Change and innovation in semiconductor technology is going to continue at a rapid pace for more than the next forty years. How can that possibly be true?
After all, every industry matures and slows its rate of change and its rate of growth. During the last ten years, mature commodity industries like oil, aluminum and cotton have grown unit volumes at rates of only 1.1 percent to 2.5 percent compounded. Even with the emergence and growth of China and India during the last decade, the compound unit growth rate of automobiles and steel has been only 3.6 percent and 5.8 percent, respectively. During that same period, computer unit growth was 11.5 percent and semiconductor unit growth was 8.4 percent compounded. But the rate of unit growth of transistors was 72 percent and the average cost per transistor decreased about 35 percent per year.
While shrinking feature sizes and growing wafer diameters have provided the largest share of that growth, and will continue to be significant contributors during the next forty years, there are many other ways to continue to reduce the cost per transistor. One largely untapped opportunity is in the third dimension i.e., growing vertically instead of shrinking in the XY plane. DRAM stacks of eight or more die are already readily manufacturable, although we havenít crossed the cost per bit parity point compared to unstacked devices. Layers in the IC manufacturing process continue to increase as well. Complex packaged systems made up of multiple heterogeneous die, memory stacked on logic and interposers to connect the dice are evolving rapidly.
So what about the demand side? There is a finite limit to the number of automobiles that a family wants to purchase. Wonít there be a limit on the number of transistors? I donít think so. Itís true that demand for most material things is less elastic (or stops increasing so fast with decreasing price) as unit volumes become large. But information is truly different from automobiles, food or most commodities. There is hardly any practical limit to the amount of information humans want. While there are limits to the amount of information they can assimilate effectively, there is a virtually unlimited desire to have access to more information if it is affordable. More than anything else, semiconductors and computers facilitate our access to information.
Figure 1 shows the learning curve for transistors since the beginning of volume production in the 1950s. The unit volume has been increasing exponentially with time for the last 50 years, although in the 1960s and early Ď70s, the rate of unit increase was less stable (causing the changes in Mooreís Law at that time from doubling every year, to every two years and finally to every 1.5 years).
To the statement - "While there are limits to the amount of information they can assimilate effectively, there is a virtually unlimited desire to have access to more information if it is affordable - I would add - if it is affordable "and is actionable knowledge"
Turning information into actionable knowledge means huge increases in processing power which means huge increases in transistors.
Of course, the algorithms for all this knowledge extraction is a different matter altogether.
We need a disruptive technology that can break the extreme capital intensive model. Until then, life under someone else's reference flow will be bad. Why stick around when the same skill can be applied to social media?
David Patterson, known for his pioneering research that led to RAID, clusters and more, is part of a team at UC Berkeley that recently made its RISC-V processor architecture an open source hardware offering. We talk with Patterson and one of his colleagues behind the effort about the opportunities they see, what new kinds of designs they hope to enable and what it means for todayís commercial processor giants such as Intel, ARM and Imagination Technologies.