Why some analyses are so misleading.
In the first article in my series on sustainability, I cited one estimate that attributed most of the electricity consumed by an integrated circuit to manufacturing, not use. Other analyses, however, come to exactly the opposite conclusion, with above 90% of lifetime energy consumption accounted for by the use phase. How can that be?
The glib answer is that industry efforts to build more efficient fabs are working. The energy used to manufacture an average die has come down, even as the well-publicized limits of voltage scaling have driven use-phase energy consumption up. This answer also happens to be true.
The real answer is more nuanced, though, and strikes at one key challenge in lifecycle analysis. What is an average die? More generally, what is the correct unit of production to use for lifecycle analysis?
The fundamental unit of IC manufacturing is the transistor. While the energy needed to manufacture a million transistors has plummeted, they can’t accomplish very much in the current era. The Intel 80486 microprocessor contained about 1.1 million transistors and used a 1 micron manufacturing process; the company’s recently announced top end Xeon Broadwell E-5 has 7.2 billion transistors on two dies, built with a 14 nm process. Moreover, each of its 22 cores is supported by 2.5 MB of low-level cache memory. Industry-wide, solid state memory devices account for vast amounts of silicon area. Is memory the fundamental unit?
Observing this dilemma, a study at University of California, Berkeley chose the “average” die as the fundamental unit of production, with “average” defined by the ITRS cost-performance CMOS logic specification. As any good statistics student knows, however, extreme high and low values can distort an “average” value so that it does not represent the whole population. As the IC industry has diversified, identifying a “typical” device has become quite difficult. Smartphones contain low power microprocessors and fast non-volatile memory, but also accelerometers, noise filtering microphones, transmitters and receivers for cellular, wireless, and Bluetooth signals, and more. At the other extreme, their capabilities are supported by massive data centers lined with racks of server-class systems. Which of these has an “average” power use profile?
Though the conceptual simplicity of life cycle analysis makes it attractive, especially for communicating with the general public, it is easy to see how it can give misleading results. Analyses focused on individual products or fabs are likely to give much more precise and actionable results.