UniversalTruth
Veteran
If available resources are utilised in this way or similar. Or even if all threads are loaded to 100% but do little helpful work, then...
AMD: Pirate Islands (R* 3** series) Speculation/Rumor Thread
- Thread starter iMacmatician
- Start date
-
- Tags
- amd
If available resources are utilised in this way or similar. Or even if all threads are loaded to 100% but do little helpful work, then...
iMacmatician
Regular
AMD's Q1 2014 earnings call transcript points towards no chips in 20 nm this year.
Could we see something like the low-end and midrange Pirate Islands chips in 2014 on 28 nm for lower cost (especially whatever succeeds Pitcairn/Curacao since that doesn't have TrueAudio) and the high-end Pirate Island chip in 2015 on 20 nm, or would it be better to go all 20 nm in 2015?
This is an article from VZ, so offcourse, i take it with all the grain of salt i can ( they mainly copy article from chineses forum anyway ) .. But if this report of the AMD sessions is true, maybe, i think, Ryan ( Anand ) should have more information... ? ( he i try )
Now the public AMD sessions as Nvidia for earning call have been allready flooded with bad informnations for strategic marketing operation ... anyway, your toughts ?
Now the public AMD sessions as Nvidia for earning call have been allready flooded with bad informnations for strategic marketing operation ... anyway, your toughts ?
Last edited by a moderator:
GF + Samsung are 14nm, not 16nm.
That's seriously disappointing, my 7950 is going to around longer than i was expecting.
Time to invest that GPU money into a 4690K.
Also;
Sooner than? TSMC?
Sooner than GloFo could without Samsung's help. That said, I don't read Lisa Su's statement as necessarily implying that there will be no 20nm GPUs this year. I also believe that HBM will be released soonish, which may not be as good as a process shrink, but should still bring a nice boost.
GF and Samsung do both, 14 and 16nm Finfet ...... but 14nm is absolutely not ready for complex chips like GPU or CPU anyway, at best for 2015 - 2016 for SOC ( not complex gpu or CPU architectures ) . Even Samsung is not using his 14nm process for their Nand right now ( 18nm ) .. Glofo today is really near to be a part of Samsung fundries ( TI, IBM, etc, nearly all funders are now part of Samsung, it seems the collaborations and associations of the big fundry, lead to a domination of Samsung in this sector ...
Alexko, i have got the main feelings, it look like Lisa Su, is not really speaking about the same thing described in the article.. ( its why i was speak about strategy and marketing, more than reality )
Last edited by a moderator:
How did you come to that conclusion?
If AMD claims that the rest of 2014 will still be 28nm, then that was a decision made at least a year ago. There's no way they're sitting on a ready for production 20nm design.
I was getting that vibe too, her comment is kind of evasive.
She answered the question without actually answering the question.
ofc they are sitting on it, TSMC is right now unable to product there work in 20nm.. but for AMD as an example, they have tap out the first core, nearly 1 year ago.... but it seems impossible to push all of that with TSMC in massproduction.. at least this is the rumor we heard, but i still have some feeling there's still a lot of wrong thing in this...
To my understanding the "14nm FinFET", no matter if it's GF, Samsung or TSMC, is combination of 16/14nm
By claiming that AMD had a 20nm design ready way before TSMC ever promised to be ready, you're essentially saying that AMD are idiots for making something that can't be produced.
You should stop reading Czech websites.
UniversalTruth
Veteran
I also find it highly improbable that 20nm chips could be ready sitting somewhere and waiting for TSMC, more likely AMD should still be working on the designs for optimisations, etc.
Very sad that in 2014 we won't see any major breakthrough.
Very sad that in 2014 we won't see any major breakthrough.
IMHO both IHVs are in the same boat right now for 20SoC. No idea about AMD's Bermuda but I'd like to stand corrected that GM200 didn't have its tape out yet.
I kinda hope that both come out with a large 28nm design, just to have a process neutral comparison of the progression of the state-of-the-art in design in 3 years. The GTX750Ti was already very nice, but a large silicon version would be even better.
IOW: I hope Erinyes is at least partially correct.
And then they should shrink the identical design asap, to have a design neutral comparison of the progression of process.
It's more a combination of 14-16nm FinFET and a 20nm back end.
However, the overriding lesson of the last 2-3 nodes is that the marketing department is a more significant factor in the number chosen than anything about the process.
Characterizing the feature length of a given process has been a frequent source of potshots between fabs, each basically saying the others are fudging numbers to look cooler.
I can't recall which 14 or 16nm process maker basically admitted they picked a smaller number so that everyone else would know that, despite the fact that it's becoming impossible to derive any real understanding of transistor density based on the node number, how much more advanced they were. Too many other things, like the design, the process target, and other sorts of foundry or IDM marketing BS wreck the math.
That's not to say it doesn't matter at all. Getting to FinFETs does become seriously needed around the 20nm range (like HKMG at ~40nm), and the lack of enthusiasm over planar 20nm is partly due to that.
Are you sure that's true for the alliance? I know 16nm from tsmc is more like 20nm + finfet (it comes out roughly a year after 20nm), but I had the impression that 14nm was a bit more than that. For instance, it seems like it's part of IBM's normal cadence (e.g. 32nm->22nm->14nm) and not just 22nm + finfet.
I don't know about IBM (their constraints are pretty different from those of other foundries) but I'm fairly sure Samsung/GloFo's 14nm process is 14nm transistors with 20nm BEOL, just like TSMC.
For time to market and ease of transition, both Samsung and Globalfoundries (both then and now) were applying 20nm metal layers to a finer transistor geometry.
IBM's path, such as it is, is a gate-first SOI planar at 22nm. This is about has far as one can go in the opposite direction as the rest of the former alliance.
The next node is IBM giving up its microelectronics division, potentially to Globalfoundries--who has just recently given up on its FinFET node transition and bought Samsung's.
(edit: Well, not giving it so much as allegedly negotiating at a price range less than it would like, with not that many bidders. )
(edit edit: The fabs, to be accurate.)