Fake Semiconductor News at IEDM 2018

[Daniel Nenni]

It was an interesting day at IEMD 2018 in San Francisco. There was a break in the rain which was nice, the sessions were great, but I heard a ridiculous piece of fake news repeated more than once and it really is disheartening. I'm still working on tracking the origin of this turd but to have it repeated at one of the premier semiconductor process forums really was horrible.

To be clear, semiconductor fabs do NOT cost $14B. In the old days disinformation was a strategic ploy used by businesses to accomplish a specific goal. Not a great way to do business but at least you could understand it. Today however fake news is often a "look at me" moment by people who are personally struggling with one thing or another. Fake news really is a cancer on society, my opinion.

If you want to project a price for a brand new 300mm leading edge fab in California it would probably be in the $15-20B range but that is not how fabs are made. From a very trusted source, the cost of the latest 300mm 7/10nm fabs is between $7-10B. TSMC is on the lower end of that scale while Intel is the upper end and Samsung is somewhere in between.

The thing you have to remember is that fabs and fab equipment are upgraded and re-used over the years. For example, TSMC used 20nm fabs for 16nm and 12nm. TSMC 10nm and 7nm are the same fabs. To be clear, TSMC did not pay $7B for a 10nm fab and $7B for 7nm, they paid $7-8B for both. The Intel fabs in Israel I believe started at 40nm and are now being used for 10nm? The Samsung fabs are the same with a significant amount of reuse and Samsung does both leading edge Logic and Memory which is a big advantage in fab building, absolutely.

So please, enough with the fake semiconductor news, a fab does not cost $14B, the end is not near, and yellow snow is not lemon flavored.

 

[Daniel Nenni]

TSMC says latest chip plant will cost around $20 bln | Reuters

FYI: TSMC builds gigafabs so let's not confuse the cost of a 10k wpm fab with one that produces 100k wpm.

 

[George Storm]

This reminder is opportune. However, IMO the usual "fake" bit in the news is not the cost of a new fab but the omission of your crucial word "upgrade" and its associated savings (or total cost).
And, in case you should visit Blighty, some of the yellow snow in my freezer may well be lemon sorbet (I also like banana, pineapple, melon, apricot, capsicum... so check the markings on the containers)

 

[benb]

Dan, if TSMC can build 60K/month of 7nm capacity for $7B, good for them. I think others are faced with costs north of $10B though.

Of course I need to offer a theory why. Partly it's Asia, some costs are lower there. But I suspect TSMC understands their processes better. You see Intel always buy the highest-end equipment and hope for the best, expecting their equipment investment to bail them out. And it didn't work at 10nm, for sure. When you understand a process better, you can buy the equipment you need, not always the highest end. It's a lean strategy based on science; TSMC is $3B or more leaner than others, because they invested in science, which allows them to reduce their investment in equipment.

 

[Staf_Verhaegen]

You see Intel always buy the highest-end equipment and hope for the best, expecting their equipment investment to bail them out. And it didn't work at 10nm, for sure. When you understand a process better, you can buy the equipment you need, not always the highest end. It's a lean strategy based on science; TSMC is $3B or more leaner than others, because they invested in science, which allows them to reduce their investment in equipment.

Implying Intel does not understand their process and TSMC does, is just ridiculous. Intel has always been pushing their processes harder than the rest but also resulting in a more costly process. They can because they are in high margin CPU business and the others need to make trade-offs for processing cost. Problem is that in recent nodes they pushed too hard and could not get the normal yield learning. Of course in hindsight it is easy to say they should have known that.