AMD recently presented at the Credit Suisse Technology Conference and though they did not reveal anything specific or concrete, some very interesting hints and implications were dropped by them. I will go over the interesting portions in this article and maybe one more. The first thing, that AMD has done is imply heavily that the upcoming Fiji GPU will be on the 28nm Node and then mentioned that they will switch directly to FinFETs from there.
AMD will be transitioning directly to FinFETs from 28nm - Pirate Islands are probably on 28nm
Most of the transcript (link given at the end) featured your usual management rhetoric, but gems such as these can be found hidden inside the jargon. The question of the 20nm process has been on the lips of many ever since 20nm first entered volume production some months ago. However, after months passed and still no mention of 20nm GPUs ever came to light, we began to wonder just what was going on. Recently, our sentiment is mostly tilted towards the fact that the Industry is making up for lost ground by abandoning the 20nm process and moving directly to 16nm FinFETs instead. Before we go any further here is the relevant quote from AMD:
....from an overall standpoint, today the bulk of our products are 28 nanometers. We will have certain products in 20 nanometers and then we'll go to FinFET from there, but our partnership with GlobalFoundries is where it comes into play in terms of what's the right point to go and intersect products with technology and introduce the parts out in the market.
Now Mr. Devinder Kumar mentions that the majority of our [AMD's] offerings are on 28nm and that this majority will transition directly to FinFET. I don't know about you but AMD's GPU lineup does make up a very significant portion of its offerings. If you put aside GPUs, what remains are old CPUs, the APUs in the consoles (which only counts as two offerings really) and APUs in general. Unless AMD plans to rebrand its old GPUs into the R9 300 Series with only the one flagship on 20nm (which is unlikely), AMD is hinting at Pirate Islands on the 28nm Node. We have already run the numbers based on initial leaks and found that the performance jump shown is very much possible on 28nm and within the 600mm^2 die limit of TSMC.
Ofcourse, the above is mostly just educated speculation on my part and it could be that Mr. Devinder was talking about the Pirate Island GPUs themselves. However, in my mind the chances of Carrizo APU being on the 20nm process are exponentially higher than Pirate Islands on the same. 28nm to 16nm FF+ makes perfect sense and you might recall that Nvidia was also recently listed at TSMC's partners for 16nm (although there was no sign of AMD). I have the mother of all hints from AMD in that regard, but I will cover it in the next writeup which should be up in a few minutes as well.
A full transcript of the proceedings can be found at SeekingAlpha.com