techPowerUp! Forums - View Single Post - GeForce GTX 660 Arrives Mid-August: Report

Benetanegia · Jul 18, 2012, 01:55 PM

http://videocardz.com/33874/geforce-...han-gts-550-ti

I'm starting to believe that the 660 might be based on GK106 after all, but it's not what everyone thought it was. That or GK106 was scrapped altogether which makes very little sense to me.

My little new theory is that GK106 took longer because it was slightly redesigned when GK100 was scrapped and GK104 took the high-end card position, in order to be able to compete in the performance class*. Early rumors speculated about GK106 being a 768 SP card, basically 2 clusters where GK104 has 4, with 2 SMX each. And it was always rumored to be a 192 bit part. Basically folowing the tradition set with fermi cards, fair enough, it made sense.

But now I think that GK106 might use the same clusters as GK110, with 3 SMX each for a total of 6, which would turn it into a 1152 SP, 192 bit chip. Exactly what rumors say for GTX660.

Like I said the alternative is that GK106 won't be released and that makes no sense to me. But with GK107 + GDDR5 taking the GTS650 name, it looks very very unlikely for it to exist if GTX660 is based on GK104 too. Yeah non Ti 660 being GK106, but only 1 SKu based on the chip is unlikely too: the non-ti 660 theory has always been supported by harvested GK106 being the 650.

Well at least August will be interesting I guess.

* You just cannot compete if your mid-range/performace GPU is half the size of your high-end. It's been half a decade since the second chip offers at least 2/3 the performance, often time 3/4. Going straight to 1/2 is suicide. Of course GK104 was THAT chip, just like GF104 and G94 senved that purpose, but since GK104 became the one to go in high-end cards a new such chip is required. IMO a 6 SMX GK106 would be it. It's good balanced, it doen't require a lot of reworking because the 3 SMX designed was already tested for GK100/110 and packing them in just 2 clusters saves a lot of space. Only "sacrifice" is geometry/tesselation compared to GK104, but that's irrelevant for mid-range because of this:

It would still beat a GTX580 on pure tesselation.

Jul 18, 2012, 01:55 PM	#66
Benetanegia Join Date: Sep 2009 Location: Reaching your left retina. Posts: 2,683 (1.99/day) Thanks: 125 Thanked 701 Times in 494 Posts	http://videocardz.com/33874/geforce-...han-gts-550-ti I'm starting to believe that the 660 might be based on GK106 after all, but it's not what everyone thought it was. That or GK106 was scrapped altogether which makes very little sense to me. My little new theory is that GK106 took longer because it was slightly redesigned when GK100 was scrapped and GK104 took the high-end card position, in order to be able to compete in the performance class. Early rumors speculated about GK106 being a 768 SP card, basically 2 clusters where GK104 has 4, with 2 SMX each. And it was always rumored to be a 192 bit part. Basically folowing the tradition set with fermi cards, fair enough, it made sense. But now I think that GK106 might use the same clusters as GK110, with 3 SMX each for a total of 6, which would turn it into a 1152 SP, 192 bit chip. Exactly what rumors say for GTX660. Like I said the alternative is that GK106 won't be released and that makes no sense to me. But with GK107 + GDDR5 taking the GTS650 name, it looks very very unlikely for it to exist if GTX660 is based on GK104 too. Yeah non Ti 660 being GK106, but only 1 SKu based on the chip is unlikely too: the non-ti 660 theory has always been supported by harvested GK106 being the 650. Well at least August will be interesting I guess. You just cannot compete if your mid-range/performace GPU is half the size of your high-end. It's been half a decade since the second chip offers at least 2/3 the performance, often time 3/4. Going straight to 1/2 is suicide. Of course GK104 was THAT chip, just like GF104 and G94 senved that purpose, but since GK104 became the one to go in high-end cards a new such chip is required. IMO a 6 SMX GK106 would be it. It's good balanced, it doen't require a lot of reworking because the 3 SMX designed was already tested for GK100/110 and packing them in just 2 clusters saves a lot of space. Only "sacrifice" is geometry/tesselation compared to GK104, but that's irrelevant for mid-range because of this: It would still beat a GTX580 on pure tesselation. Last edited by Benetanegia; Jul 18, 2012 at 02:01 PM.