From Charlie Demerjian at SemiAccurate
["The only problem we have with the roadmap as it stands is that 14XM really seems like a cheat, and that image is unlikely to be shaken until the chips that use it come out. "]
Global Foundries laid out their roadmap for the next few years at Common Platform, and this time it went to the 10nm node. Lets take a look at it in greater detail, there are a few non-obvious points to talk about.
In their morning Common Platform (CP) keynote, Global Foundries Mike Noonen, Executive VP, Global Sales, Marketing, Quality, and Design did something unusual, he admitted the company botched something. More interestingly, he did it publicly and directly when they said their 32nm HKMG process was not exactly what anyone wanted it to be when it was released. This basically says that they have fixed the internal processes that lead to the little ‘whoopsie’ with Llano and are confident that those dark days are behind them. We shall see, but we hope they are right, the foundry business needs more competition.
The roadmap laid out in the keynote out had a few new additions since last year, mainly the 20LPM, 14XM, and 10XM processes. Of these, two are what they sound like, one is not. In case you are wondering, XM stands for Extreme Mobility, XM just sounds better to the ‘Mountain Dew’ generation than EM which is probably why it was chosen.
The roadmap in coarse detail
All of the 32nm, 28nm, and 20nm processes have been announced previously, although a few suffixes have been added. The 20nm process is about what you expected, the next step in planar transistors shrunk to 20nm. There are no FinFETs like Intel is using at 22nm, but GloFo and the CP partners are moving to gate last this time around. This entire process is pretty evolutionary.
That changes quite a bit on the 14XM process, it is most definitely not evolutionary on the transistor front but the rest is. Because of the way the last one was announced, we didn’t bother to write it up at the time but it is worth looking at. Unlike what the name implies, 14XM is not really a full shrink of the 20LPM node to 14nm, it is only half way there. 14XM uses the middle and back end of line (MEOL and BEOL) from the 20nm node couples to the new 14nm FinFETs.
Global Foundries says this effective reuse of the entirety of the non-transistor bits of 20LPM will pull in 14XM, lower cost presumably through the reuse of existing equipment, and lower customer design costs because most of the software for 20nm just carries over. While we have no doubt this is true, it seems like a bit of cheating to do it this way. Since we are not nearly versed enough in the minutia of these 20nm and 14nm processes to say what the exact technical trade-offs are, we will wait until more information comes out to say for sure.
Depending on your views about what GloFo and the CP partners are doing, you could consider 14Xm to be a full shrink of 20LPM or a minor technical half-step. Whatever you classify 14XM to be, the 10XM process is really a full shrink. Starting with test runs in late 2015, 10XM will be a second generation FinFET process with a true 10nm MEOL and BEOL. All that 10XM carries over from 14XM is the number 1 and the letters XM.
Overall, GloFo did a good job in their keynote. They essentially came clean on the early 32/28nm problems that were nagging their reputation and laid out a clear path forward. The pulling in of 14nm may leave some people nervous, but given how much of the process is carried over from 20nm, as long as 20nm is OK, 14′s pull in should not be a big worry. The only problem we have with the roadmap as it stands is that 14XM really seems like a cheat, and that image is unlikely to be shaken until the chips that use it come out. At that time, the results will speak for themselves, until then, whispers will rule.
Sentiment: Strong Buy
The race is for LP. XM makes sense as its an incremental step and it gets them in the temporal ballpark of Intel's accelerated 14nm LP rollout at the end of 2014. At the end of the day, what matters is how much power the transistors are drawing and what speed you can switch them at. If they can get half-way in bridging the performance gap between 20 nm planar and 14nm FinFET, that should be good enough to make them competitive. It's a half step in the right direction and a good parry to Intel's acceleration. If they can pull this off which seems likely, they'll be competitive. The big advantage of the foundries is they don't need the huge margins on processor ASPs that Intel needs to maintain status quo.
The race is for LP. XM makes sense as its an incremental step and it gets them in the temporal ballpark of Intel's accelerated 14nm LP rollout at the end of 2014.
Gloflo just announced FDSOI - planar
they don't seem to know which side is up and which side is down.
The big advantage of the foundries is they don't need the huge margins on processor ASPs that Intel needs to maintain status quo.
Probably one of the most stupid posts -
foundries are OK with low margins but Intel not
"The race is for LP. XM makes sense as its an incremental step and it gets them in the temporal ballpark of Intel's accelerated 14nm LP rollout at the end of 2014."
[What kind of bs is "temporal ballpark"? It doesn't get them in the ball park at all. It will produce a poor economics half-step that is too little, too late. A half-step two years behind Intel and then only if things go perfectly. How did things go for TSMC with 28nm? Would you say perfectly? LOL Don't expect 20nm or the half-step to 14nm to go perfectly. What evidence shows at this point is that each ARM node jump ramps slower and slower. Intel is already ramping FinFet like a champ. The ARM foundries have no chance of doing the same. You need to add a couple of years to all of the ARM camp projections... ]
"At the end of the day, what matters is how much power the transistors are drawing and what speed you can switch them at. If they can get half-way in bridging the performance gap between 20 nm planar and 14nm FinFET, that should be good enough to make them competitive."
[Yeah, if it wasn't going to be two years late. ARM if they are exceedingly lucky will move a half-step to 14nm while Intel is moving a full step to 10nm. The gap is going to get bigger, not smaller. It's just basic math. ARM is moving slower and in half-steps. The fabrication fade is in full swing...]
"It's a half step in the right direction and a good parry to Intel's acceleration. If they can pull this off which seems likely, they'll be competitive."
[TSMCs experience at 28nm says it's unlikely. Glofo's experience at 32/28nm says it's unlikely. What are you basing likely on other than wishful thinking? The ARM foundry track record disputes your conjecture...]
"The big advantage of the foundries is they don't need the huge margins on processor ASPs that Intel needs to maintain status quo."
[No, the ARM foundries don't HAVE the huge margins that Intel has and this makes them less competitive. Intel gets “paid twice” for its products, since it both designs and builds. If you have less margin then you have less to give away. You can't be as flexible on pricing. There is no advantage for the foundries here - the advantage is all Intel's...]
Sentiment: Strong Buy
GloFo admits to problems at 32/28nm but says they have fixed all of them. And now based on having fixed everything they will jump directly to 14nm production.
Hmmmm. Thinking. Thinking.
Well, shouldn't that fixing have resulted in volume production at 28nm? It hasn't.
And don't they have to do 20nm planar before doing 14nm anything since they are copying most of it in order to think that 14nm might actually work? They haven't.
So, if you are willing to accept not one but three leaps of faith, then you might conclude that GloFo will have the half-step to 14nm in the next couple of years.Based on their track record I'm not willing to accept any leaps of faith. I'm gonna be from Missouri on all three suppositions. Show me.
In general we are being told that the ARM foundries are going to develop 20nm planar and the half-step to 14nm at the same time. Except the 14nm has to be based on successful completion of the 20nm node jump since it will use the middle and back end of line (MEOL and BEOL) from the 20nm node. So, um - exactly how do you develop these at the same time? Does this make sense to anyone because it really sounds like they are making this up as they go along.
Given the amount of smoke the ARM foundries are blowing, I'm surprised the whole world doesn't look like Beijing. The ARM Bubble has begun its descent into the land of sharp objects. Even the 14nm half-step looks like a very sharp object. Intel's full-step 14nm FinFET looks like concertina...
Sentiment: Strong Buy