GTX680 pics

[Silverforce11]

A full node shrink at the same die size will typically give 60-80% performance increases. The fact that gk104 is very small should already be hint enough that its performance is not going to be anywhere near that.

Heck, you can assume perfect scaling and make that a 100% increase, it still wouldn't put gk104 at gtx590 levels like some of the absurdness being posted here.

If its near 7970 speed is a good design for NV. Better perf/mm and perf/w is not something NV has been able to produce, for them to do so on a new process, its amazing.

Edit: Transistor count: http://www.techpowerup.com/162341/GK104-Transistor-Count-and-Exact-Die-Size-Revealed.html

Notice how the 6970 -> 7970, there's an increase of ~63% on transistors. It roughly yields an overall 40% increase in performance, approaching >60% in newer dx11 games. Then look at the the gtx580 -> gk104 transition, a ~12% increase. Not a perfect comparison or correlation, but enough to hint at whats coming.

 

[blackened23]

A full node shrink at the same die size will typically give 60-80% performance increases. The fact that gk104 is very small should already be hint enough that its performance is not going to be anywhere near that.

Heck, you can assume perfect scaling and make that a 100% increase, it still wouldn't put gk104 at gtx590 levels like some of the absurdness being posted here.

If its near 7970 speed is a good design for NV. Better perf/mm and perf/w is not something NV has been able to produce, for them to do so on a new process, its amazing.

Edit: Transistor count: http://www.techpowerup.com/162341/GK104-Transistor-Count-and-Exact-Die-Size-Revealed.html

Notice how the 6970 -> 7970, there's an increase of ~63% on transistors. It roughly yields an overall 40% increase in performance, approaching >60% in newer dx11 games. Then look at the the gtx580 -> gk104 transition, a ~12% increase. Not a perfect comparison or correlation, but enough to hint at whats coming.

I think i'm done speculating and looking at possibly fake charts, next thursday needs to get here faster. :thumbsup:

 

[Keysplayr]

A full node shrink at the same die size will typically give 60-80% performance increases. The fact that gk104 is very small should already be hint enough that its performance is not going to be anywhere near that.

Heck, you can assume perfect scaling and make that a 100% increase, it still wouldn't put gk104 at gtx590 levels like some of the absurdness being posted here.

If its near 7970 speed is a good design for NV. Better perf/mm and perf/w is not something NV has been able to produce, for them to do so on a new process, its amazing.

Edit: Transistor count: http://www.techpowerup.com/162341/GK104-Transistor-Count-and-Exact-Die-Size-Revealed.html

Notice how the 6970 -> 7970, there's an increase of ~63% on transistors. It roughly yields an overall 40% increase in performance, approaching >60% in newer dx11 games. Then look at the the gtx580 -> gk104 transition, a ~12% increase. Not a perfect comparison or correlation, but enough to hint at whats coming.

This is of course assuming that 6970 or 7970 has anything to do with GTX580 or gk104.

 

[blackened23]

This is of course assuming that 6970 or 7970 has anything to do with GTX580 or gk104.

Are you saying that transistor count isn't indicative of performance? It doesn't always scale linearly, however performance always scales with transistor count on a given architecture.

 

[Keysplayr]

Are you saying that transistor count isn't indicative of performance? It doesn't always scale linearly, however performance always scales with transistor count on a given architecture.

What I'm saying is, "This is of course assuming that 6970 or 7970 has anything to do with GTX580 or gk104."

No, transistor count isn't indicative of performance. You say yourself it doesn't scale linearly, and finally, on a given architecture? Isn't the 7970 a new GCN based architecture? Isn't Kepler supposed to be a new architecture? What's your point?

 

[BallaTheFeared]

But according to all the rumors Kepler is a redesigned uArch so how can you compare it to anything?

 

[Termie]

Well, here's more rampant speculation for you all. Based on the laptop Kepler numbers we've seen, I estimate the gk104 will be 50% faster than the gtx580, making it about 25% faster than the 7970.

Ok, I admit that this sounds a bit high, but I do think we can say that the ~1500 shaders are about half as efficient as the 580's, making it like a 750 shader 580.

 

[Grooveriding]

Expecting a 300mm2 die to perform 30% above a 340mm2 die is a bold expectation and pretty close to a pipe dream. It would be impressive enough if it just manages to match the 340mm2 die. The $300 sweet dream is over, now we are on to 100% improvements in perf/mm2.

 

[notty22]

Are you saying that transistor count isn't indicative of performance? It doesn't always scale linearly, however performance always scales with transistor count on a given architecture.

No it does not. There are many cases where it does and many where it does not.
Fore example the transistor count on the gtx 580 is actually lower than the gtx 480.
http://www.xbitlabs.com/articles/graphics/display/geforce-gtx-580_2.html

 

[Grooveriding]

No it does not. There are many cases where it does and many where it does not.
Fore example the transistor count on the gtx 580 is actually lower than the gtx 480.

Yes, there is a small difference, but the 480 has a disabled cluster that is enabled on the 580. That's why their die sizes are near identical but the 580 is faster.

 

[notty22]

It can be argued that GF110 was a tweaked gpu die.
There was obviously more areas fused off on a gtx 470 and its specs as having the same transistors as a gtx 480.
I'm of the opinion the GF110 was a tweaked design and had less transistors. Resulting in less leakage, better yields, higher clock speeds etc. while maintaining the Fermi architecture.
http://www.xbitlabs.com/articles/graphics/display/gigabyte-gf-gtx400_2.html

 

[blackened23]

So disabled transistors have a meaningful measurement to you? Interesting. Your comparison is not valid because you're not taking the disabled CUs into consideration.

 

[Grooveriding]

It can be argued that GF110 was a tweaked gpu die.
There was obviously more areas fused off on a gtx 470 and its specs as having the same transistors as a gtx 480.
I'm of the opinion the GF110 was a tweaked design and had less transistors. Resulting in less leakage, better yields, higher clock speeds etc. while maintaining the Fermi architecture.
http://www.xbitlabs.com/articles/graphics/display/gigabyte-gf-gtx400_2.html

Fair enough to have your opinion, but that is now wholly unrelated to what you were trying to say earlier comparing the 580 to 480 die size. The 580 and 480 show almost the same die size and a 6% performance disparity in favour of the 580 at the same clocks because the 580 has the extra cluster enabled.

Same goes for the 570 vs 470 and the situation is the same with the 570 vs 580 - they have the same die size as well, but the 580 as we all know is faster than a 570 because of disabled areas on the 570 chip.

 

[Silverforce11]

This is of course assuming that 6970 or 7970 has anything to do with GTX580 or gk104.

As I've said, not a perfect correlation but a good indication, die size, transistor count and density, tdp, all adds up.

If its gtx580 +20%, thats already an awesome design given its characteristics. Obviously not awesome pricing.. but meh.

 

[Tempered81]

 

[Medu]

But according to all the rumors Kepler is a redesigned uArch so how can you compare it to anything?

Is the GTX680 a GPGPU or has it been stripped of those parts to maximise it's GPU ability? That might explain how Nvidia have leap frogged AMD in performance per mm2 and performance per watt- although I suspect that NV are been a little more liberal with their TDP than AMD.

 

[ShadowOfMyself]

People using arguments like die size, power consumption and such to justify the 680 being a mid range card make me laugh... You realize all of that also applies to the 7970? By your logic, there are no high end cards this generation

 

[Bobisuruncle54]

People using arguments like die size, power consumption and such to justify the 680 being a mid range card make me laugh... You realize all of that also applies to the 7970? By your logic, there are no high end cards this generation

Exactly. The only factor that determines whether a graphics card is high end or not is the price.

 

[Silverforce11]

People using arguments like die size, power consumption and such to justify the 680 being a mid range card make me laugh... You realize all of that also applies to the 7970? By your logic, there are no high end cards this generation

There ARE NO high-end cards on 28nm so far. You didn't know this?

They just decided to charge you enthusiast pricing for a mid-range product.

 

[boxleitnerb]

Exactly. The only factor that determines whether a graphics card is high end or not is the price.

No, that is nonsense. If I make and sell a 7770 for $500 it is high end? The price is too dependent on supply/demand to be any accurate measure for the actual product placement.

A card that is 30% faster (at stock speeds) than the previous high end is not high end, it's a better refresh - speedwise that is.

Now IF that 30% increase becomes the standard over the next few generations, THEN we can lable those cards high end. But we cannot change the definition over night, that is just stupid. Cards with TDPs around 130W, with one additional power connector and a small PCB by todays standards were considered high end once, now they are found in the performance segment. That change of classification happened over time, though, not form one day to the next.

High end is what is (within reasonable margins) economically and technologically feasible, i.e. currently a 450-500mm2 die with an according performance increase over a smaller chip like the GK104.

Imagine this:
Had Nvidia brought a beefed up GTX460 (aka GTX560) in 2009, proclaimed it the new high end/enthusiast card and sold it at $549, wouldn't you have asked:
"That's it?"
"Where is my 60-70% performance increase over the GTX285?"
Wouldn't you have missed a card like the GTX480/580? I certainly would have!

 

[Gloomy]

500 dollars is high end territory. Deal with it nerd.

You can use any word you want, but an attack is an attack and we won't allow it.
-ViRGE

 

[GaiaHunter]

No, that is nonsense. If I make and sell a 7770 for $500 it is high end? The price is too dependent on supply/demand to be any accurate measure for the actual product placement.

A card that is 30% faster (at stock speeds) than the previous high end is not high end, it's a better refresh - speedwise that is.

Now IF that 30% increase becomes the standard over the next few generations, THEN we can lable those cards high end. But we cannot change the definition over night, that is just stupid. Cards with TDPs around 130W, with one additional power connector and a small PCB by todays standards were considered high end once, now they are found in the performance segment. That change of classification happened over time, though, not form one day to the next.

High end is what is (within reasonable margins) economically and technologically feasible, i.e. currently a 450-500mm2 die with an according performance increase over a smaller chip like the GK104.

Imagine this:
Had Nvidia brought a beefed up GTX460 (aka GTX560) in 2009, proclaimed it the new high end/enthusiast card and sold it at $549, wouldn't you have asked:
"That's it?"
"Where is my 60-70% performance increase over the GTX285?"
Wouldn't you have missed a card like the GTX480/580? I certainly would have!

If the fastest card available was the 7770, it would be high end.

High-end is the fastest card. And the faster card can command higher prices.

Replying to you imagination, NVIDIA could try to sell it for $549, but if AMD was selling the 6970 at that price and NVIDIA was also selling the GTX480, only idiots would buy the GTX560 at that price.

The 28nm >400 mm² will only be proved (both economically and technically) feasible when there is one out. Currently it isn't since it is being commercialized.

 

[boxleitnerb]

Why do you insist on making it about absolute performance, only? If you had an increase of just 5% each generation, would that be high end? We need to view these parts in relation to the older parts.

Since 2006 every generation (at least from Nvidia) brought a 60% performance increase over the fastest card. Now if you suddenly leave out the big chip you're telling me you're missing nothing in the lineup? A card with a 384bit interface, 3GB VRAM that beats the GTX580 by 60, better 70%. GK104 isn't that card.

I feel some of you just have too low standards and are content with anything the companies are putting in your lap and taking your hard earned money for. Sad.

500 dollars is high end territory. Deal with it nerd.

Nerd? Maybe. At least I'm not gullible and question what I actually get for that $500

 

[Keysplayr]

As I've said, not a perfect correlation but a good indication, die size, transistor count and density, tdp, all adds up.

If its gtx580 +20%, thats already an awesome design given its characteristics. Obviously not awesome pricing.. but meh.

And it all adds up to..? What? Ever consider that AMD had to add a bunch of transistors for their entry with a "Fermi-like" GPU? This is really their first true effort at GPGPU oriented chip. There are a lot of Xtors dedicated to GPGPU amiright? Nvidia started this back with G80 and have since made their strides in improving GPGPU and now probably made a breakthrough, not nearly needing as many Xtors to get the GPGPU job done and maybe even better than Fermi?
I really don't know why you gents don't consider these things. You didn't with Fermi either. You based ONLY gaming performance on die size (GTX480 vs HD5870 and GTX580 vs HD6970) but continually ignored GPGPU purposes when considering Fermi's larger die size. This has been happening since GTX280 I think.
Nvidia has been at this for a while. 7970 is AMD's first true attempt at a GPGPU. Yes, their prior cards provided SOME GPGPU functionality. Great at bit coin mining, Milkyway, a few specialized apps. Nothing like what Nv had. Now they are smartly trying to head in that direction.
IT's AMD's turn to have the bigger die size for a few gens while they figure out just what they need and don't need to have a kick ass GPGPU.

 

[KCfromNC]

And it all adds up to..? What? Ever consider that AMD had to add a bunch of transistors for their entry with a "Fermi-like" GPU? This is really their first true effort at GPGPU oriented chip. There are a lot of Xtors dedicated to GPGPU amiright? Nvidia started this back with G80 and have since made their strides in improving GPGPU and now probably made a breakthrough, not nearly needing as many Xtors to get the GPGPU job done and maybe even better than Fermi?

Could be. Maybe it's not. We'll know just as soon as nVidia manages to release a Kepler desktop card. Until that happens it's all just guessing.