Moore's Law Timeline

Question

I find it incredible how transistors per area keeps increasing. How has it been achieved so far? My ignorance tells me that if ICs were designed properly from stage to stage they should have got this far in a lot shorter a time, but at the same time I'm sure it's been a vast number of incremental improvements. The question is, what improvements were they? Were they all variations on a theme, or completely different improvements, it was probably a mix, but some enlightenment as to what kind of improvements they were, and why it's done in so many small increments.

Was it mostly improvements in the photo-lithography? Or transistor/circuit designs that allowed greater tolerance of imperfections? Or material science improvements which allowed higher quality materials in the transistors, tracks, and layering? Any other facets?

Thanks

Answer 1

If you have ever worked on a seriously complex technical project, you will know that it's basically impossible to design something properly from the beginning.

Think about it. If cave men had just thought properly, then they should have been walking on the moon 100,000 years ago.

Manufacturing modern semiconductors is a seriously difficult business, and it involved so many engineering challenges that had to be overcome in order to make it possible. You can't overcome these challenges simply by designing something right in the first place. The only way to do it is to take baby steps. Get a new technology running. It won't be very good to start with. There will be lots of imperfections in the process, and the yield will be low. Slowly people work out how to optimise the process variables in order to make the process reliable, and get the yield closer to 100%. Then you take another baby step.

In theory there's no difference between theory and practice, but in practice it is.

In order to progress from the integrated circuit to today's multicore CPU took innovations in:

• Chemistry: coatings, ultra pure crystal growth
• Optics: How do you focus photons which are larger than the features you are making? How do you generate a light source bright enough and at the short wavelength you need. That light source can be one of the largest consumers of power in a semiconductor fab.
• Mechanical aspects: techniques for polishing silicon wafers ultra flat. Accurately registering (positioning) the wafers for repeated exposures.
• Computing: You need a powerful computer to be able to design a powerful CPU. Catch 22.
• Construction: Massively complex and expensive fabs had to be build in order to build these things reliably and economically.

" they should have got this far in a lot shorter a time "

Really? It's been only 53 years since the first integrated circuit was patented in 1959. That's amazingly quick, considering humans have been around for hundreds of thousands of years, and most of this time they made no progress at all in integrated circuits.

Answer 2

One of the improvements is not electronic, but rather optical. Wafer steppers which are used to project the patterns for different layers onto the wafer's photoresists use optical lenses. In the 80's when feature sizes of a few microns were common it was feared that at feature sizes below about 400 nm (the limit for visible light) the optical system used wouldn't suffice anymore.

Today we have feature sizes down to 22 nm, and the steppers still use optics to transfer the patterns. But not the 80s' optics, they weren't good enough for this kind of resolution.

Answer 3

This is a very competitive industry. If some company could have made 100 nm devices in 1985 they would have. It is precisely because of this competitiveness that Moore's law continues to hold.

Shrinking linear dimensions by a factor of 2 isn't just about one thing. Advancements need to be made on a number of fronts to make a real world profitable chip possible. One of the technology limits, as Steven mentioned, has been photolithography, but there have been many others. I'm not a chip or fab designer so I don't know all the details. I do know that the investment in a new smaller feature size fab process is huge. Usually companies build whole new fabs for a new process because it's not as simple as just replacing a single machine with a better one. Just the air handling alone is a big issue, and there are many others.

Making smaller transistors is only part of making smaller chips. You have to consider the electrical properties of the transistors as they get smaller. The dissipation per area increases, which drives the operating voltage down, but that gives a lower ratio between FET on and off current. That in turn drives up leakage current, which increases quiescient dissipation. Better thermal conductivity to the case is needed, and better heat transfer on the board, etc. This goes on and on and on with many interacting parameters.

I'm old enough to remember several "barriers" where basic physics supposedly said we couldn't go any further and Moore's law was doomed to stall. Each time clever people found a way to do something different to get around the physics. I don't know enough myself to have a good idea when the pace of advancement will slow down. Having watched this process since the mid 1970s I have been really impressed how many cycles of Moore's law there have already been, and how vastly computing has changed in a fraction of a lifetime.

Answer 4

1. Economics dictates a new wafer fabrication process change every 2 years. New equipment costs Billions while construction & tuning process and design takes time to optimize high yields and then must be amortized during production. Intel & IBM are leaders in this game with R&D patents and process capabilities.
2. Design changes include flash memory going from binary to N quantization levels so using DAC<>ADC can get log N more density per cell, but this added overhead and large ECC codecs. every other area improved as well.

3. IBM have now invented RAM cells that may take 5~10 yrs to produce 150 TB chips going from 1e6 atoms down to 12 atoms using antiferromagnetic crystal lattice

4. Improvements include many material changes such as;

   • strained silicon, introduced with the 90nm process in 2003
   • hafnium-based gate-last high-k metal gate (HKMG)

There are too many changes to summarize realization of Moore's Law, but it is accomplished in every layer and department; funding, research, design, architecture, fabrication, materials, processes, redundancy and error correction.

Funny thing, it is not a law of Physics, just a peculiar pattern of growth or shrinkage depending how you look at it.

Gordon Moore is 83, retired / Chairman Emeritus, co-founder and former Chairman and CEO of Intel Corporation.

added

A huge part of growth in CPU's has to be granted to the cost reductions in $/GB of RAM. In addition to area density, hierarchical architecture there are dozens of other factors such as cycle time reduction from 100 hrs to 36hrs in the 90's for making each chip.

The major Asian memory companies have competed and continue to succeed in this area. This article details some interesting reasons that are relevant to the challenges of "Moore's law" and Memory.