> Even with current low cost silicon, there is still a high rejection > rate. That, as was said , cannot be afforded with a bigger chip.
Actually, bigger chips increase the number of rejects incredibly. The reason is very simple: the basic idea behind chips in most cases relies on all chip components being operational, i.e. there is no fault tolerance. There are some exceptions, though, see below. Take a silicon wafer. Now immagine there are specks of unusable silicon, say a couple of um across randomly distributed with a density of one on every 2 square cm. There are two factors that govern the yield: 1) size of geometry - if this is >> than the size of the anomaly, there is a good chance an anomaly will only produce a degraded component, not a completely faulty one. However, today, almost all geometry used is far smaller than the sizes of the anomalies, so we run into problem: 2) size of the chip. Obviously, if the chip size is >2cm square, statistically ALL of them will have a fault, i.e. yield will be 0. As the chip gets smaller, somewhere around half of the 2cm square area, the yield suddenly goes up quickly. For very small chips, the number of failures approaches area_of_anomaly/area_of_wafer * 100 in %, i.e. the yield becomes virtually 100%. This is why small signal transistors, having a very small die, cost pennies, but a CPU that has a die of 100 times the size, does not cost 100 pennies - all sorts of additional processing are necessary to even get >0 yield on these, and it has to be payed for. In reality, anomalies on silicon are not the only problem, there is a vast number of different pollutants that can affect the process of making a chip, but the basic behaviour is the same. This is why chip prices are extremely dependant on chip size, and why all the manufacturers try to squeeze the size of the die down as much as possible. This problem first became evident with memory, as these were traditionally the largest chips. For a long time it held back the jump from 64k bits to 256k bits for dynamic RAM. Finally, someone figured out that providing extra RAM 'rows' that were programmable will include a dose of fault tolerance. However, this came at a price - traditionally, with a shrinkage of geometry, comes a corresponding shrinkage of delays, i.e. speed increases. But since the 256kb DRAM needed 'programmable' rather than fixed row decoders, some of the speed benefit was lost, so the 256kb DRAM chips were of the same speed grade as 64kb chis available earlyer, so one speed grade jump was 'missed'. Today all memory produced has some fault tolerance, it is tested and then appropriately programmed at the factory. Furthermore, even CPUs have a dose of fault tolerance. It is mostly evident with chips that come in different falvours, such as different speed grades, or cache sizes (an asside: being memory, most caches in todays big CPUs also have 'extra' memory cells to provide fault tolerance, or, in some cases, error detection and correction schemes). This same problem reared it's head in the manufacture of active matrix LCDs, which are the most extreme form of a chip - the size of the whole screen. Cost effective sizes ginally jumped from about 9" diagonal to more once they figured out how to make the displays line by line - the lines are produced on a drum and 'stuck onto' the glass, one by one. They are tested as they get stuck, and if found defective, the whole line is scrubbed off, then replaced by a new one from the drum - instead of throwing the whle screen away. In parallel with these technology, material technology advances also, so as time goes by, and prodcts mature, they actually move downwards in the technology chain. For instance, to get the first 15" LCDs, the drum technology was required. Advances in materials made it possible to produce 15" screens using traditional technology today, but at the same time, combined with the drum thing, now you can make 19" screens. Nasta