Intel (INTC +3.86%) usually launches a new processor family on an annual basis. Each year, Intel adds new features and functionality. This allows personal computer vendors to market new features, higher performance, and so on, which helps keep the PC upgrade cycle going.
New chip features are often added by increasing the number of transistors a chip is made up of. Chip companies need to be deliberate in choosing how many transistors they will use at a particular generation because more transistors translates into larger chip area, which impacts costs.
This article will look at how Intel chooses transistor counts and die sizes for its PC processors from a financial perspective.
Looking for a pattern
The following table includes the die sizes of a number of Intel processors. To keep the comparison apples-to-apples, I'll look at chips with two cores and what Intel calls a "GT2" graphics configuration. This allows us to see the progression in transistor count and chip sizes over the generations on comparably positioned parts.
|
Year |
Processor Family |
Lithography |
Transistor Count |
Die Size |
|---|---|---|---|---|
|
2011 |
Sandy Bridge |
32nm |
624 million |
149mm^2 |
|
2012 |
Ivy Bridge |
22nm |
830 million |
118mm^2 |
|
2013 |
Haswell |
22nm |
960 million |
131mm^2 |
|
2014 |
Broadwell |
14nm |
1300 million |
82mm^2 |
Sources: AnandTech, Intel, Wikipedia, SemiAccurate.
Each lithography transition roughly doubles the number of transistors that can be placed in a given area. Generally speaking, new manufacturing technologies cost more on a per-wafer basis, but the density increase reduces the cost per transistor significantly. This allows Intel to place more transistors (adding more functionality) on a chip while keeping the cost structure roughly the same generation-on-generation.
Consider the following table, which illustrates the generation-on-generation changes in transistor counts and die sizes:
|
Year |
Processor Family |
Lithography |
% change in transistor count from previous |
% change in area |
|---|---|---|---|---|
|
2011 |
Sandy Bridge |
32nm |
N/A |
N/A |
|
2012 |
Ivy Bridge |
22nm |
+33.01% |
-20.8% |
|
2013 |
Haswell |
22nm |
+15.66% |
+11.01% |
|
2014 |
Broadwell |
14nm |
+35.4% |
-37.4% |
Source: author calculations.
Now, look at the following chart Intel showed at its November 2014 investor meeting:
Source: Intel.
Notice that when Haswell ramped up production in the third quarter of 2013, its cost structure (for comparably positioned parts) was roughly on par with where Ivy Bridge was in the first quarter of 2013. Also note that as Haswell production progressed, it became noticeably cheaper to build than Ivy Bridge, despite being built on the same manufacturing technology and having a larger die size.
Additionally, notice that by the third quarter of 2015, Intel's 14-nanometer Broadwell is expected to be slightly cheaper to build than Haswell was by the second quarter of 2014.
Here's how Intel chooses its die sizes
From the data here, it seems that Intel takes advantage of more mature yields and/or improved manufacturing technologies to bring new features to market. However, Intel seems to restrict its transistor budgets for new products to allow for future products to keep production costs flat-to-down relative to prior generation products.
The pattern goes something like this: When Intel moves to an entirely new manufacturing technology, it can afford a significant step up in transistor count as the area per transistor drops significantly. Then, as that manufacturing technology matures, Intel can release a follow-up architecture with more transistors and a slightly larger chip size, and higher-yield rates at that point will more than offset the cost increase associated with a larger chip sizes.
