20 Nanometer Node Remains Broken for GPUs – 2015 To See Mostly 28nm Products
A particularly interesting editorial over at PC Perspective was recently published - one I don't feel the slightest regret, echoing here. The write up, simply put, is basically an aggregation of process updates so far and compounds on the in-feasibility of the 20nm process as a viable base for the Fiji XT GPU. AMD rolled out Tonga on the 28nm node last year, so the probability of Fiji XT being on the 28nm is pretty significant. We already know Fiji will feature HBM memory on board via an interposer, but just how a process shrink is going to fit in (if at all) with the grand scheme of things remains to be seen.
20nm GPUs remain just as infeasible as ever - 2015 to be the domain of 28 nanometer
The prime argument against the 20nm process is one our readers should be pretty familiar with now (seeing its extensively used in our comments section) - that of Silicon Edge's wafer calculator ( a neat little program that calculates the amount of usable dies on a round wafer given a specified die size and process node.) The editorial pointed out that at the size that Fiji is purported to be, AMD will be seeing approximately 98 usable dies if the size turns out to be 550mm^2 (on a 12 inch wafer). This number will increase significantly if the die size is lower and yield will decrease the number if the process is lower.
There have been multiple rumors of AMD expanding its wafer contracts and of TSMC loosing a significant portion of its 28nm customers. At this point, both 20 and 28nm process are more or less equally probable (in my opinion) where the Fiji XT GPU is concerned. This is not because of existing evidence, but because of company claims themselves. However, hard evidence is very hard to argue against, and this is why PCPer's ed makes such an impact. It notes transistors per square mm and how AMD has been able to push up the amount of transistors on a die of a given size by alot.
To be honest, the evidence is very much stacked against AMD bringing out a 20nm GPU - everything from heat dissipation to clocks becomes a problem. As the article rightfully points out, Intel the absolute giant of the silicon Industry stumbled at the sub 20nm node - so physics is about to become a real problem for foundries very fast. It also means that speaking in terms of sheer practicality, 28nm HKMG would be perfect for a high end, water cooled GPU. Ofcourse, that would mean AMD would need to work on its architectures solely, or go with the 20nm process and let the consumers suffer lower than normal clock rates.