A decade ago, we entered the age of multicore in terms of computer architectures. The clock frequency could not be raised anymore, but we still can put more transistors onto a chip. So multiple cores were put onto single chips. A rapid increase of cores was predicted. Today, multicores are everywhere, but the number of cores is only a single digit. Of course, you can buy a Xeon Phi, but where is the PC with 30 cores? The answer is that we "hit the Power Wall".

To explain that, let us look at how processors are scaled. Suppose we have a chip with 16 cores built with 180nm technology. If we use the twice as fine 90nm technology, we get four times as many transistors, since area scales quadratically relative to length. So, we can put 64 cores onto the chip. Additionally, transistors also become faster, if they get smaller. The electrons have to travel 90nm instead of 180nm, although it is not a straight way, so the factor is roughly 1.4 times. Both effects combined we should see computations 2.8 times faster.

Now for the bad news. Since there are four times as many transistors working, they also draw more power, but nowadays power must be kept constant. For mobile devices, we must not drain the batteries. For stationary devices, leading off the heat is the limiting factor. Instead, we can lower the voltage and keep the frequency up. Unfortunately, there is a lower limit to the voltage due to leakage. These effects are increasingly canceling out the computation improvements. So scaling further becomes useless at some point and this point we call the "Power Wall".

There are four possible ways around this wall. First, we could refrain from building ever more cores into chips and just shrink the chips keeping their performance constant. Second, we could lower the clock frequency, so they do not run that hot, but are slower individually. Third, we put a lot of specialized cores on the chip, which are as efficient for special tasks as many generic cores. Fourth, we find a better transistor technology than the current CMOS approach. All of those are currently pursued by various researchers and companies.

© 2013-02-24
qznc