The Fermi Thread - Part 3

[kaktus1907]

If this is true, we have a deference from 470 to 480 32sp witch seems to be a +20% advantage for GTX480, if fermi was with its whole sp512 then maybe we have seen xtra +20% advantage from 5870.
Hope GTX480 ultra or something hits the marker.

not only 32 shaders, plus 8 ROPs + 15% clock speed that means some difference than 470.. rops are more important imo..

[Musho]

To be honest, I'm not impressed at all. 5870 seems to be far better value for money. You have to add in both the fact that they are using 10.2 instead of 10.3 and the fact that those benches are from nVidia. I'm sure they picked the best settings/games/ingame locations possible to make Fermi look good. I'm quite certain the reviews to come will show a smaller advantage for Fermi.

[weston]

Exactly. ATI has had what, 7 months or so to refine their drivers? Nvidia should easily add about 15 to 20% (possibly more) increase in performance from driver optimizations alone over the next few months..

I'm betting the 200.xx drivers will be the first to have real G100 optimizations..

lol I have heard these exact arguments about the R600 when it first came out. Easily add 20%? It might happen, but not easily. You act like drivers are somehow as magical as steve jobs' iPad

It's crazy that the GTX 295 may still be the fastest card offered by nVidia for a while longer.

[trinibwoy]

My problem is the meagre bandwidth increase over the 285 - 12%. And in terms of SLI well when we're talking about < 20 fps does it matter?

[Sn0wm@n]

Yeah, I'm on a 30 incher myself and those SLI numbers if true look really good..

And I'm sure they'll get better as well in the future.

But it's hard to justify spending 1,000 bucks on graphics alone, especially when only a handful of games require that sort of juice to play at 2560x1600 maxed settings..

then get another 30inch monitor and enjoy nvidia's eyefinity

[japamd]

I've made the calculations:

GTX 480 beats the HD5870 %24 average on 1920 + 4xAA, and 23% average on 2560 + 4xAA.

GTX 470 performs nearly completely equal to HD5870 on both settings.

if those benches are correct.

Let's make easier to compare

Scores normalized to HD 5850 = 100%

Only 1920x1200 AA 4x e AA 8x

[GoldenTiger]

Let's make easier to compare

Scores normalized to HD 5850 = 100%

Only 1920x1200 AA 4x e AA 8x

[*IMG]http://i44.tinypic.com/14b47ec.png[/IMG]

Nice .

[annihilat0r]

Great job

[Carfax]

lol I have heard these exact arguments about the R600 when it first came out. Easily add 20%? It might happen, but not easily. You act like drivers are somehow as magical as steve jobs' iPad

I don't know about ATI, but this is fairly routine with Nvidia.

The G200's release drivers were the 170xx, but it wasn't until the 180xx drivers that the architecture received it's first significant boost....to the tune of 10 > 15%.. Of course the recent 190xx drivers added even more, so all in all I'd say my GTX 285 is about 20% faster on average right now than it was on release day.

And remember that the G200 isn't as radical a change in architecture as Fermi is..

[96redformula]

Let's make easier to compare

Scores normalized to HD 5850 = 100%

Only 1920x1200 AA 4x e AA 8x

Now that's a slick graph. That makes comparing cards across the board very nice to see IMO.

[annihilat0r]

Would be better if it was normalized to 5870 I think, looks like the 470 is competing with it also.

[Carfax]

then get another 30inch monitor and enjoy nvidia's eyefinity

Yeah, my wife would love me sinking 2,500 bucks on something she knows I don't "need."

[metro.cl]

I wouldn't say it reflects market share more than anything else. I'd say market share might be a small factor though. Regardless of how much, the end result is still that nv's drivers were inferior to ATi's, so the anti-ATi driver crusade has no merit.



When the earthquake hit Chile, you were the first one I thought of (i'm not familiar with anyone else from Chile). I hope things are OK down there for you and yours.

Hi Mate,

Thnks for asking, yes we are doing OK, i'm far away from the epicenter, like 500KMs but we felt it pretty hard like 7.5-8.0 richter. For this huge earthquake we did pretty good overall.

Regards

[MrDiaz]

Yeah, my wife would love me sinking 2,500 bucks on something she knows I don't "need."

We hear that all the time...

[japamd]

Would be better if it was normalized to 5870 I think, looks like the 470 is competing with it also.

Done

[Ramon Zarat]

The GF100 architecture is VERY different from anything that has been produced by Nvidia before. So much so, the emphasis is clearly put on GPGPU performance, not gaming. Therefore, I don't expect tremendous driver optimization to be even possible this time around. At least, surely not in the almost "magical" 20-30% range. And I support the view that Nvidia final silicon has been available at least 3-4 months ago to the driver developer. I believe the drivers to be quite refined already.

So, no dual GPU, no clock increase unless you go water cooling, little to no drivers optimization: I'll say it again: Unless Nvidia comes out with a REV2 silicon that draw far less juice and allow 512sp at higher clock (very doubtful in the short and even middle term), performance wise, what you'll see on the 26 is what you'll get for a VERY long time from these cards.

I really hope for Nvidia sake that the 480 is way over 20% better than the 5870 in the vast majority of benchmark because ATI will simply strike back with a 2GB 1Ghz 5890, undercut 480 price (RV870 is a LOT cheaper manufacture than GF100), and that's gone be the end of it.

[Roger_D25]

JapAMD - Thanks for taking the time to put those graphs together, much appreciated!

[eric66]

battleforge results are unbelievable though never thought that ati will loose so much in its own favorite game lol

[ShadowFox19]

GTX480...come to Butthead.

[Yvese]

The GF100 architecture is VERY different from anything that has been produced by Nvidia before. So much so, the emphasis is clearly put on GPGPU performance, not gaming. Therefore, I don't expect tremendous driver optimization to be even possible this time around. At least, surely not in the almost "magical" 20-30% range. And I support the view that Nvidia final silicon has been available at least 3-4 months ago to the driver developer. I believe the drivers to be quite refined already.

So, no dual GPU, no clock increase unless you go water cooling, little to no drivers optimization: I'll say it again: Unless Nvidia comes out with a REV2 silicon that draw far less juice and allow 512sp at higher clock (very doubtful in the short and even middle term), performance wise, what you'll see on the 26 is what you'll get for a VERY long time from these cards.

I really hope for Nvidia sake that the 480 is way over 20% better than the 5870 in the vast majority of benchmark because ATI will simply strike back with a 2GB 1Ghz 5890, undercut 480 price (RV870 is a LOT cheaper manufacture than GF100), and that's gone be the end of it.

Not really. Nvidia fanboys have taught us they'll buy anything even if it's overpriced. The GTX470 comes to mind. Same MSRP as the 5870, yet performs less and people are still willing to jump all over it.

You can't stop the green fanboys.

[Dami3n]

battleforge results are unbelievable though never thought that ati will loose so much in its own favorite game lol

Ati winning in Crysis and Vantage, and nvidia winning in battleforge and Dirt2... surely the end of the world as we know it is near

[Sushi Warrior]

The GF100 architecture is VERY different from anything that has been produced by Nvidia before. So much so, the emphasis is clearly put on GPGPU performance, not gaming. Therefore, I don't expect tremendous driver optimization to be even possible this time around. At least, surely not in the almost "magical" 20-30% range. And I support the view that Nvidia final silicon has been available at least 3-4 months ago to the driver developer. I believe the drivers to be quite refined already.

So, no dual GPU, no clock increase unless you go water cooling, little to no drivers optimization: I'll say it again: Unless Nvidia comes out with a REV2 silicon that draw far less juice and allow 512sp at higher clock (very doubtful in the short and even middle term), performance wise, what you'll see on the 26 is what you'll get for a VERY long time from these cards.

I really hope for Nvidia sake that the 480 is way over 20% better than the 5870 in the vast majority of benchmark because ATI will simply strike back with a 2GB 1Ghz 5890, undercut 480 price (RV870 is a LOT cheaper manufacture than GF100), and that's gone be the end of it.

Don't you think that would make it more likely for better drivers? It looks really impressive on paper but the FPS is not reflecting that, maybe it just isn't using all of its power.

[Chumbucket843]

if the leaks are right here is a graphic of the gtx480 i made bc i have too much free time. it has a 24% advantage overall so it could sell at $480vand offer same perf/dollar as 5870. i would have included 5970 but im having trouble finding reviews with these games.

idk why -1% ends up +1% on crysis.

[***Deimos***]

From Nvidia:
[IMG*]http://img10.imageshack.us/img10/4164/73529807.jpg[/IMG]
[IMG*]http://img522.imageshack.us/img522/6378/gtx4802.jpg[/IMG]

These look *very* real. The other blue/gray graphs look totally fake.

Far Cry 2 2560x1600 4x/16x: 40.5% advantage
Batman AA 2560x1600 4x/16x: 39.3% advantage
Stalker Clear Sky 2560x1600 4x/16x: 27.4% advantage
Battleforge 2560x1600 4x/16x: 23.5% advantage
Aliens vs. Predator 2560x1600 4x/16x: 25% advantage
Dirt 2 2560x1600 4x/16x: 10.4% advantage

1. DX11 games do very well. Very very good sign.
2. Fps remains high at high AA/2560. 99% sure new drivers will improve Battleforge and Dirt2 further.
3. Fermi has bigger lead over 58xx than G200 had over 48xx!!
4. Air OC wont get GTX470 to catchup to GTX480 - worse PCB, power ceiling and gap is too big. GTX470 HIGHER than 5870 would be HUGE WIN for nVidia.
5. Half the games are quite CPU limited (compare teh SLI Fermi score).
6. 2GB 5870 OC wont do much. Performance is NOT proportional to clock. Even +18% 1000Mhz (possible with new PCB), would only typically get +5-7%fps
7. But, what if these are non-retail scores!! (ie 512SP 750Mhz)

PS: 700Mhz for HUGE die is impressive. Still NOBODY knows how yield situation will play out. Perhaps even a special "39nm" respin stepping with <225W.

2 to 3 months maybe less....GF110 is taped out

Sigh of relief? What if similar delays to Fermi? Silicon respins? RV770 - ie best case scenario - was like 140days = 4+ months. Same with Redwood.

No rumours or slides yet... will be interesting to see configuration (guessing800/1600/1200 256SP 128bit??)
EDIT: I predicted driver optimization contraversies before - nVidia definetly wants to be #1 EVERYWHERE.. IQ still unknown.. for all we know that +40% in FC2 is because of some glitches...

[Ramon Zarat]

Don't you think that would make it more likely for better drivers? It looks really impressive on paper but the FPS is not reflecting that, maybe it just isn't using all of its power.

You have a point. The new architecture is maybe so alien that even after 4 month of hard work from the Nvidia driver team, there is still 20-30% of untaped power to be unbleached.