|Home page||Services||Past achievements||Contact||Site
|Page d'accueil||Services||Réalisations précédentes||Contact|
The advantages of implementing complete circuits on the semiconductor (as integrated circuits) rather than as discrete devices soldered together on the printed circuit board are manifold.
The advantages of implementing large scale integrated (LSI) circuits on the semiconductor rather than as small scale integrated (SSI) circuits soldered together on the printed circuit board are then the same. Thus, there is a constant drive towards larger and larger-scale integration. The ultimate aim would be to be able to use the area of the whole semiconductor wafer for laying out the system. This is the goal of wafer scale integration (WSI). But even then, the drive would continue, to use larger and larger wafer diameters.
Perhaps there could now be a revived case for adopting WSI, in the delay between 450mm wafers taking over from 300mm wafers.
Unfortunately, there are several limiting factors on the maximum size of an integrated circuit:
Integrated circuit yield
In many ways, the most serious of these limits is the first one (the use of ball grid arrays, and asynchronous logic, can be used to address the other problems). It manifests itself particularly in the parameter called yield. If the area of the chip is made too large, the alignment tolerances will be exceeded at the extremities, and parts of the circuit will fail to work. Just a single fault is sufficient to render the complete chip useless. Similarly, reducing the feature size (to enable more transistors to be fabricated within the same chip area) causes the tolerances to be tightened, and hence the chances of faults to be increased, even within a small area. Indeed, to a first approximation, the yield is exponentially dependant on the product of the integrated circuit area and the defect density for the given fabrication technology and feature size.
This is, in fact, the main reason why wafer scale integration, despite being initially researched in the 1970s, is still just a distant goal. The photolithographic, and associated, technology has been improving over those decades, but so has the technology to handle ever-larger semiconductor wafers.
One way of achieving larger integrated circuit areas is to allow faults within the integrated circuit to be tolerated. This is possible if redundancy is employed: so that parts of the circuit that are discovered as being faulty can be switched out, and replaced by spare parts that were fabricated at the same time as the original circuit.
However, this introduces a completely new set of trade-offs (Shute 1988):
Intuitive derivation of the Poisson Model
The probability, on testing an integrated circuit, chosen at random on a newly fabricated wafer, and of finding it to be working is the same as the yield, Y, of those integrated circuits (the number that are working divided by the total number that were fabricated).
We would expect that, in general, large integrated circuits are more likely to contain faults than smaller ones. Indeed, if the number of defects per unit area, D, is small, we would expect the probability of any given integrated circuit containing one of these defects to be directly proportional to its area, A. Halving the defect denisity would then halve the probability still further; so, when the values are very small, we would expect the probability of an integrated circuit to contain a defect to be equal to D.A, and hence the probability of it to be functioning correctly, to be equal to 1–D.A (that is, Y=1–D.A). More generally, we would expect Y to be a monotonically rising function of D and A, Y(D,A), but only becoming a linear function at the low values suggested above.
If we have a large integrated circuit, such as a microprocessor, composed of two major components, such as a block for the CPU and a block for the memory, we would expect the yields of those two components, Y1 and Y2, to be a function of D and their respective areas A1 and A2. The yield of the whole microprocessor device will then be the product of the separate values (since this is the probability that both one component will be found to be working and that of the other component), while the area of the new device is the sum of the separate values.
This suggests that the yield function, Y(D,A), is an exponential one. The argument of an exponential function must be dimensionless, as in Y(D,A)=bk.D.A, which can be rewritten as Y(D,A)=ek.ln(b).D.A
As noted earlier, when the values of D and A are very small, the number of faults on the wafer, of area Aw, would be about D.Aw, and distributed so thinly as to be little likelihood of more than one falling within any given integrated circuit. But one fault is all that is required to wreck an integrated circuit completely. Thus, for small values, D.Aw also represents the number of faulty chips on the wafer. Since there are about Aw/A chips fabricated on the wafer (because A is small), the proportion of faulty chips to chips fabricated would tend to D.A, and the yield function, Y(D,A), would thus tend to 1–D.A. Considering the Taylor expansion of the exponential function, and comparing this to the expression for small values of D.A, the overall yield function, at least to a first approximation, turns out to need the constant, k.ln(b), to be –1, and hence the final expression to be:
This is the Poison model for integrated circuit yield. It does assumes, somewhat simplistically, that defects are points of zero size, occurring completely independently of each other, with equal impact regardless of where they occur in the integrated circuit. Over the years, other models, such as Seeds, Murphy and Bose-Einstein, have been proposed to address this over simplification.