The Fermi Thread - Part 3

[Milos-stancene]

I want to put an end to the "dead horse beating".

R600 vs Fermi
both: hot, new DX, late
diff: Fermi is faster, not broken, amazing DX11 features/performance for future games, good MSRP (markup another story).

fx vs Fermi
both: hot, new DX
diff: Fermi is faster, not broken, no questionable "optimization" (yet) and doesn't sound like turbine

GeForce vs Fermi
both: hot, new DX, foundation for new technology lineup.

Basically,
ALL new DX cards are always hotter. That's the trend. Get over it.
Even RV870 was gotter than RV770. Its not somehow a new issue only Fermi has.

Somehow, in the rosy glamourous faded memories, many people forget 9700pro was ridiculed for requiring extra power connector with ATI's classic Win2000 driver issues and at best performance on par with GF4. Ofcourse a year later, with issues fixed and DX9 launced, everybody was buying one.


Die shrink will be too late. 512 SP not important. Its like 5% difference.

There is 1, and only 1, critical "to-do".
DEVELOPERS.

If Fermi can get +20% in BattleForge, an "AMD game", image the performance if it was TWIMTBP!
Developer support is more critical than ever. All the consoles are DX9. There's little motivation to do "hard work" to make DX11 games.
It doesnt matter if its 480 SP or 512 SP, if nVidia isn't there to baby-hold and show 1-2-3 how to use it. Because if it never gets used, there's no benefit of having it.

nicely said

[LiquidReactor]

I don't think you'd screw up something that fundamental such as fillrate with release drivers.

What drivers do optimize are probably shader code and texture formats, perhaps a bit of how games render here and there. Feature tests however should be extremely close to practical values from day 1.

So....is Fermi supposed to be a mediocre gpu yet a great gpgpu?

[Mungri]

So....is Fermi supposed to be a mediocre gpu yet a great gpgpu?

Nope:

[acicula]

Nope:

it does better in tests that involve floating point stuff(this is pure integer stuff)

also cuda 3.0 support seems pretty awesome with unified memory architecture(awesome to program)

edit in general gpgpu application perf is very dependent on the application, meaning there is no fastests gpu for gpgu as each have applications in which they are strong due to the architecture.

[LiquidReactor]

Nope:

That is just one test. I've seen tests where the performance difference is as lopsided but favoring Fermi / GTX 480. If you can't post without bias, don't post.

[Mungri]

Oh, like in folding:



So you can fold while making hamburgers

[ElSel10]

Oh, like in folding:



So you can fold while making hamburgers

You seem to be a troll. The password cracking is an extremely simple GPGPU application, probably the only one that can make use of the theoretical 2.7TFlop of the 5870 (although 2.2Tflop in practice) and remain useful.

Let's have a look at some more:

[Mungri]

I just dont know anything about GPGPU. Ive never been bothered about it.

[Macadamia]

You seem to be a troll. The password cracking is an extremely simple GPGPU application, probably the only one that can make use of the theoretical 2.7TFlop of the 5870 (although 2.2Tflop in practice) and remain useful.

Let's have a look at some more:


[images snipped]

Urgh purleeze. The version that Anandtech used 1) Didn't have any vectorization optimizations/experimentation 2) Was not done with the -local switch. nVidia's OCL compiler probably adds it automatically by now, you need to specify it on the Radeons.

But again you don't seem to know much more about GPGPU either, with the cherrypicking of benchmarks that suit your agenda.

[E30M3]

Any ideas on how much drivers might improve Fermi arch performance?

20-30% is what they normally gets

[ElSel10]

Urgh purleeze. The version that Anandtech used 1) Didn't have any vectorization optimizations/experimentation 2) Was not done with the -local switch. nVidia's OCL compiler probably adds it automatically by now, you need to specify it on the Radeons.

But again you don't seem to know much more about GPGPU either, with the cherrypicking of benchmarks that suit your agenda.

Cherry picking? I posted the rest of the benchmarks from anandtech that were not already posted.

So you don't know for sure that nvidia adds it? Let's say it does. I thought OpenCL was ATI's forte? Not only did nvidia have full support before ATI but they appear to be faster as well.

Do you have a 5870? Why don't you or somebody that has one run this benchmark with the -local switch then?

[Sushi Warrior]

The really, really annoying thing is that I feel like I am required to defend Nvidia because the ATI fanboi-ism is SO blatant. It wasn't an epic win for Nvidia so everyone who likes ATI interprets that as "FAIL FAIL FAIL". It isn't a black and white win or lose, if you want FPS (and to a much lesser extent FPS/$) go GTX470/480. If you want efficiency, go 5870/5850. If you want ULTIMATE, go 5970....

And on the whole PhysX thing, it is an advantage.... because Nvidia has a decent lineup of PhysX games (3-5 good ones) while ATI doesn't have any games with exclusive physics... something is better than nothing.

[Movieman]

nVidia reference board.

EDIT : Here's the stock untouched GTX480 BIOS. Nibitor doesn't seem to work but maybe someone can figure out the way to mod it.

http://www.megaupload.com/?d=1SIHTCTP

edit2:

Vantage default at 825/2000 with the highest recorded temps.

How'd you get a cpu score of 68,000+ in Vantage with a bloomfield at 4300?

[N19h7m4r3]

How'd you get a cpu score of 68,000+ in Vantage with a bloomfield at 4300?

PhysX Enabled on the GPU greatly boosts the CPU score in Vantage

[NaMcO]

How'd you get a cpu score of 68,000+ in Vantage with a bloomfield at 4300?

The Physics test is probably running on the GTX480 and that's an incredible boost.

[Mungri]

ATI fanboys sure are Intelligent people.
bad losers???

GPGPU is just another gimmick that I have no interest in, like Physx and 3D vision. Simply not wanting to read about things that I dont need doesnt imply anything about my intelligence.

And I wouldnt ever state that ATi have in anyway lost to the Fermi, in fact I think its the other way around. ATi sales actually increased since friday because plenty of people who were waiting for Fermi saw the results and chose to buy a 5850 / 5870 instead.

Oh, and I'm not an ATIfanboy. Ive owned several Nvidia cards up untill the HD 3000 range was released, at which I decided to start buying based on price to performance.

[Movieman]

PhysX Enabled on the GPU greatly boosts the CPU score in Vantage

That's what I thought but isn't it supposed to be run with it off so all are equal?

[ElSel10]

GPGPU is just another gimmick that I have no interest in, like Physx and 3D vision. Simply not wanting to read about things that I dont need doesnt imply anything about my intelligence.

You just called GPGPU a gimmick, fyi.

[Luka_Aveiro]

That's what I thought but isn't it supposed to be run with it off so all are equal?

Yes, but that's an user's score, not a Reviewer's one.

He forgot do disable PhysX in nVidia drivers.

ORB doesn't take account of PhysX improved GPU scores.

[Movieman]

Yes, but that's an user's score, not a Reviewer's one.

He forgot do disable PhysX in nVidia drivers.

ORB doesn't take account of PhysX improved GPU scores.

Ok, but then doesn't the number become meaningless unless the person specifically states that it's off or on?
See where I'm going here?

[Movieman]

I've PM'd bhavv to remove himself from this thread.
If anyone sees another post from him here please PM me and I'll send him on a nice vacation..

[Sushi Warrior]

Oh, and I'm not an ATIfanboy. Ive owned several Nvidia cards up untill the HD 3000 range was released, at which I decided to start buying based on price to performance.

You may not be an ATI fanboi but I would certainly consider you an ATI troll.

Did anyone mention Nvidia's lack of double-redundancy between layers? Apparently that is the cause of the horrible leakage, because TSMC isn't able to properly make the metal-metal interconnects and ATI added double redundancy due to experience with RV740.

[Luka_Aveiro]

Ok, but then doesn't the number become meaningless unless the person specifically states that it's off or on?
See where I'm going here?

GPU scores are not affected by PhysX activated, only CPU score and consequentially overall score.

So, GPU score is the one and only vaild score you can count with to have fair comparison...

[570091D]

Nope:

I just dont know anything about GPGPU. Ive never been bothered about it.

'nuff said.

You may not be an ATI fanboi but I would certainly consider you an ATI troll.

Did anyone mention Nvidia's lack of double-redundancy between layers? Apparently that is the cause of the horrible leakage, because TSMC isn't able to properly make the metal-metal interconnects and ATI added double redundancy due to experience with RV740.

i believe you're thinking of the vias, when tsmc first began producing 40nm wafers they had problems with via failure and channel width. the via problem has since been fixed (amd's rv740 had double vias), but the channel width is still too large. wide channels cause leakage and extra heat, but can also provide A LOT of overclocking headroom (think amd's TWKR chips) with the right cooling. i'm looking forward to k|ngp|n's ln2 results with the 480's.

btw, word on the street is that Bx revisions of gf100 will fix the channel width problem.

[***Deimos***]

the card crashes at 875Mhz almost instantly in '06, at 850Mhz it crashes in the second test but it should do 835/840Mhz core. I haven't pushed the memory yet. I'll do that now.

So many times I prediced that GTX 480 would barely OC. That nVidia would have an impossible time reaching 750Mhz.

If so many people are getting 800Mhz+, perhaps 480SP and 700Mhz had more to do with limiting power consumption than yields.

If you had chips easily able to do 800Mhz, why else would you sell them at 700Mhz?

By using LN2 cooling on a GeForce GTX 480 video card he was able to take the core clock from 700MHz up to 1165MHz, which is an amazing 66% overclock! He then used an Intel Xeon E5640 processor at 4GHz and was able to score just shy of 30,000 3DMarks in the Performance test of 3DMark Vantage

Holy Smokes Batman. Can somebody do the math on how much current must be going through those 6 voltage regulation circuits!!?

Me thinks at whatever ungodly voltage Shamino used for 1165, "normal" 1200W-1500W PSU could be not enough for the system.

At stock voltage it seems LN2 alone gets 900Mhz on the core - ie cooling seems to be the main issue

http://hwbot.org/hardware/videocard/geforce_gtx_480