SEMI describes themselves as “the global industry association serving the manufacturing supply chain for the micro- and nano-electronics industries.” That is a pretty broad remit. One of the things that they do as a neutral party is produce the World Fab Forecast. This is actually a bottom-up database that tracks fabs as they are built, equipped, ramped to volume, upgraded and expanded and, eventually, closed. Of course this is of great interest to people selling the equipment and material required, but also it impacts the entire semiconductor ecosystem. This is especially true in the fabless/foundry ecosystem. If TSMC builds a 16FF+ Gigafab then Apple, Qualcomm, nVidia, Xilinx and lots of others are affected, not to mention Samsung, Global Foundries, UMC and Intel who compete with them.
Christian Dieseldorff maintains the database and also presented at the latest SEMI Silicon Valley SEMI Breakfast Meeting. The forecast covers 3 years of detailed information with a future forecast of 1½ years. To give an idea of how extensive the database is, it covers 1174 fabs, what are officially called “front-end facilities” to distinguish them from test, package and assembly and are nothing to do with FEOL or what we call front-end in EDA. They are owned by 500 companies. There are 58 future facilities starting high-volume manufacturing (HVM) in 2015 or later, with 218 facilities currently either in construction or equipping.
Another aspect to be tracked is consolidation. Recent events (many of which we have covered here at Semiwiki) affect 46 fabs from 4” to 12”. Here they are:
- Fujitsu winds down chip production
- Global Foundries “buys” IBM’s semiconductor business
- Grace and Hua Hong merge
- Cypress and Spansion merge
- Infineon buys International Rectifier
- Triquint and RFMD merge (Qorvo)
- Freescale and NXP (will) merge
Where are all the big fabs? In 2015 there are 1999 fabs equipping with 16 of them investing over $800M including Samsung, SK Hynix, Sandisk/Toshiba, TSMC, Intel, GloFo and SMIC.
Fab equipment has been a rollercoaster. But actually as a rollercoaster it would be a bit boring since it is too predictable, two years up, followed by two years down. This has gone on every 4 years since 1999. But it looks like this year will be the second up year and, for the first time in living memory, 2016 should be an up year too.
One area where historically there has been overinvestment in fabs has been DRAM. In the last few years this has moderated resulting in DRAM prices firming (which is a large part of the recent growth in the overall semiconductor market). In fact there were 40 dedicated DRAM fabs in 2007 and now we are down to 15. But the capacity of new fabs is huge compared to the old, with the 5 new fabs this year adding over 30K wpm with the 10 closing dropping about the same amount.
One area with a big decline is not the number of fabs or capacity but the number of companies building fabs. In the 2004-7 era there were nearly 50 companies building fabs but in the 2014-16 era that is down to 13. Obviously this is largely due to fables (and fab-lite) but also consolidation and companies going out of business too.
One new phenomenon is the IoT market. As I have said before, this is primarily a market for old processes, and old processes typically run in old fabs which are mostly 8”. In fact in 2007 there were 200 8” fabs with a capacity of 5.7M wpm, which dropped to 183 by 2013 with a capacity drop of 11% to 5.1 wpm. However, capacity will increase 4% by 2018 to 5.4M wpm although the number of fabs will again drop to 180.
Bottom line is that fab expansion rates slowed to below 2% in 2012 and 2013. They should increase around 3% year on year from 2015 to 2018. This reflects growth of the overall semiconductor market of 9% in 2014, forecasts for 3-8% in 2015 and 3-9% in 2016.