Last week we touched on a few of the laws governing algorithm performance on computers. These laws talked a lot about the nature of computers, how they work and communicate with each other, and the impact that this interaction has on the performance of software. This week we are going to talk about a few of the laws that govern what is possible in the development of computer hardware.
The most well-known of these laws is Moore’s Law, which recently reached an end of its usefulness. In 1965, Gordon Moore, the co-founder and CEO of Intel, observed that the density of transistors on an integrated circuit doubled about every two years. There are several arguments that the usefulness of this in regards to computational power ended in 2001, but Intel as a leader in the processor industry argues that it is still in effect and improving computing hardware today. For a couple of reasons, I am in the group that believes Moore’s Law died in 2001. The first is that in 2001, we produced the fastest single core processor every created, and we have not gotten any faster. In fact, the single core speed is now about half as fast as what it was in 2001. The second is that starting in 2001, the speed benefit of smaller transistors went away; as they got smaller they also got slower, but Intel as well as other processor manufactures started putting more cores on a single chip. We now have processors with 40 or more cores in a chip, but the speed is around half the speed of the single core chips of the 1990s. It has been agreed that Moore’s Law no longer applies as transistors have reached sizes nearing the size of a single atom and can no longer get smaller, so Moore’s Law is in effect dead, but still worthy of mention.
Koomey’s Law is very similar to Moore’s Law but is related to the energy consumption of a processor. Jonathan Koomey observed that as the transistors got smaller, they were getting more energy efficient at an average rate of doubling in efficiency every 1.5 years, but the efficiency has fallen off as Moore’s law slowed, resulting in a current doubling of efficiency only every 2.6 years.
Dennard’s Law, referred to as Dennard Scaling, was an observation in 1974 by Robert H. Dennard and his colleagues. He was attempting to provide an explanation as to how processor manufacturers were able to increase the clock frequency, and thus the speed of the processor, without significantly increasing the power consumption. Dennard basically discovered that, “As transistors get smaller, their power density stays constant, so that the power use stays in proportion with area; both voltage and current scale downward with the length of the transistor.” Dennard scaling broke down in the 2005-2006 era as it was ignoring some key factors to the overall performance of the transistor. These factors include the leakage current, which is the amount of current loss across the gate of the transistor, and the threshold voltage, which is the minimum voltage necessary to open the transistor gate. This established a minimum power consumption per transistor, regardless of the size.
Rock’s Law is an economic factor related to the cost of processor manufacturing, and probably one of the main reasons the scaling of Moore’s law was over years and not months. Arthur Rock, an investor in many early processor companies, observed that the cost of semiconductor fabrication plants doubles every four years. Rock’s Law is also known as Moore’s second law and is the economic flip side of Moore’s Law.
The last law I will talk about on computer processor design is Bell’s Law. In 1972, Gordon Bell observed that over time, low cost, general purpose computing architectures evolve, become mainstream, and eventually die out. You can see this from your cell phone; most people replace their cell phones every two years. However, his law was dealing with larger scale systems. For example, roughly every decade a new class of computers results in new usage and establishes new industry – the 1960s mainframes, 1970s minicomputers, 1980s personal computers, 1990s worldwide web, 2000s cloud computing, and 2010s handheld devices and wireless networks. Predictions indicate that the 2020s will be the decade of quantum computing.