NVIDIA GF100 & Friends speculation

[Psycho]

That 3dmark06 SM2/3 scores looks more or less like a 5770 with a similar cpu... probably not very realistic...

 

[rpg.314]

GTX 480 to have a 300W TDP,GTX 470 at around 220W

300W TDP

Does no one care for PCIe stickers anymore?

If this trend continues @ Moore's law's pace, how long do you think will it be before GPU's evolve into ENIAC's shape, diet and size

 

[Mindfury]

300W TDP

Does no one care for PCIe stickers anymore?

If this trend continues @ Moore's law's pace, how long do you think will it be before GPU's evolve into ENIAC's shape, diet and size

I think NV can't make dual-fermi on 40nm,so they need to push GTX480 clock as high as possible...

 

[mczak]

Yeah, but by all accounts, a 3GPC chip will be substantially faster than Juniper. I guess, it will fit the 5770-5850 hole.

Well, a 3GPC chip with 32 rops / 256bit memory interface by all accounts could be as fast as gtx470, if they can get clocks up a bit.

My understanding is that putting 3 SMs/GPC will involve more work than is suggested by the gap between GF10x and GF100. Doing it makes the whole fermi modularity kinda redundant.

I'm not sure how this needs to be done, but certainly GF100 can already do this somehow, how else would 32 cores be disabled? In fact with GF100 it's more strange as it results in a asymmetric chip.

 

[DavidGraham]

That 3dmark06 SM2/3 scores looks more or less like a 5770 with a similar cpu... probably not very realistic...

The benchmark is done using a dual core CPU running at 2660MHz , 3D Mark06 is CPU sensitive .

 

[trinibwoy]

We are not going to have cards where the TDP is essentially equal to the max power available on the line. Come on guys....I know things are bad but a little common sense never hurt.

The benchmark is done using a dual core CPU running at 2660MHz , 3D Mark06 is CPU sensitive .

That same CPU achieves a similiar score with an overclocked GTS 250. Granted that card has more TMUs and probably a much higher clock than the 470.

 

[Jawed]

And a 5870 would be nearly 20% faster with a similar cpu.

Ignoring the nonsense overall score and looking at the individual tests, GTX280 scores higher:

http://www.technic3d.com/article-966,5-amd-radeon-hd5870-grafikkarte-generationswechsel.htm

• SM2.0 - 6964 v 5699
• HDR - 8426 v 7569

Can't say I'm willing to believe GF100 is that slow on this test.

Jawed

 

[mczak]

Another (credible) source for those gtx470 benchmarks / clocks:
http://www.heise.de/newsticker/meld...ung-der-GeForce-GTX-470-enthuellt-946411.html

1255Mhz shader clock
800Mhz gddr5 (base) clock

They say though specs are not final (specs should be final next monday).

 

[rpg.314]

I'm not sure how this needs to be done, but certainly GF100 can already do this somehow, how else would 32 cores be disabled? In fact with GF100 it's more strange as it results in a asymmetric chip.

I think disabling a few SMs will be different from designing a GPC to have less SMs. But yes, it should be possible to have 3 SMs/GPC.

 

[Dave Baumann]

We are not going to have cards where the TDP is essentially equal to the max power available on the line. Come on guys....I know things are bad but a little common sense never hurt.

Why are you so sure of this? This can be a reasn why a single rail supply is being recommended.

 

[Mindfury]

That 3Dmark06 score came from a random poster at PCI, probably fake.



heise.de (in german) has new information about the GTX470:
http://translate.google.com/transla...e-GTX-470-enthuellt-946411.html&sl=auto&tl=en

shader@1255 MHz
RAM@1600 MHz

Vantage-X 7511 (compare GTX285/HD5850/HD5870: 6002/6430/8730)
Vantage-P 17156 (compare HD5850/HD5870: 14300/17303)

Unigine benchmark
4xAA [fps]: 29/22/27 (GTX470/HD5850/HD5870)
8xAA [fps]: 20/19/23 (GTX470/HD5850/HD5870)

 

[Alexko]

We are not going to have cards where the TDP is essentially equal to the max power available on the line. Come on guys....I know things are bad but a little common sense never hurt.

Probably not, but does it really make a big difference whether it's 280, 290, 299 or 300W?

 

[trinibwoy]

Why are you so sure of this? This can be a reasn why a single rail supply is being recommended.

I'm not. Hence the common sense bit. There's simply no precedent and that scenario would have zero headroom.

edit: actually what's the TDP on the GTX 295? Is it 290+? If so, then yeah ignore me...

 

[Alexko]

I'm not. Hence the common sense bit. There's simply no precedent and that scenario would have zero headroom.

edit: actually what's the TDP on the GTX 295? Is it 290+? If so, then yeah ignore me...

http://www.nvidia.com/object/product_geforce_gtx_295_us.html
Maximum Graphics Card Power (W) 289 W

And for the HD 5970:

http://www.amd.com/us/products/desk.../Pages/ati-radeon-hd-5970-specifications.aspx
Maximum board power: 294 Watts

 

[Silus]

If the GTX 470 is really that close to the HD 5850, then it's a disaster of epic proportions. If true, it will barely be faster than the GTX 285.
I would wager that the 470 should be closer to the GTX 295 and not the GTX 285, but if not, then this is definitely not looking good.
A new architecture that brings almost no performance advantage over the previous generation, is definitely a flop.

Still, some of the numbers posted are very fishy. I would like to see someone actually "credible" with some leaked benchmarks. It's wait and see.

 

[thop]

Heise is credible, the numbers are real no doubt.

 

[mczak]

edit: actually what's the TDP on the GTX 295? Is it 290+? If so, then yeah ignore me...

Nope it's 289W Though it exceeded it in practice. Maybe it's much closer to the truth this time - kinda like what AMD did, the 4870X2 also exceeded its TDP, whereas the 5970 does not.

 

[Jawed]

I would like to see someone actually "credible" with some leaked benchmarks.

I suggest you name who you think such credible leakers are.

Jawed

 

[Psycho]

The benchmark is done using a dual core CPU running at 2660MHz , 3D Mark06 is CPU sensitive .

I know, and as I wrote, with the same cpu, a little OC'ed 5770 could have made that score - the 57x0 shows a similar difference in the scores, unlike most other cards of that range (eg. existing nvidia cards scores more in SM2 and less in SM3).
And ofcourse we should just look at SM2/3 scores and not the even more cpu dependent overall score.

 

[Silus]

Heise is credible, the numbers are real no doubt.

Wouldn't be so sure. The Unigine numbers for the 4xAA scenario are the same as the chinese site posted earlier, with the addition of the HD 5850 numbers, so this seems to be just an extension of what the chinese posted before.