AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread
UniversalTruth
Veteran
Of course, the hype was built to explosive levels and now when people are returned back to reality, everybody will be disappointed...
I hope they stress on the best performance per dollar ratio more than everything else
pjbliverpool said:
Well, it would only have 160 TMUs, so it would still be behind the 780 and Titan in texture fill.
I would guess it will, on average, be slightly faster than a 780, and if it retails for $549 or $499 that is a relatively great deal. If it retails for for $599 or $649, it will just be an AMD flavored 780 and not terribly exciting.
~80% the size of GK110.
I'm guessing around 440mm2.
Supposedly has a rectangular die which I believe someone here, B3D, mentioned that is done to increase perimeter for the pads/interface.
If curacao is indeed faster than pitcain (making it very much a direct gk104 competitor) it seems quite strange to keep tahiti in production. The curacao part of those rumours doesn't seem that solid though..
"nearly 30%" is probably slightly more than 25% and higher area-efficiency would probably bring the xt quite close to titan (but titan isn't the full gk110, so bring on the ultra..)
"nearly 30%" is probably slightly more than 25%
and higher area-efficiency would probably bring the xt quite close to titan (but titan isn't the full gk110, so bring on the ultra..)If curacao is indeed faster than pitcain (making it very much a direct gk104 competitor) it seems quite strange to keep tahiti in production. The curacao part of those rumours doesn't seem that solid though..
"nearly 30%" is probably slightly more than 25%
So ~430-440mm2 with somewhere around 5.5-6b transistors, I'm guessing around 5.8b, compared to the 4.31b in Tahiti.
iMacmatician
Regular
At least it wasn't built up as high overall as the Cayman hype.Of course, the hype was built to explosive levels and now when people are returned back to reality, everybody will be disappointed...
I still blame the Cayman hype on the Ibiza performance targets getting out.
The most intriguing part was the following quote: “Another thing I can tell you is about the process node: this GPU is in 28nm. Some have speculated that it was 20nm and it’s not for a specific reason: At 28nm for an enthusiast GPU, we can achieve higher clock speeds and higher absolute performance.”
I'm not sure what this means and I wish the interviewer had followed up on it. If he's just saying: "For a given design, 28nm will clock higher" then OK, fine.
But on 20nm you could make the GPU wider. So is it a yields/process readiness issue? Is 20nm too expensive? Or is 20nm really so crappy that even with a wider GPU, you couldn't clock it high enough to match a 28nm design?
Dave, are you at liberty to elaborate?
I'm not sure what this means and I wish the interviewer had followed up on it. If he's just saying: "For a given design, 28nm will clock higher" then OK, fine.
But on 20nm you could make the GPU wider. So is it a yields/process readiness issue? Is 20nm too expensive? Or is 20nm really so crappy that even with a wider GPU, you couldn't clock it high enough to match a 28nm design?
Dave, are you at liberty to elaborate?
UniversalTruth
Veteran
I guess all of the above
20 nm is simply too early to be shown for mass production...
20 nm is simply too early to be shown for mass production...
UniversalTruth
Veteran
Here another guess....
I guess it is the same story as in Intel where the management simply doesn't give green light for more serious die size construction... As simple (and sad, and perhaps wrong as well) as this...
I guess it is the same story as in Intel where the management simply doesn't give green light for more serious die size construction... As simple (and sad, and perhaps wrong as well) as this...
They couldn't, really, since $/transistor on 20nm is the same or even higher than on 28nm... If 28nm clocks higher and is as cheap/cheaper then it is a better solution.
Also, will be easier to cool, since a larger chip at a particular performance level puts out less W/mm2 than a smaller.
How? If it's 440 and tahiti is 365..
But Tahiti is very conservative - if we apply the same transitor density (/redundency) as Pitcain (which is also a pretty early 28nm chip, so should be possible to keep close to that ratio on a bigger chip today) it would only be 325.
Keeping the same IO intact should leave a LOT of additional space then - looks like gddr5/MiscIO takes up a 3rd of the die.. Meaning we could upscale the rest by 50% (simplified ofcourse, but..).
(sorry, can't find fellix' original post of the above tahiti shot here)
Last edited by a moderator:
I'm quite sure the fact that they're not doing them doesn't mean they wouldn't know how to make one.
--
Only reason that it sounds like it could be slower than Titan (which is faster than the 780 btw) is that he says "best performance per mm^2" rather than "best performance"
NVIDIA did say that, albeit a while ago, but that's not the argument M. Skynner put forward; nor did he say that 20nm wasn't ready.
I just find it peculiar that he should address the node issue without being prompted to do so, and in such specific, yet odd terms.
DuckThor Evil
Legend
Perhaps they would have had to been more conservative with the clocks on that new "unknown" node, instead of with the well known old node. Kind of like they did with the 7970 vs Pitcairn and Cape Verde. And perhaps he wasn't really comparing the nodes itself, but only this particular chip design on those two nodes at the release window timeframe. Who knows, but I wish they can put some pressure to nVidia anyway.
I dont think textures filling is so much important nowadays.. i mean ofc there's a difference between got 80 and 160 textures Units ( providing they are nearly working at same level ).. not so much between 160 and 180TMU.. ( outside some specific benchs ). ROPS have substantially more impact in the end. For the GK110, this TMU number is more a consequences of the number of SMX. For a "compute oriented card", TMU is not really what you think at first. ( they have 16 TMU / SMX, ofc they could have put only 8/SMX but this would have divide them by 2 on the final count ).
Anyway: " They’re coming in Q4." .. so i think we have now a confirmation ( on a 3 months largeness )..
As for the 28 vs 20nm.. let say, today with 28nm, they know how far they can go on clock speed, for sure it is not the case with the 20nm.. and this still let them do a die shrink at anytimes then in current 2014.
Last edited by a moderator:
However, they are comparing it to the GK110 die rather than the Geforce Titan in particular.
The reference Geforce Titan is clocked somewhat conservatively, and can be overclocked a fair amount when adequately cooled.
If AMD can achieve similar performance/die size scaling as Pitcairn with Hawaii it is certainly achievable to beat a current reference Titan, but they may be reluctant to generate hype when Nvidia could deflate it with an upclock to Titan.
DuckThor Evil
Legend
I don't believe for a second that the scaling will be as good as what a die size jump from Pitcairn to ~430mm2 would suggest, but hopefully the performance will at least match or be little better than what Tahiti has per mm2.