Gordon Moore, co-founder of Intel drew a graph in 1965 predicting that transistor density on fabricated ICs would double about every two years.
Source: http://www.intel.com/technology/mooreslaw/
His prediction was based on observations of the technology used in semiconductor fabrication which was still in its early years. Moore's prediction has stood the test of time quite well. The graph on Wikipedia shows that in fact the number has doubled in fact about every two years. http://en.wikipedia.org/wiki/Moore
How did Moore get this right? First of all, it's quite doubtful that he was looking very far into the future at the time and simply projected past rates of increase based on his experience in the indusrty. In the early 60's a lot of advancement was going on in the semiconductor industry and there was huge competition fueling it. Most of the new developments were well known - secrets in terms of semiconductor dimensions were impossible to keep - you simply cut the top off your competitor's device and put it under a microscope. There you could measure the dimensions of the device geometry. If the device was sliced and stained you could also measure the amount of diffusion down into the semiconductor too. So the results of any companies work were there to see and the physics and machinery required to create the diffusion masks were all well known - expensive but well known. Since there was so much money involved, it was reasonable that steady development would continue and Moore had a pretty good chance of being right for at least a few years.
It could have come out quite differently. There have been few major changes in the way semiconductor masking is done. Mostly it's been evolution with contributions from many areas of science and engineering, ranging from mechanical to optical to chemical. If we'd had one of those stunning breakthroughs that happen very so often Moore could have massively understated the case. Imagine that instead of taking a decade for the engineering and science to evolve, someone just had a thought one day and skipped forward ten years. If that had happened there would be no curve - just one big step.
So from some point of view, Moore got lucky because engineering didn’t advance very fast.
If we look at other areas of science and engineering we can see massive increases over time. The first atomic bomb called Trinity was exploded in 1945. The yield was 19 kilotons. In 1961, just 16 years later a 50,000 kiloton bomb was detonated by the Russians - that's more than 2,600 times as powerful as the first one. If we'd doubled bomb power ever two years it would have been only 256 times as powerful. So by that standard the semiconductor industry has done very poorly. Of course, it's a grossly unfair comparison but it illustrates the math.
So will Moore's law continue on its current trend? There are possibly two major factors: heat dissipation in the device and the need for higher transistor densities. To take the second factor first, the need seems likely to grow since computing power is like garage space or vehicle horsepower, the more we have, the more we seem to want. The heat factor is a real problem though. As the frequency of operation goes up so does heat generation caused by leakage in the devices. Leakage gets worse as the device dimensions get smaller because of tunneling effects. At very small sizes the device needs to be viewed from a quantum point of view. Barriers are not solid - they are walls with probability distributions and the probability of electrons crossing the walls goes up as the walls get thinner. This is a gross oversimplification but the fact is that as the devices get smaller, leakage increases and with it heat generation. The overall problem with the heat is how to get rid of it. If we start making 3D devices the heat problem will get much worse because we will be reducing the surface area to volume ratio and the devices in the inner most parts will need to dissipate heat by conduction through the outer devices - this is one reason why we may find it hard to make 3D devices of any real size. Changes in semiconductor technology may help but I am not hopeful. Not so long ago it was predicted that gallium-arsenide (and other III-V) mixes would quickly replace Silicon because III-V devices are much faster. We see GaAs used in LEDs and some other applications but Silicon still rules today. So a clear speed advantage didn't produce a winner. I suspect we'll see Silicon technology continue to develop for a while but device geometries can’t shrink much more without some new advance.
Instead of increasing chip density we can use more chips. The trend towards parallel computing is really just getting started. We stand more chance of building really powerful computers if we can develop better ways to write parallel programs. If we get to be even moderately good at that we may not need higher density devices at all and Moore's curve will flatten out not because of limited technology but because technology took a left turn. Time will tell.
5 comments:
I am gratified to see Moore's Law on your blog, Nigel. *grins* I had anticipated it.
Love the new blog, analysis and source information.
Checkout the examples from past classes on my blog.
But is it Moore's law fitting the technology, or people busting their butts not to break Moore's law?
A bit of both I suspect. There has been a big financial advantage to higher density devices. Take a look at what Xilinx can do with an FPGA. Two Power PC cores plus a huge array of programmable logic. You can do an awful lot with one of their devices.
And keeping the device count down also minimizes interconnects which are a source of cost and failure.
So why has the Processor chip Industry stagnated on the 32 bit process by developing dual and quad core rather than embracing 64 bit and 128 bit processing chips? Prior to Digital Equipment Corporation (DEC) acquisition by Compaq they were prototyping a 128 bit and 256 bit processors as the next generation to their Alpha systems. In regards to Moore’s law the processor chip has been molded into the paradigm.
32-bit vs 64-bit is not a process decision that's a market decision and has nothing much to do with chip density. Sure, a 64-bit device has a bus that's twice as wide and we can argue that the overall device takes twice as many transistors to fabricate but we could have 128-bit devices with less complex ALUs. The graphics cards today use such devices. Bus width is only required for high-precision math and that's not a big market. The move to 64-bit is taking place in the server market because it's needed to address larger disk devices and larger memory systems than is possible with 32 bits. But size management grows exponentially with bit size so the need for more bits in bus width will taper off over time.
Nigel
Post a Comment