Author: Robert Hyatt
Date: 10:31:54 02/27/03
Go up one level in this thread
On February 26, 2003 at 12:03:42, Steve J wrote: >>5. I am also looking for some predictions/information about processor speed in >>20-30 years from now. For micro's Moore's law still holds. So 21 years is 7 >>doublings of speed or 128 times as fast as today. > > > I've spent 25 years in manufacturing side of the semiconductor industry and >would like to introduce what I call "No-Moore's Law". It describes the physical >limitations that silicon (or any other compound) will run out of gas and can be >shrunk no more. It also talks about some of the financial limitations of >shrinking die. > What you have written can't be true. Because if you read this forum long enough, you realize that many are convinced that designing chips is a trial-and-error process where the engineers don't know anything at all about how fast a chip will run until it is produced and tested. No ideas about the expected wafer defect rate. Etc. It was a nice explanation of a known issue, but it can't be right because it implies that the engineers really know what they are doing, rather than relying on blind luck to get something to work. :) Of course, all the engineers I personally know are repeating your story and they are sticking to it. But they must all be mistaken. After all engineering isn't a true science, it is based mostly on serendipity. :) Bob > There are several key points. 1) There are physical limitations to what >degree a transistor size can be shrunk. This is based on the size of the atom, >and 2) There are exponential increases in the costs of fabs and mask sets as >each reduction takes place. These will cause a practical end to the amount that >die sizes can be shrunk. > > > First, let's take a look at the existing "90 nanometer" process. The >operations of the circuits relies on tightly controlled processes. While the >circuits and processes are controlled very tightly, a 10% mismatch between >components on a given die can be fatal. > The size of a silicon atom is .3 nanometers. This means that existing >processes are about 300 atoms across. At this size, a one atom variance is .3%. > However, if the transistor size is halved, for example, five time, then it will >be 300/(2^5) =~ 10 atoms across. This means only one atom variance will cause a >mismatch of 10%! Added to that is normal processing variance which will makes >the product not manufacturable. > If we assume that the size is halved every two years, then there is about 10 >years left in Moore's law. > > > Compounding the problem is the exponential cost of making fabrication lines of >finer size transistors. From a historical perspective, 25 years ago it cost >under 10 Million dollars to put together a fab (equipment, extra cost for a >clean environment, etc). EACH product that was made in a fab would have a >dedicated mask set as tooling to make the product in that fab. This tooling >cost in the range of $10k to $15k. > At that time not only did every company, but, every product line within a >company that had more than $50 Million in sales would have their own fab line. >Many companies would have 6 to 8 (or more) fab lines. > The costs of leading edge processes have increase dramatically. There are no >more $10 Million fabs being made. Many new fabs are costing $1 Billion or more! > This has caused a dramatic shift in fab investments. Not too surprisingly, >very few companies can afford to make a leading edge fab, and instead, rely on >companies like TSMC and UMC to make the large investments and allocate the cost >into the costs into the sales price of the wafers. > While this has provided a working business model, as the cost of fabs continue >to double, there will be a point at which the incremental savings from a new >process technology will be too expensive to justify the cost of the fab. > > The mask set tooling has also increased dramatically. Instead of $10k to $15k >dollars for a set, maturing processes of today cost $100k. Products that are in >design right now are forecasted to have mask set tooling costs in excess of >$500k. Given that the entire annual budget for smaller companies (including >salaries, rent, etc) can be $5 Million, it will not take long before fewer and >fewer companies will be able to make a run at the market with new products. > > > The bottom line is that physical and financial constraint will bring an end to >Moore's law. Realistically, it will not be an abrupt halt, but, instead from >doubling every two years, to double every four years, then to 5% increase per >year. > My bet is that we will see a dramatic slowing in 7 to 10 years. > Beyond that, we will rely increasingly on more processors per system and other >techniques instead of more transistors per processor die.
This page took 0.05 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.