Megafabs: Why they build them?

The Chip Insider™

January 17, 2008 – From the Front Lines:

Megafabs: Why they build them? One of the most interesting questions to answer is why are Megafabs so popular? Megafabs, or fabs with 80K or more WSM* have been the rage in Asia since 300mm. In general, Asia has led the trend to larger fabs since Japan led the market in the 80’s. The reason used to be simple: economy of scale. Moving from 5K WSM to 10K lowers Cost Per Wafer (CPW) by roughly 25%. Going from 10K to 20K nets about a 10% savings, which goes directly to your bottom line.  20K to 40K WSM nets about 4-5% lower cost. But going to 80K WSM or more nets less than a 1% gain, which is statistically insignificant, given the accuracy of capacity planning models.

So, simple economies of scale do not explain the difference. The actual reasons are mostly managerial:

• It’s easier to tap unused capacity in a mega-fab:
  • Most fabs begin to lose efficiency above 93% utilization
    • At 100K WSM that’s 7K WSM to play with
  • Tool downs are less disruptive
  • Hot lots are less disruptive
  • Capacity can be added incrementally in empty cleanroom space thanks to the introduction of mini-environments

• You build fewer fabs
  • This allows one to meet delivery demand faster without the disruptive factor of building a new fab
  • If you are a foundry, where wafer prices can fall as low as $1K, 100K WSM is only equivalent to $300M of additional revenue per month when prices are so low
    • But this is more a reason why foundries prefer 200mm fabs, because depreciation alone on a 300mm fab runs $1250/wafer
  • You go to the bankers less often for capital infusion

• There are labor economies of scale:
  • Fewer fabs means fewer fab managers
    • Saves pay and maintains hierarchies
  • Easier to manage at the corporate level
    • Creates fewer political issues internally
    • Does not threaten established hierarchies

• There are information economies of scale:
  • Yield problems and solutions come faster
    • More engineering resources can be brought to bear
  • Yield learning is easier to propagate.
    • Solutions execute without fab-to-fab NIH* issues
    • Saves on expensive IT resources
    • Allows continued reliance on human information transfer, lowering transition risk

• There are competitive economies of scale:
  • In memories, price wars are very common
  • Being able to stuff the market with lots of parts
    • guarantees you get earlier prices, which are higher
    • forces losses at competitors because they get later prices, which are lower
  • Prices can fall at a 5% rate per week
    • A 1 week production ramp advantage equates to a 20.4% margin advantage per quarter — that’s enough to bury your competition

Some of these do have economic benefits, but they would be very hard to measure in contrast to a world-class Western IDM. One reason is that these companies’ products are far different. The differences have emerged more for cultural reasons, in that decisions were made that set one versus the other down different information automation paths. In the case of Asia, the scale of information automation pales in contrast to the best western IDMs. They rely heavily on sneaker-net. In the U.S. this failed miserably as T.J. Rodgers proved, who coined the term.

At the foundries, yields are notoriously low. For example, GPUs average 70%, which wouldn’t allow you to pass into manufacturing at most microprocessor fabs. Another problem they face is high Mixed Signal content in their design mixes. Analog contributes to 50% of the respin reasons in Mixed Signal designs, which means you are always in an unpredictable scramble to market. So the business is very different with lots of variability on the line, which a large, sneaker-net oriented Asian fab does best.

In memory, the patterns are uniform and design/SKU mix is low. Plus, one always needs to bring huge volumes to bear on a market. This is especially true due to the highly volatile pricing, as shown above. It does put the chip maker in a race to the bottom, but we all love to race and that’s a race you never want to lose, which is why it continues to be so popular (it also keeps us from ever becoming truly mature, which is why Art’s Ferrari dealership is doing so well in Silicon Valley — nothing like being in the red to get your blood running).

There is a huge downside to these Megafabs, as they concentrate huge amounts of capacity in one region. It’s quite feasible that a natural disaster or pandemic could take out 60-80% of fabless supply or 30-50% of memory supply. That leaves the industry at tremendous risk. Nevertheless, this is not so much due to Megafabs as it is the result of heavy government subsidies that have concentrated capacity in some rather small countries. So, it’s just a risk we live with. Nevertheless, Megafabs might not have happened without these subsidies.

*WSM: Wafer Starts per Month   NIH: Not Invented Here   GPU: Graphic Processor Unit