NVIDIA Kepler speculation thread

I've heard the 192 alu per SP figure from a source who I think actually knows.

 

How do you know?

 

I dont know, just speculating i like the idea what mczak says.. Everything is permitted in order to reach 3xDP FLOPs/W goal lol..

 

Let the leaks start

 

GK107 mobile GDDR5 versions: http://ideapad.zol.com.cn/58/160_573716.html

GTX 660M - 5Gbps
GT 650M - 4Gbps

Core clocks could not be recognized.

 

640M is 625MHz @1.8GHz Effective MC..
650M 4.00GHz
660M 5.00GHz

i guess they halved the ROP count on 640M so it's okay with 1.8Gbps DDR3, say 8ROPs.. and 650M and 660M are 16 ROPs but 650M is clocked 20% lower than 660M.. i guess 650M @775MHz and 660M 925MHz pretty much GTX550Ti level performance..

 

GT 650M is DDR3 or GDDR5.

3DM11 score of GT 650M DDR3: http://nbbbs.zol.com.cn/2/98_12128.html specs: http://nbbbs.zol.com.cn/2/98_12127.html
...around ~15% faster than GT 640M DDR3.

And halved ROPs? This should not be possible, if they coupled like in Fermi.

 

Disable half of block 8 ROPs and attach to 2 MCs.. nvidia did more exotic changes on GF108 -4 ROPs on 128bit..

 

But GF108 was designed in this way and not deactivated afterwards.
Maybe Kepler is more flexible in this area.

 

I liked mczak's idea too and I think that's how it is going to be done - makes room for more CCs "generating" the peak FLOP values and reduces the need for data movement, "earning" some power savings too.

I just thought you found some confirmation of this already.

 

What does that mean for a potential GK106? Doesn't exist?

 

GTX660 for the desktop maybe?

 

I know. Thats why I suggested GK106 as a desktop GTX660.

 

GK106 may replace GF114(670/680M) as 675/685M when it's due.. just like GF104(485M) replaced GF100(480M)..

 

Might actually be real for a change.

Better density than AMD? I hope it is because they fattened up the caches. Though seems unlikely considering ALUs jumped by 4x.

 

The die size of GK104 is correct. I dont know about the transistor count, but if they've got the die size correct odds are the transistor count should be as well.

As discussed, we cant really compare GK104 with Tahiti, the right comparison would be with Pitcairn. And though the gap has shrunk, AMD still leads in density.

 

Where were the transistor and die sizes confirmed? All I saw was that post on Chiphell and that's about it

And all this GK110 take makes my head hurt. So much speculation about this unicorn chip that has few confirmed details

 

GF114: GTX 670M/GTX 675M.

This roadmap says GTX 680M 4GiB in Q2 2012: http://forum.notebookreview.com/gam...600-series-not-just-fermi-37.html#post8379171

So its probably GK104 or GK106 with some memory tricks to bind 4GiB@192-Bit.

It could be the first time Nvidia could beat AMD in the core count: 1536 > 1280.