My performance simulation of GeForce 9800GX2

[The0men]

Man I like the theory behind your tests, I certainly do... But I truely beleive the Card itself will have a fair amount of variation to your findings. Its not that your doing anything wrong, but i think if nvidia wanted people to get results like this, they would have just left it to people to have 2 GTS to use.

[Devious]

Not everyone has a SLI mobo though.

[The0men]

Not everyone has a SLI mobo though.

Yer thats good point... I guess I am actually hoping that this card preforms a bit better than this to be honest.

[MuffinFlavored]

Just get good drivers, 512-bit interface (is that 256+256 for cards, no?), and 1GB of memory, underclock all you want, and leave it to the Xtreme guys to get good cooling, clock it way up.

And thats how it should be done.
Atleast, that's what she said.

[cantankerous]

ah no, actually the cards don't double their bitrate and memory amount. The card will still run at 256bit with 512mb of memory. Unfortunately with SLI you cannot double them, the specs for one card is what is used for them both when running, they are just running parallel to each other.

[Monkeywoman]

why would u unerclock? they reduced the size of the gpu to lower heat out put. my money is that the clocks will be the same if not a little higher.

[Warboy]

ah no, actually the cards don't double their bitrate and memory amount. The card will still run at 256bit with 512mb of memory. Unfortunately with SLI you cannot double them, the specs for one card is what is used for them both when running, they are just running parallel to each other.

Just get good drivers, 512-bit interface (is that 256+256 for cards, no?), and 1GB of memory, underclock all you want, and leave it to the Xtreme guys to get good cooling, clock it way up.

And thats how it should be done.
Atleast, that's what she said.

The Card is Rumored to have Full 512bit with 1GB of Memory, Its not shared from my understanding.

[JoeBar]

Why ppl can't wait...?

[HKPolice]

9800GTX will not be a 'new' chip. It will still be a G92 but this time FULLY unlocked, probably with 144 or 160 shaders and 24 ROPs (384bit ram) + even higher default clocks.

G80 has 690Million transistors (including the 9 million in the NVIO chip)
G92 has 754Million transistors but the current 8800GTS only has 16 ROPs compared to 24 on the G80.

This makes no sense, why would NV make a chip that is basically identical to the G80 design with 64 Million extra transistors but 8 less ROPs??

Sure, the G92 has double the TMUs of G80, but if you look at the 7800GTX vs 7900GTX; they have identical chip functionality but 7800GTX is 302 million transistors vs 278 million in the 7900GTX!! Nvidia was able to shave 24 million transistors off the 7800GTX while keeping all the features. I have no doubt that they were able to pull off a similar trick with the G80 vs G92 since they had about a year to work on the design.

Obviously, there are still some parts of the G92 chip that are still hard locked. This would also explain why there are such delays with supplying G92 chips for the 8800GT and 8800GTS because each chip is being binned so thoroughly. It also makes more sense this way because designing a new chip specifically for the 9800GTX line would be too costly especially since it's one of the ultra high end cards and volumes will be very low.

Nvidia is very smart taking this approach because it gives the highest yields possible. With the RV670 chip, if only 319 of the 320 shaders are working, then the whole chip is useless and scrapped. With the G92 chip, if all the shaders work, they bin it as a 9800GTX, if only 129 of them work: 8800GTS, if 113 of them work: 8800GT and if only 97 of them work: 8800GS! Very few wasted chips.

[Cooper]

The Card is Rumored to have Full 512bit with 1GB of Memory, Its not shared from my understanding.

You are correct - that's rumors. technologicaly this is not possible with current implementation. it still remains 2x256bit and 2xwhatever-memory-amount-per-card.

[Shocker003]

Remember, this was just a simulation, not real results.

You are damn right . We shouldn´t throw punches over a simulation, wait until these 9800 cards to hit the retail shops.

[zerazax]

9800GTX will not be a 'new' chip. It will still be a G92 but this time FULLY unlocked, probably with 144 or 160 shaders and 24 ROPs (384bit ram) + even higher default clocks.

G80 has 690Million transistors (including the 9 million in the NVIO chip)
G92 has 754Million transistors but the current 8800GTS only has 16 ROPs compared to 24 on the G80.

This makes no sense, why would NV make a chip that is basically identical to the G80 design with 64 Million extra transistors but 8 less ROPs??

Sure, the G92 has double the TMUs of G80, but if you look at the 7800GTX vs 7900GTX; they have identical chip functionality but 7800GTX is 302 million transistors vs 278 million in the 7900GTX!! Nvidia was able to shave 24 million transistors off the 7800GTX while keeping all the features. I have no doubt that they were able to pull off a similar trick with the G80 vs G92 since they had about a year to work on the design.

Obviously, there are still some parts of the G92 chip that are still hard locked. This would also explain why there are such delays with supplying G92 chips for the 8800GT and 8800GTS because each chip is being binned so thoroughly. It also makes more sense this way because designing a new chip specifically for the 9800GTX line would be too costly especially since it's one of the ultra high end cards and volumes will be very low.

Nvidia is very smart taking this approach because it gives the highest yields possible. With the RV670 chip, if only 319 of the 320 shaders are working, then the whole chip is useless and scrapped. With the G92 chip, if all the shaders work, they bin it as a 9800GTX, if only 129 of them work: 8800GTS, if 113 of them work: 8800GT and if only 97 of them work: 8800GS! Very few wasted chips.

I hope that's true, and it would make sense, but I thought the G92 was 128SP's max unless they derived another chip from it

[saaya]

According to several reviews, the 8x pcie 1.0 versions have plenty of bandwidth except for o/c'd Ultras and SLI configs. 16x and 16x X2 are PLENTY of bandwidth for every card under every condition possible. The best part about pcie 2.0 right now is the increase in power to the slot. Until cards start maxing out 16x bandwidth, the extra it gives is useless.

Kind of like ATI's decision to bring dx10.1 and sm4.1 in their cards now when its only a minor update......and an update that wont even be out until SP1 is officially released....and devs wont even be using these features until at least the end of 08 into 09.

Can anyone tell I HATED that ATI tried to market those features?!?!?

yeah, the difference is small, but people here go nuts if they find a way to gain 0.1% extra performance... we can only guess what the difference from 2x16 2.0 and 1x 16 2.0 is for 2 cards in sli, but id say its 1-5% depending on the app. and if you use professional apps like 3dmax and catia then you DO see big differences between pciE 16 and 8 and pciE 1.0 and 2.0 afaik.

about dx10.1... afaik it only makes some features of dx10.0 mandatory... its just atis way of showing off that they support 1 or 2 features nvidia doesnt or doesnt guarantee

With the RV670 chip, if only 319 of the 320 shaders are working, then the whole chip is useless and scrapped.

that doesnt make much sense, why would ati throw away gpus?
im sure they are recycled and used as something else.
and what makes you think rv670 only has 320 shader processors after all?
i wouldnt be surprised if it has 360-400 shader processors overall

about g92, i think it has more than 128 stream processors...
and H and other sources already confirmed this:

256 Stream Processors Total

[zerazax]

about g92, i think it has more than 128 stream processors...
and H and other sources already confirmed this:

I thought they meant the 9800GX2 would have 256 SP's total, which is correct if it is 2 x 8800GTS-512 (full G92) cores put together, as that's 2 x 128...

It's the same as saying there's 1GB of RAM when it's really just 2 x 512MB's, 512 for each core

[HKPolice]

I hope that's true, and it would make sense, but I thought the G92 was 128SP's max unless they derived another chip from it

Ya well everyone thought that G92 only had 112SPs when 8800GT first came out

[HKPolice]

yeah, the difference is small, but people here go nuts if they find a way to gain 0.1% extra performance... we can only guess what the difference from 2x16 2.0 and 1x 16 2.0 is for 2 cards in sli, but id say its 1-5% depending on the app. and if you use professional apps like 3dmax and catia then you DO see big differences between pciE 16 and 8 and pciE 1.0 and 2.0 afaik.

about dx10.1... afaik it only makes some features of dx10.0 mandatory... its just atis way of showing off that they support 1 or 2 features nvidia doesnt or doesnt guarantee


that doesnt make much sense, why would ati throw away gpus?
im sure they are recycled and used as something else.
and what makes you think rv670 only has 320 shader processors after all?
i wouldnt be surprised if it has 360-400 shader processors overall

about g92, i think it has more than 128 stream processors...
and H and other sources already confirmed this:

Uhhh do you even understand how chips are made? Not every single chip on a wafer of silicon is perfect. Many of them have defects so parts of the chip have to be disabled or the whole chip gets thrown away if there are too many defects.

There are no other ATi cards based on the RV670 chip that has fewer shaders than 320. Maybe they're saving up all the defective chips with less than 320 shaders and ATi might release a lower end card using these chips later, but so far there has been nothing.

H only confirmed that the 9800GX2 has 256 shaders, NOT the 9800GTX. You are confused.

[zerazax]

Ya well everyone thought that G92 only had 112SPs when 8800GT first came out

Uh who thought it was just 112SP? I sure as hell didn't, and most people knew that they wouldn't cut the real SP down from 128, which was what the 8800GTX had. After all, if G92 is a derivative of G80, why would they cut SP's?

And most reviewers and people who went to the unveiling knew that the architecture was mostly the same, and knew it was still 128 but had parts disabled.

[zerazax]

The Card is Rumored to have Full 512bit with 1GB of Memory, Its not shared from my understanding.

If it's anything like the 7950GX2, it's not shared.

Current GPU architecture and design hasn't gotten around that yet. We'll see what R700 does though, if it lives up to all its hype

[Warboy]

If it's anything like the 7950GX2, it's not shared.

Current GPU architecture and design hasn't gotten around that yet. We'll see what R700 does though, if it lives up to all its hype

Yes I know, But If it really does have a 512bit bus, and 1GB memory. Then It will be a 8800GTX/Ultra killer.

[bakalu]

I think, 9800GTX will still be a G92. It has 128 SP, 64 TMUs, 768MB Ram 384 bits, 24 ROPs.

[Leeghoofd]

8800GTS 512 SLI slower than 8800GT SLI?

Didn't see that ? one of us needs glasses or a bigger screen

Secondly I hope the driver support for the 8800X2 is better than the 7950x2 I had, as it really was slower in many games than my 7900GTX....

Do you like the Asus OBR ? might sell my Maximus and go for the P5N-T Dlx

[STaRGaZeR]

Didn't see that ? one of us needs glasses or a bigger screen

I was talking about the simulation, 8800GTS 512 SLI with lowered clocks to match future 9800GX2 specifications... One of us seems to need interpretation lessons

First, in the simulation is slower than 8800GT SLI and 8800GTS 512 SLI. Second, NVIDIA will suck in the driver development, so maybe we will see even more perfomance hit. All of this 9800GX2 crap looks like exists only to face the 3870X2, at marketing level.

[saaya]

I think, 9800GTX will still be a G92. It has 128 SP, 64 TMUs, 768MB Ram 384 bits, 24 ROPs.

no, 512mbit ram per gpu only afaik... makes sense, to get more mem you basically have to pay 2x as much as with a normal videocard cause the mem isnt shared. so when going from 512mb to 1024mb for a single card means 50$ extra, but for a dual card it would mean 100$ extra...

but its dissapointing... i thought the gx2 is supposed to be a high end card, the fastest possible card... well for that you would expect at least 768mb per gpu to have reasonable fps at high res with high aa. a 256 vs 512 vs 1024 compare has already shown that at high res plus high aa 1024mb per g92 gpu means fps bosts of 50%+ ...

[gdogg]

Thanks for the benches, any idea where crossfire 2900xt 512MB's will fit in there? Got them 16x16x, I can do my own benching, but config is different.

[Leeghoofd]

IOne of us seems to need interpretation lessons

Really not much to interpret from that phrase "Originally Posted by STaRGaZeR : 8800GTS 512 SLI slower than 8800GT SLI ? " : I'm just a belgian , you know from that country where we had elections and 140 days later we still don't have a real government lol...
Just a joke mate I knew what ya ment.

On topic now : another thread came up, at Tomshardware ( reliable or not) this card is posted to hit the shelves at a retail price of 450 dollars ???? lol my 8800GTS 512 is outdated again and it didn't even reach my house looooooooooooooool

Link to that news (no review though nor performance) here :

http://www.xtremesystems.org/forums/...d.php?t=172440

Could that price be real ? This card could do some heavy pounding in SLI config on the 3dmarks... if they get their drivers sorted and hopefully most games will support it... ( heavy doubts there though)