pawn]

I was actually taking back at first that perhaps the 9800GTX was a new architecture, instead of just a G92 refresh of the 8800GTX. Guess i was wrong. LOL

I agree with the pci-e bus frequency. At only 16x, it will be saturated very shortly even with a single gpu. That's why there is the x32 lanes from the pci-e 2. It'd be like a 8800ULTRA w/ a 128bit-mem bus. Overkill.

 

[knightwolf6543]

G80 was a new arc.
G90 was the die shrink
D9 should be a new arc.

tick tock remeber. (dammit intel....)

 

[ziggo0]

I got my hopes up for a second there....

 

[Elluzion]

Tri SLI is obsolete? lmfao.

 

[4od]

They already tried this with the 7950 GX2 and it wasn't as good as SLI because of the PCIe bottleneck. I can't see that problem being alleviated by using 4 8800 Ultras instead of a pair of 7950s

 

[Caust1c]

ya i think thats in the roadmap for nVidia. I believe its next in line to be the new GTXs in the 9000 line. (assuming its called that) and the core is supposedly called D9 or D9P

 

[ratbuddy]

They already tried this with the 7950 GX2 and it wasn't as good as SLI because of the PCIe bottleneck. I can't see that problem being alleviated by using 4 8800 Ultras instead of a pair of 7950s

Uhh, PCIe 2.0.

 

[deathman20]

Uhh, PCIe 2.0.

SLI currently doesn't saturate the PCI-E bus as long as its x8 or x16 PCI-E speeds 2 cards runs at its maximum potential in a PCI-E 1.0a slot

 

[jason4207]

I thought PCIe 2.0 only gave you x32 lanes, but on 2 separate slots. So it's basically x16 + x16, the same thing 680i has had for a while. Am I wrong? Does PCIe 2.0 allow x32 lanes in 1 slot?

 

[nd4spdbh2]

Im pretty sure there's an issue of throughput there which necessitates SLI in the firstplace...

The whole point of having a multi core cpu at this point is that we've reached a sufficient system bus width to permit that without excessive bottlenecking but we're talking about communication via a PCIx16 bus- once you hit the bandwidth limitation of that bus you can have 89 GPUs on your card and it won't matter. You have to GET to the data to process it.

They're working on a new PCI specification because of this.

TBH we arnt even close to bottlenecking a single pcie 16x 1.1 slot... let alone twice the bandwidth pcie 16x 2.0 slot... i mean a 8800gts 640 is only 10% bottlenecked by a 4x slot over a 16x (1.1 spec) that said you could have 4 8800gtx's running over the same pcie 16x 2.0 slot without bandwidth problems...

 

[Gid]

The title on this thread is grossly misleading.

QFT