[wccftech] Another leaked Kepler performance slide 760/770/780/790...

[Arzachel]

Do they give any reasoning? Or are they just calling "fake" out of hand?
What are they saying?

I think the slide being fake is a smaller leap of logic than Nvidia droping hotclocks and trying to fit 2304 Cuda cores on a die, which is plain impossible on 28nm unless they overhaul the arch.

 

[MrTeal]

I think the slide being fake is a smaller leap of logic than Nvidia droping hotclocks and trying to fit 2304 Cuda cores on a die, which is plain impossible on 28nm unless they overhaul the arch.

If the slides are true, Kepler would have to be by far the biggest GPU die ever produced, and yields would be beyond terrible. Even if they could reduce the size off each CUDA core by a 1/3 on the same node just through arch tweaks, AND they had perfect scaling going to 28nm, the core section would still be 50% larger in mm^2 than it is in Fermi. After Fermi, I can't imagine anyone at NVIDIA not currently sitting in a methadone clinic has any desire to start playing with 600+mm^2 (up to more like 1000mm^2 realistically) dies on a brand new node.

Either that or Kepler's a ground up redesign and a new CUDA core has little in common with the existing products. Like you said, until I see differently I'm going to assume this table is fake, or at least parts of it are.

 

[boxleitnerb]

Hotclocking costs alot of die space, so I have read. Let's say they lose 40% and another 40% with the 28nm process, 2000 shaders should fit quite nicely.

 

[3DVagabond]

Some are assuming that Kepler shaders are equivalent in size to Fermi.

 

[boxleitnerb]

Who?

 

[formulav8]

Unrealistic expectations in that slide. How could anything 1 gen newer be over 100% faster than last gen flagship. That hasn't happened in a long time.

^
That

The die would be Very Huge

 

[Arzachel]

Some are assuming that Kepler shaders are equivalent in size to Fermi.

(Re-)Read MrTeal's post, he said it far better than I can.

Also, Nvidia has said absolutely nothing to indicate, that Kepler is anything more than refinement of Fermi.

 

[Arkadrel]

Hotclocking costs alot of die space, so I have read. Let's say they lose 40% and another 40% with the 28nm process, 2000 shaders should fit quite nicely.

I thought the idea behinde hotclocks, is that *if* you run something twice as fast, you need less of them to do the same work?

In short, nvidia uses hotclocking to keep their chip size down, at a slight cost to performance/watt.
Otherwise why do it?

 

[boxleitnerb]

I thought the idea behinde hotclocks, is that *if* you run something twice as fast, you need less of them to do the same work?

In short, nvidia uses hotclocking to keep their chip size down, at a slight cost to performance/watt.
Otherwise why do it?

Yes, the absolute number is smaller, but maybe you need to make each ALU larger (compared to lower clocked ones) to drive it to a higher frequency?

News from OBR, take it with a grain of salt:

WWW.OBR-HARDWARE.COM - OBRovsky Blog

 

[AdamK47]

Unless there are good sources, I don't buy it. These sites post BS specs just to get visitors to their site. It works.

 

[Arkadrel]

@boxleitnerb

Normally Id say because its ORB, its likely to completly fake and made for page hits. However that looks pretty reasonable I think, or atleast like its something that could possibly happend (unlike the other charts/slides posted before in here).

A 104GK like that would probably be a tiny bit slower than a 7970-7950 or so.
A competitor for the rumored 7890?

It ll probably be 7890 @~350$ and this GK104 abit faster @~399$.

 

[Lonyo]

@boxleitnerb

Normally Id say because its ORB, its likely to completly fake and made for page hits. However that looks pretty reasonable I think, or atleast like its something that could possibly happend (unlike the other charts/slides posted before in here).

A 104GK like that would probably be a tiny bit slower than a 7970-7950 or so.
A competitor for the rumored 7890?

It ll probably be 7890 @~350$ and this GK104 abit faster @~399$.

HD7950 is $399 for the 1.5GB version. If GK104 was slower, then it would need to be less.

 

[Vesku]

I think from a consumer perspective it will be great if NVIDIA continues to offer the biggest die they can get on a node. AMD has clearly opted for a moderate die approach so it's left to NVIDIA in the big as it can get department.

It would be quite the coup if they can get anything like the performance gains going around the rumor mill, though.

 

[Arkadrel]

Valid point...
Your probably right, might be closer to the 7950 or above it.

 

[Lonyo]

It isn't great to compare the the optimum 55nm GPU from NV to its first gen 40nm because of the major issues NV had with 40nm. It kind of assumes NV will have similar troubles with 28nm to their 40nm troubles.

Worst case scenario they might struggle, but consider the improvement from the GTX285 to the GTX580, 55nm to 40nm, and if they manage to repeat that going from 40nm to 28nm they could do quite a lot more than looking at the 285 to 480 would suggest.

 

[Vesku]

That's a good point Lonyo, I made the comparison more to try to dampen the pre-release uproar about how the 7970 didn't trounce the 580 enough. IMO, that forum conversation never really settled on a logically consistent argument and I still think it mainly flared up due to people's fears of what 780 MSRP might be if the 7970 stays in the ~$500 range.

The size and MSRP are the same for the 480 and 580 so you'd just have to add in the 8-15% gain to the performance (~50% from 285?). Maybe I should have cut out a little square to indicate the disabled section of the 480?

 

[3DVagabond]

WWW.OBR-HARDWARE.COM - OBRovsky Blog[/url

@[URL="http://forums.anandtech.com/member.php?u=300481"]boxleitnerb

Normally Id say because its ORB, its likely to completly fake and made for page hits. However that looks pretty reasonable I think, or atleast like its something that could possibly happend (unlike the other charts/slides posted before in here).

A 104GK like that would probably be a tiny bit slower than a 7970-7950 or so.
A competitor for the rumored 7890?

It ll probably be 7890 @~350$ and this GK104 abit faster @~399$.

With hotclocks it should be a bit slower than the 7900. How about without? 12.5% more shaders @ ~59% of the speed. If that's the case this chip isn't going to even be in the same neighborhood as the 580, never mind the 7900's. It would have to be monstrously more efficient.

 

[Vesku]

There was talk of an enhanced Fermi being the 560 equivalent for the new series, that "GK104" certainly looks to be a die shrink+. Wouldn't rule out hotclocks entirely.

Will be interesting to see how NVIDIA does with the flagship 28nm part.

 

[Arkadrel]

@3DVagabond

Do you honestly believe Nvidia is gonna drop its hotclocks? to live up to claims about GPGPU compute performance/watt?

I dont think nvidia will use more shaders, that are smaller/simplier than their current fermi ones either. To me that would be a step backwards for them, unless that way of doing things proves to be more effecient (maybe THEN nvidia could take a approch more akin to AMDs, but I doubt that ll happend, because so far their design philosopy is able to keep up fine in performance/watt, sure its abit below AMDs but not by insane margins).

That Spec sheet ORB either heard about, or made up, looks very possible at the very least
(id be assumeing it has hotclocks as well, and performs abit under a 7950, based on those specs).

 

[3DVagabond]

@3DVagabond

Do you honestly believe Nvidia is gonna drop its hotclocks? to live up to claims about GPGPU compute performance/watt?

I dont think nvidia will use more shaders, that are smaller/simplier than their current fermi ones either. To me that would be a step backwards for them, unless that way of doing things proves to be more effecient (maybe THEN nvidia could take a approch more akin to AMDs, but I doubt that ll happend, because so far their design philosopy is able to keep up fine in performance/watt, sure its abit below AMDs but not by insane margins).

That Spec sheet ORB either heard about, or made up, looks very possible at the very least
(id be assumeing it has hotclocks as well, and performs abit under a 7950, based on those specs).

Obviously, I don't know for certain whether or not they're dropping hotclocks, but, yes, I do believe they are dropping hotclocks. Maybe for more power efficiency, as you say. Or, maybe because the clock speeds needed are unattainable. We'll have to wait and see.

 

[BallaTheFeared]

Someone said cuda cores don't make a huge difference in (gaming) performance, it's more about TMU's and ROP's.

Can anyone confirm that?

 

[Arkadrel]

@BellaThefear

Not sure if this is right or wrong, but.... how about trying to keep core clocks @ stock, and lowering your shader clocks? really low, and see if performance drops from it?

You could test it yourself with your 470.
Download and run 3Dmark, once with stock everything, and once with stock everything but shaders down clocked alot, and looking at the 2 results.

 

[BallaTheFeared]

I can only do that with my 9800GT.

Fermi is locked at x2.

 

[Arkadrel]

I googled <.<

this was all I could find:

from 1242 -> 1566 (~26&#37; increase), results in a ~7% increase here.

However I suspect it ll depend on the game, and whats the limiting factor in fps in said game.
Maybe Farcry was one of those games where it didnt effect it much.

 

[3DVagabond]

I googled <.<

this was all I could find:

from 1242 -> 1566 (~26% increase), results in a ~7% increase here.

However I suspect it ll depend on the game, and whats the limiting factor in fps in said game.
Maybe Farcry was one of those games where it didnt effect it much.

Or there could have been a bottleneck elsewhere holding it back. It's something, but not enough info.