NVIDIA Kepler speculation thread

[Ailuros]

Well if it manages to outperform a 7970 with only 60% of the bandwidth color me impressed.... I have a hard time believing that though.

That's what everyone is thinking when he first hits the specs. However the real juice is as always not in the sterile unit count, but deeper architectural changes no one but real insiders know at the moment. Just for the record I still don't believe it'll outperform a XT. Best case scenario IMHO a PRO.

 

[CarstenS]

That hardly correlates with "world's fastest DX11 GPU" now, does it?

 

[mczak]

Let's fill in some blanks then:

GTX680 = GK104
3*32SPs/SM

Hmm. Not 2*48? Or just 6x16?

That's what everyone is thinking when he first hits the specs. However the real juice is as always not in the sterile unit count, but deeper architectural changes no one but real insiders know at the moment. Just for the record I still don't believe it'll outperform a XT. Best case scenario IMHO a PRO.

Well a 7970 with 256bit memory bus would still be about as fast as a 7950, so even assuming similar bandwidth efficiency it's clearly possible. We'll see if it's really that good.

 

[Ailuros]

That hardly correlates with "world's fastest DX11 GPU" now, does it?

Remove the white block from the slide and it'll be obvious what marketing has been cooking.

Hmm. Not 2*48? Or just 6x16?

Or 12*8 or 24*4....

 

[LordEC911]

That hardly correlates with "world's fastest DX11 GPU" now, does it?

19x12 with 4x/16x
Haven't been paying attention to drivers but is Cat11.12 the current 7970 driver?

 

[Man from Atlantis]

replace old games with their successors
http://forums.overclockers.co.uk/showpost.php?p=21223264&postcount=214
use a GTX550Ti card picture for background

http://daisyoixpcsupport.blogspot.com/2011/03/nvidia-geforce-gtx-550-ti-redefines-pc.html
a little dirty msoffice/ps job voila.. no doubt slides are fake..

 

[trinibwoy]

It's fake!? No way!

 

[itsmydamnation]

It's fake!? No way!

i never saw that picture as being double confirmed........

 

[CarstenS]

Remove the white block from the slide and it'll be obvious what marketing has been cooking.

Ah, sure. The Radeons quite obviously are DX11.1 GPUs, so they don't count. *SCNR*

 

[Kaotik]

19x12 with 4x/16x
Haven't been paying attention to drivers but is Cat11.12 the current 7970 driver?

No, there's no Cat xx.yy for 7970 yet, just RC's with the actual driver build name instead of xx.yy designation

 

[CarstenS]

The 7970 launch driver actually thinks it is of the 11.12 breed.

 

[Otto Dafe]

Maybe NV decided to make a dual GK104 card to compete for the top slot while they work on GK110. That could make all the wild speculation sort of accurate.

</crazy hypothesis>

 

[Rangers]

It's fake!? No way!

My guess is any rumor/slide claiming Nvidia got rid of the hot clock is fake. Just a hunch.

 

[whitetiger]

My guess is any rumor/slide claiming Nvidia got rid of the hot clock is fake. Just a hunch.

I would say that the hot-clock is a failed idea, and they should drop it
- it must burn a lot of energy, and cause them masses of problem with layout
- therefore, hopefully, the rumors are true!

 

[A1xLLcqAgt0qc2RyMz0y]

Maybe NV decided to make a dual GK104 card to compete for the top slot while they work on GK110. That could make all the wild speculation sort of accurate.

</crazy hypothesis>

Not crazy at all. In fact that is exactly what they should do.

 

[seahawk]

Well if it manages to outperform a 7970 with only 60% of the bandwidth color me impressed.... I have a hard time believing that though.

Well, say equal to slightly ahead for anything up to 4xMSAA/16AF 1920x1020, falling behind at 8xMSAA/16AF 2560. Comes at 299$ and power consumption is equal to slightly better. (Say around 170W for typical games) Hard to believe?

 

[snarfbot]

so 4048 shaders at 850mhz, 256 tmu's 64 rops, that would make gk110 about 4x as fast as gf110 on paper.

silly season indeed, so how do the vliw rumors fit into all this, if at all?

 

[AlexV]

so 4048 shaders at 850mhz, 256 tmu's 64 rops, that would make gk110 about 4x as fast as gf110 on paper.

silly season indeed, so how do the vliw rumors fit into all this, if at all?

It's clearly 5-way VLIW with a special unit for transcendentals and other stuff. Oh, wait....

 

[Ailuros]

I would say that the hot-clock is a failed idea, and they should drop it
- it must burn a lot of energy, and cause them masses of problem with layout
- therefore, hopefully, the rumors are true!

And the evidence that it burns a lot more power is where? On GF1xx the hot- to core clock ratio is on a steady 2:1 ratio. Else the 512SPs@1.544GHz of a GTX580 assuming perfect frequency scaling would be equivalent to 1024SPs@772MHz. Cayman has 1536SPs@880MHz about 14% less transistors than GF110, about 26% less die area with a somewhat comparable TDP to the former but burns if memory serves well in real time scenarios up to 20% less depending on case. On top of framebuffer differences it still stands that FLOPs != FLOPs between architectures and of course that FLOPs by themselves are only one part of a GPUs 3D performance.

I don't think NV's real problem so far was the hotclock per se, but rather the fact that AMD's performance chips came along a lot smaller with quite high efficiency.

 

[Kaotik]

Well, say equal to slightly ahead for anything up to 4xMSAA/16AF 1920x1020, falling behind at 8xMSAA/16AF 2560. Comes at 299$ and power consumption is equal to slightly better. (Say around 170W for typical games) Hard to believe?

Yes, it is. Just take a look at past generations, it's extremely rare that one is so much ahead of the other