Moore’s Law states that the number of transistors that can be squeezed into a given slice of silicon doubles every two years (or 18 months) – something I wrote about recently and where I declared “More transistors means greater speed, more and cheaper memory and so on … ”
Except, maybe not. As the graph below, shamelessly grabbed from Herb Stutter’s “The Free Lunch Is Over: A Fundamental Turn Toward Concurrency in Software“, shows, while Moore’s Law (the green graph) holds true, the other associated improvements that we have come to expect to parallel it, such as a similar increase in efficiency per watt (royal blue graph) and clock speed (navy blue) have not. In short, we can build cheaper chips but they are not necessarily much faster.
And, as this article recounts, we are now talking about “dark silcon” – bits of chips that have to remain unpowered while other parts are in use so as to ensure the whole chip does not fry or fail due to too high power consumption.
So, if we have reached the point of “peak silicon” what can we do about it?
The chip manufacturers have responded by packing more and more cores into their devices and that works up to a point – we do not even need to have very parallel coding outside the operating system to take some advantage of that on even a typical multitasking desktop machine. But none of us are doubling the number of video renderers, MP3 decoders, database queries and spreadsheet calculations we run in parallel every 18 months, so the “Moore paradigm” of computing power doubling in that time will be lost.
A more fundamental alternative is to rewrite our software so that it becomes inherently designed to take advantage of multicore machines. Writing effective parallel software is not easy, but it can be done for lots of tasks. But even then there are limits – “Amdahl’s law” reminds us that parallelisation will only speed the parts of a program that can be run in parallel: if say we had a piece of code that must be run in serial and takes 5 seconds, and some code that currently takes 55 seconds but could be made perfectly parallel, then if we had 2 processors it takes 5 seconds (serial time), plus 27.5 seconds for the parallel code, doubling the processors but not quite halving the time, with a 46% saving. Doubling the number of processors again (to 4) cuts total computing time to 18.75 seconds but the proportional saving has dropped to 42%. In other words, the “Moore paradigm” also disappears.
The third thing we can do is look for better algorithms: the recent announcement of a vastly improved fast fourier transform (FFT) algorithm shows what can be done here – algorithmic improvement can vastly outstrip hardware speedup. But currently for many problems (those in NP space) there is no prior known algorithm available and computing power can be simply dedicated to going through all the possible algorithms looking for the one that works (we do not know what algorithms solves an NP problem but once a solution is found we can verify it ‘easily’). Assuming, as most mathematicians are said to do, that P does not equal NP (ie there is no yet to be discovered algorithm that cracks NP problems) this at least means that “peak silicon” will keep internet commerce safe for the foreseeable future but it is bad news in lots of other ways.
There is a fourth option, of course, which is to get a better physics – either for silcon fabrication, quantum computing or some other physics based innovation. Right now, though, these are probably still the least likely options but as the links below show, lots of people are working .