| | | Just to put a more concrete face on this, let's examine a hypothetical case of using 12" wafer for a 1" processor. And let's say there are 4 unrepairable defects on the wafer. This means that you can fit only 100 CPUs on that wafer (a 10x10 square inside the round wafer is all that is usable to your 1" chip). You will lose 4 to defects and end up with 96 chips you can sell.
Now suppose that the true feature size halves and your chip, with nearly exactly the same design, now takes only 0.5". This means you can now have an 11x11 usable square, within you will fit 484 chips. You still lose 4 to defects and end up with 480 chips you can sell. So in this case (given these dimensions for the chip - it will vary if you had different sizes) in this case your yield went from 96 to 480 and you got 5x more chips for using the latest fab tech. What is more, again without nearly any changes to your design at all, your new chip is twice as fast because the signals have to travel half the length they used to (not completely true, but we're making rough estimates for illustration purposes).
This made moving on with the latest fab tech a must. Every generation brought huge increases in speed and costs and as a chip maker, if you did not use this new tech, you gave your competition nearly a 10x advantage and went out of business.
The point here is that it was the economy that drove the upgrade cycle.
However, this is not really the case anymore. The new machines are astronomically expensive. And when TSM tells you that they are moving to a 1nm tech, they are not really delivering 80x increase in yield from the 9nm tech. It's just marketing BS. As a result, nobody wants to use the advanced fab unless they have to. And they only have to if they are building AI chips or something similar.
I am old enough to remember processor speeds being measured in single digit megahertz at 16 bits (which was really a fake 2x 8bits). That was around 1986. Sixteen years later, Intel's speed went from 5 MHz to 3Ghz and the bus became 32bit. So you gained a 1000x increase in speed while doubling your data pipe.
IF we had kept pace, by now your CPU would have been running at terahertz speeds. But guess what, it is still the same 3 GHz speed as before. Your improvements have come in terms of denser bigger chips that run 64bit bus. It's nice. But it pales in comparison to the previous progress.
And what I am saying is to extrapolate this same deceleration into the future because we are progressing asymptotically to the physical limits of silicon chips.
Any future progress has to come from architecture and material science, not from feature size reduction. And again, this has been true for a while, but this too getting harder and harder to achieve.
And of course you can see this in the stock performances too. The classic chip maker (INTC and MU) are still below their all time high of 25 years ago. The return long term return by most equipment makers (AMAT or KLIC) is below the SPX and even below the consumer staples.
Too many of older generation view the chips with the 1990s lens. But those times are long gone and are unlikely to return. Perhaps if/when quantum computing takes off or if/when we come up with new materials that opens new venues this will change. But for now, most chips are closer to consumer cyclicals than high premium high tech companies, even though it takes a lot of R&D.
Here's 27 years of performance. These charts percent returns, scaled to 100 at the start of the chart.
Intel
Applied Material to Consumer Staples index (you know, the food and grocery stuff)
AMAT to Costco (let's compare a leader in Fab equipment with a leader in staples/retail [not going to compare to AMZN])
 |
|
