The Apple A9X in all of its glory. Image source: Chipworks.
In an earlier column, I tried to provide an estimate of the manufacturing costs of the Apple (AAPL +0.11%) A9 chip found inside of the iPhone 6s and 6s Plus. The estimate worked out to something in the range of $22 to $24, in line with an estimate provided by IHS iSuppli.
Thanks to the folks at Chipworks, we now know how large the A9X chip inside of Apple's iPad Pro is. Armed with this knowledge, I'd like to offer up an estimate of how much it costs to build that particular chip.
Putting together all of the variables
As I wrote in the previously linked column, there are three major variables that go into determining the manufacturing cost of a chip:
- Finished wafer cost
- Chip die size
- Manufacturing yields
The impact of silicon wafer cost should be obvious. The chip die size impacts both the total number of chips that can be physically built on a given wafer as well as the manufacturing yields of the chip (there is more that can "go wrong" on a larger chip than on a smaller chip; see here).
And, of course, manufacturing yields impact costs, as the more good chips a company can extract per wafer, the lower the effective cost per chip is.
For wafer cost, I once again assume $8,400 for a Taiwan Semiconductor Manufacturing Co. (TSM 0.72%) 16-nanometer wafer (this estimate comes from IBS' Handel Jones, and is what Apple pays for the wafer, which includes Taiwan Semi's margins). For die size, I am using 147 square millimeters (a measurement provided by Chipworks).
The last variable here is manufacturing yields, a number that's difficult to estimate. With the A9, I assumed yields of 80% (citing a remark from an AnandTech article discussing yields of the Taiwan Semi-built 16nm HiSilicon Kirin 950 chip).
The A9X is a larger chip, which is the major factor that leads me to believe yields will be lower. The A9X also runs the CPU complex at a more aggressive frequency, which could mean lower "parametric yields" (i.e. it's easier to build a CPU that runs at 1.85 GHz than the same CPU at 2.26 GHz).
Although I must admit that there is a lot of margin for error here, I'm going to estimate manufacturing yields on the A9X at approximately 65% (though I expect them to improve over the lifetime of the A9X).
Doing the math ... again
Using Silicon Edge's dies-per-wafer estimator, I get approximately 401 dies per wafer. Assuming yields of 65% gives us about 261 good dies per wafer. Dividing the wafer cost estimate of $8400 by the estimated number of good dies yields a die cost of around $32.30.
In the A9 article, I assumed packaging costs of $3 to $5 per chip. For the sake of conservatism, I'm going to assume the high end of that range here and call it $5.
With that, I arrive at an estimate of $37.30 to manufacture the A9X.
Does this make sense? A quick sanity check
At $37.30, the A9X is significantly more expensive than the A9, which I estimate to cost Apple somewhere in the range of $22 to $24. If we assume that the A9 costs Apple $23, then the A9X costs Apple 62% more to build, which makes sense given that the A9X chip die itself is around 40% larger than the A9's (die cost increasing faster than die size on a given process makes sense).
I think this estimate passes a sanity check and is within the right ballpark (both on an absolute basis and relatively speaking), though I must caution that there is room for error here as my yield rate estimates for both the A9 and A9X represent merely educated guesses.
