NVIDIA Kepler speculation thread

[Picao84]

Au contraire, IMO this fits nVidia's mindset very well. What's the most significant competetive advantage nV currently has over AMD? It's PhysX! And if they can't achieve advantages in other areas, like raw rendering power, then what do you do?

Charlie wasn't the first with this rumour anyways: http://news.techeye.net/chips/nvidias-kepler-suffers-wobbly-perturbations

BTW: Do you guys remember the BF3 slide showing massive advantages of Kepler vs. Fermi? : http://wccftech.com/nvidias-kepler-...orce-500-series-generation-performance-chart/

Notice what it says in the lower right corner......... interesting, isn't it?

All that only makes sense if Physx starts to be synonymous with something else, other than their proprietary physx engine, since Battlefield 3 does not use Physx...

 

[CarstenS]

BTW: Do you guys remember the BF3 slide showing massive advantages of Kepler vs. Fermi? : http://wccftech.com/nvidias-kepler-...orce-500-series-generation-performance-chart/

Notice what it says in the lower right corner......... interesting, isn't it?

So... does that mean you consider Battlefield a soon-to-be-physx'ed title? All their in-house destruction stuff thrown overboard and all? I have a hard time to believe that.

 

[Kaotik]

Au contraire, IMO this fits nVidia's mindset very well. What's the most significant competetive advantage nV currently has over AMD? It's PhysX! And if they can't achieve advantages in other areas, like raw rendering power, then what do you do?

Charlie wasn't the first with this rumour anyways: http://news.techeye.net/chips/nvidias-kepler-suffers-wobbly-perturbations

BTW: Do you guys remember the BF3 slide showing massive advantages of Kepler vs. Fermi? : http://wccftech.com/nvidias-kepler-...orce-500-series-generation-performance-chart/

Notice what it says in the lower right corner......... interesting, isn't it?

You mean the fake slide?

 

[DarthShader]

I am only pointing out an observation, since we had a round of lulz when that slide first appeared.

If we had Crysis2 being DX11'ed post release, being a high prestige title, then I am guessing it could be possible another high prestige title, BF3, get's Physx'ed later on. And since not all cards will do great with it, the in-house destruction stuff won't go to waste. Heck, the PhysX effects could be something completely new added, water, smoke, moar particles etc.

Edit: Sure, my dot connecting could be way wrong here.

 

[Arty]

That is what we refer to as double confirmation of a rumor, dont we?

 

[silent_guy]

If they add compute related features that also benefit PhysX and market it as improved PhysX HW acceleration, I say: yes, please.

 

[DavidGraham]

The only way this could be true ,is if NVIDIA planned to accelerate all physics engine on the market , which sounds even more ridiculous than Carlie's bullsh!t !

 

[ninelven]

This all seems like a non-story. Did people expect GPUs not to improve at various workloads or to get worse?

 

[silent_guy]

With the loss of the so called “Hot Clocked” shaders, this leaves two main paths to go down, two CUs plus hardware PhysX unit or three. Since there is no dedicated hardware physics block, the math says each shader unit will probably do two SP FLOPs per clock or one DP FLOPs.

I'm speechless...

Does he use the MIT automatic paper generator?
http://pdos.csail.mit.edu/scigen/

 

[psurge]

If these (alleged) new features drastically improve e.g. OpenCL performance across the board on Kepler HW, I don't see how that's ridiculous.

Another possibility is that new features might allow for different implementation strategies/algorithms that previously resulted in poor performance (e.g. one can fully utilize the GPU with a smaller thread count, irregular control flow has less of a penalty, etc...). In other words, PhysX compute kernels might be expressible in a portable way, but not in a performance portable way.

Finally, extra compute features could simply be unavailable outside of a new cuda rev that requires >= Kepler HW. In that case, if I were a physics engine developer, writing a custom code path for an initially tiny slice of one GPU maker's install base probably wouldn't be at the top of my todo list.

Not that I'm buying Charlie's article, mind you...

 

[Alexko]

It actually sounds a lot more like he doesn't have a clue. It'll be faster, unless it's slower.

Well he does say:

"Benchmarks for GK104 shown to SemiAccurate have the card running about 10-20% slower than Tahiti."

 

[AlphaWolf]

Well he does say:

"Benchmarks for GK104 shown to SemiAccurate have the card running about 10-20% slower than Tahiti."

sure, right before he says

That brings up the fundamental question of whether the card is really performing to that level,

 

[psurge]

And then he also says:

All of the benchmark numbers shown by Nvidia, and later to SemiAccurate, were overwhelmingly positive. How overwhelmingly positive? Far faster than an AMD HD7970/Tahiti for a chip with far less die area and power use, and it blew an overclocked 580GTX out of the water by unbelievable margins.

It's totally discombobulated.

 

[rpg.314]

If it performs like Tahiti, then why is it priced like Pitcairn? It's not like nvidia is hurting for marketshare. Serious question. Could it be that they anticipate AMD unlocking some overclocking headroom in near future? Even if they don't now, it seems 28 nm refresh would clock much better than now.

As for the PhysX block. Ha bloody ha.

It's likely to be a bunch of instructions (by themselves or acting as an interface to a ff block) to manage irregular data structures/mem ops.

It can't be much smaller than ~250 mm2, otherwise it would be pad limited. But being ~30% smaller on a brand new process can save them a lot of money.

 

[trinibwoy]

If it performs like Tahiti, then why is it priced like Pitcairn?

Let's assume it performs like Tahiti. What should they price it as to leave room for GK110?

 

[rpg.314]

Let's assume it performs like Tahiti.

Alright.

What should they price it as to leave room for GK110?

In that case, GK110 is the alpha dog and has unbounded room for price. It will be sold in larger quantities as Tesla and Quadro anyway.

 

[thatdude90210]

Wasn't there a GTX ultra in the 8800 gen that was ridiculously priced?

 

[AlphaWolf]

Wasn't there a GTX ultra in the 8800 gen that was ridiculously priced?

It must have been something more than ridiculous as apparently $550 is ridiculous. It was $830.

 

[Arty]

It can't be much smaller than ~250 mm2, otherwise it would be pad limited. But being ~30% smaller on a brand new process can save them a lot of money.

G71 was 196mm2, 256-bit ASIC.

 

[AlphaWolf]

G71 was 196mm2, 256-bit ASIC.

I recall some debate about the size of G71, but doesn't matter. RV670 was even smaller at 190mm2.