AMD: R9xx Speculation

[Kaotik]

The first two wafers had too big dies to be pipe cleaning parts. Dunn about the third one. BTW, charlie is on record saying NI is @TSMC. May be they'll manufacture it at both TSMC and GF to reduce migration risk.

Or just high end NI @ TSMC, midrange/lowend @ GF, or vice versa

 

[rpg.314]

The front 2 wafers have >300mm2 dies.

 

[Silent_Buddha]

Or just high end NI @ TSMC, midrange/lowend @ GF, or vice versa

If I were to bet, the best low risk proposition while also giving it a good test for eventual high end use would be using something similar to Rv740.

You wouldn't want to run into significant problems with either your high volume budget parts nor your large die performance/enthusiast parts. So taking a middle part that straddles performance boundaries is a safer bet. One with a relatively larger die than a budget part (so problems related to larger dies can be worked out), while also not requiring high volumes like a budget part. And finally not adversely affecting or delaying your low volume "Halo" generating parts.

Regards,
SB

 

[rpg.314]

After the 40 nm fiasco, both IHV's will be hesitant to dip in 28 nm without sacrificial chips. But if NI has to come in 2010, there simply isn't enough time to do pipe-cleaning. I guess Juniper's replacement as the first chip seems a better bet.

 

[eastmen]

But if you have to do a redesign for a full process drop, why not add/improve stuff at the same time? Going 40nm -> 28nm was always going to be more than just a shrink. I doubt AMD will stand still in the face of continued innovation from Nvidia and Intel. AMD know you can't get away by just shrinking the same old design and hope that no one else comes up with something new.

Amd who needs money would rather sell a 200mm2 chip at $600 , $400 , $300 than a 330mm2 chip at those prices and would certianly have a huge advantage against a 500mm+ chip

I'm thinking AMD will want to press the process advantage to make cash. Esp if the gf100s are on par with the 40nm verisons of cypress

 

[Squilliam]

If I were to bet, the best low risk proposition while also giving it a good test for eventual high end use would be using something similar to Rv740.

You wouldn't want to run into significant problems with either your high volume budget parts nor your large die performance/enthusiast parts. So taking a middle part that straddles performance boundaries is a safer bet. One with a relatively larger die than a budget part (so problems related to larger dies can be worked out), while also not requiring high volumes like a budget part. And finally not adversely affecting or delaying your low volume "Halo" generating parts.

Regards,
SB

Something similar to the RV740 would be Juniper. They could effectively replace the 56xx and the 57xx parts and in addition to this it would be a killer part for laptops. They already have a plentiful supply from TSMC if they are slow at getting volume through the door and as they are paid for performance/watt not per mm^2 they could get some pretty large margins on the part so early volume need not concern them too much either.

 

[tannat]

If they are this aggressive going towards 28 nm I still believe it would be smart to do one chip at TCSM and one at GF. It would ad flexibility if one of the foundries mess up the node and since the foundries try out one line, it takes time to ramp up. They would not be that bottlenecked at NI launch.

 

[Bouncing Zabaglione Bros.]

If they are this aggressive going towards 28 nm I still believe it would be smart to do one chip at TCSM and one at GF. It would ad flexibility if one of the foundries mess up the node and since the foundries try out one line, it takes time to ramp up. They would not be that bottlenecked at NI launch.

Then AMD would have to design two chips, as foundries use different libraries, even for the same process nodes. You can't design a chip for processes at GF and then get it made at TSMC (or vice versa).

 

[tannat]

Then AMD would have to design two chips, as foundries use different libraries, even for the same process nodes. You can't design a chip for processes at GF and then get it made at TSMC (or vice versa).

I'm aware of that. However early adaption to next node is something ATi has excelled in earlier and why not use their strengths to advantage? The pain of doing it for two processes is large but the advantages sounds huge to me. Safety net, higher volumes at early stage of the node. They already had different teams on juniper and cypress in order to be able to launch them back to back. Their main problem were getting out the chips and that they were so in the hands of the TSMC node maturity (or lack of) .

 

[Kaotik]

Then AMD would have to design two chips, as foundries use different libraries, even for the same process nodes. You can't design a chip for processes at GF and then get it made at TSMC (or vice versa).

Even though the current chips are similar, there's some differences between them anyway, only the highest end has DP capability, so it's not perhaps too farfetched that they would design the highest end for one library, and the rest for the other.
Or then there's of course the distant possibility that they have 2 "Northern Islands" or 2 "R9xx"s, one being Evergreen, just a more of it, and designed for TSMC 28nm libraries, and then "completely new" architecture (think of R3xx > R6xx switch) designed for GF 28nm libraries, and they'd go for which ever works better.
Or even then a mix of both, the new architecture would be used for either high or lower end chips, and designed for one library, and the rest of the lineup would be "more of the same" and designed for the other

 

[jimbo75]

http://www.semiaccurate.com/forums/showpost.php?p=33815&postcount=13

It's the middle wafer apparently. It also looks similar to the one Charlie wrote an article about 2 months ago. http://www.semiaccurate.com/2010/01/09/global-foundries-28nm-wafer-spotted/

 

[3dilettante]

I thought the wafer in the earlier story was later confirmed to be test structures after all.

 

[jimbo75]

I'm not sure, I just put 2+2 together and it seems to make sense. Both pics taken at CES, the wafers look similar anyway, both 28nm, both ~300mm2. If that's an ATI chip at 28nm it must be an absolute monster though, light years ahead of anything around right now.

 

[Silent_Buddha]

If it actually is the second one, I'll eat a hat. Those dies look absolutely HUGE.

Regards,
SB

 

[jimbo75]

If it actually is the second one, I'll eat a hat. Those dies look absolutely HUGE.

Regards,
SB

I agree. What else could it be though?

 

[Silent_Buddha]

I was originally thinking it was possibly the 3rd wafer in the shot.

Regards,
SB

 

[jimbo75]

http://www.brightsideofnews.com/news/2010/1/13/globalfoundries-knows-how-to-party!.aspx

Theo says they are all 28nm, Daniel Nenni says the middle one is AMD/ATI.

It's a total mystery to me now, anybody got an opinion on what we're looking at here?

 

[HVAC]

http://www.brightsideofnews.com/news/2010/1/13/globalfoundries-knows-how-to-party!.aspx

Theo says they are all 28nm, Daniel Nenni says the middle one is AMD/ATI.

It's a total mystery to me now, anybody got an opinion on what we're looking at here?

how about fusion ???

 

[PSU-failure]

First one probably is an hex Stars CPU, but its size and proportions are strange.

Perhaps a 32nm shrink of Thuban as it's ~300mm2.

Others are too small to see details.

 

[rpg.314]

Amd who needs money would rather sell a 200mm2 chip at $600 , $400 , $300 than a 330mm2 chip at those prices and would certianly have a huge advantage against a 500mm+ chip

I'm thinking AMD will want to press the process advantage to make cash. Esp if the gf100s are on par with the 40nm verisons of cypress

After ~2 yrs of ass kicking, nv won't let AMD have time-to-new-node advantage. Fermi2 will be on 28nm, it is most certainly slotted for 4q10 (ofc, delays are another matter) and it will be >550mm2.