TSMC's 20nm node appears to be on its way out even before it took off. The planar node simply came too late, too close to TSMC's 16nm FinFET (16FF) node, which will also be updated to 16FF+ in just a few quarters.
We talked to a couple of sources familiar with the segment and heard an interesting take on TSMC's decision to proceed with 20nm anyway. As always, the company burned a lot of money getting the new node ready, despite the fact that it was aware many outfits would skip it and wait for FinFET.
However, TSMC knew it would be worth the effort, since it would get Qualcomm and Apple on board.
Sweetheart 20nm deals with Qualcomm and Apple
While some segments have the luxury of skipping this node, others do not. Nvidia and AMD can afford to ditch 20nm for their GPUs, but companies like Qualcomm, Apple and Samsung cannot.
With that in mind, TSMC knew that it would be worthwhile - Qualcomm needed a new node and so did Apple (for different reasons). Samsung also has a 20nm node and Apple uses both for A8 SoC production. Samsung also uses its node for the latest Exynos series application processors, but Qualcomm would have been in trouble had it not been for TSMC's node.
As far as the rest of the industry goes, we did not hear a lot of kind words about the 20nm node - with descriptions ranging from "pointless" to "broken" or any of a number of dirty words. However, TSMC needed it for Qualcomm - without it, the node made next to no sense at all.
16FF and 16FF+ are not even in the same league
The unfortunate timing and lack of big performance increases rendered the 20nm node relatively unpopular.
TSMC expects about 60 16nm tapeouts by the end of 2015, but we are not sure whether this figure includes 16FF+ designs, hence the number could go up later. On the other hand, the foundry currently has about 30 tapeouts of 20nm designs and chances are it will not get a lot more.
However, the figures may be misleading, as at least two of TSMC's 20nm design wins are huge (Apple A8 and Snapdragon 810). TSMC should be able to score a few more high-volume 20nm design wins, but they will be dwarfed by 16FF/16FF+.
There are also some economic considerations - Apple and Qualcomm may be willing to pay a bigger premium for 20nm than companies like Nvidia and AMD. They simply have more to gain from it and since 20nm capacity is limited for obvious reasons, there is really not a lot of wafers to go around and TSMC has to make its money back. Since 20nm will be a limited, short-lived node, it is reasonable to conclude it carries a hefty premium. Otherwise it would make no sense for TSMC.
Why the industry didn't take to 20nm
So what doomed the 20nm node? A number of factors were involved. The industry has been pushing FinFET for a while, so developing designs for a planar node did not make much sense in a number of niches. Redesigning old 28nm planar parts for 20nm makes sense, but then again the performance gain is just not that big and the SoC lifecycle is too short.
The TSMC 20nm process uses double patterning to increase gate density in planar transistors, thus allowing almost twice the gate density relative to 28nm. The process also offers significant leakage reduction and 25% better power efficiency than 28nm.
However, there are a number of downsides. The very small planar transistors limit speed, making the node less than ideal for high performance chips. A smaller die and a 25% reduction in power consumption make perfect sense for mobile SoCs, but not so much for discrete GPUs and other big cores. 16FF+ should deliver power savings of 50% and a 40% performance gain relative to 20nm.
Major SoC makers like Apple, Qualcomm and Samsung have a lot to gain from transitioning to 20nm in high-end parts, but smaller SoC makers geared toward the mid-range market do not, hence they will wait for 14/16nm FinFET processes to become available.
20nm will give the big boys a slight competitive edge in 2015 and to some extent it's their node, since there's a chance it wouldn't have even materialised had it not been for them. However, the advantage won't last very long, as competitors will focus their efforts on fast FinFET SoC rollouts (we are looking at your MediaTek).