Question about processor yields and wafers

[Shmorq]

It seems that manufacturers produce a bunch of identical chips and then sort them by the speed based on their tests. Does anyone know what percentage of them fail? And if they do fail, are they thrown out or is there anyway to repair them?

Also, why are the silicon wafers circular? It seems to make more sense to make them into a rectangular figure so less of the area is wasted.

 

[pm]

<< It seems that manufacturers produce a bunch of identical chips and then sort them by the speed based on their tests. >>

This is exactly what happens.

<< Does anyone know what percentage of them fail? >>

This depends primarily on how big the chip is. As an explanation: imagine that you have a bunch of grains of rice and throw them on a wafer. Then imagine that the wafer is divided into big squares, say 16 total. The odds are fairly high that a grain of rice (a defect) landed in a square because the squares are very large. If you reduce the size of the squares, more of them are likely to be rice-free.

<< And if they do fail, are they thrown out or is there anyway to repair them? >>

Outside of memory (in SRAMs, DRAMS, and in microprocessor caches), it is not impossible to repair them, but it would be a tedious affair with a random defect and would certainly cost many hundreds (if not possibly thousands) of times more than simply manufacturing another. Random defects, except in memory elements, are not repaired. In certain memory structures (DRAMs and microprocessor caches), redudancy is built into the structure and is tested and fuses are blown to kill defective rows.

<< Also, why are the silicon wafers circular? It seems to make more sense to make them into a rectangular figure so less of the area is wasted. >>

This showed up in another current thread and Outersquare had a good reply, but I can't remember the title of the thread. Wafers are circular due to the method by which they are produced. For silicon chips, all of the silicon atoms need to be aligned in the same direction. So not only do you need a very pure material, but everything must line up exactly. It turns out that there's a relatively &quot;easy&quot; way to do this, you take a chunk of aligned silicon and use it as a 'seed', dip it into a vat of silicon at just the right temperature and pull it out very slowly. You end up with something that looks like a silicon sausage. Then you cut the ends off and slice it cleanly into very thin slices which are then wafers. No one has found a way to do this such that the wafers are rectangular and this method is so much easier than other methods that could produce square or rectangular wafers.

 

[DeeK]

<< Also, why are the silicon wafers circular? It seems to make more sense to make them into a rectangular figure so less of the area is wasted. >>


For minimal silicon crystal structure defects, the rods from which wafers are cut are made using the Czochralski method. A seed crystal is rotated at the end of a spindle and pulled out of molten silicon to create a cylindrical silicon ingot. The rotational speed and the lift rate can be adjusted to vary the diameter of the ingot. The wafers are cut from the ingot using a diamond saw. There's a basic illustration of that method here.

 

[MaxImuMAdVaNtAgE]

i just read an interesting article about a supercomputer HP developed called teramac which uses defective processors and utilizes built-in redundancies to make them perfectly usable...pretty cool idea, and it's cheap cuz they use cpus that they normaly wouldn't

MaxImuM

 

[Passions]

I am pretty sure that processor yields are protected secrets, HEE HEE.

 

[Shmorq]

This is all very interesting.

For processors, is redunancy only used on the cache since it takes up the most room (at least I think it does) or because it's tougher to have redundancy for the rest of the CPU?

I just looked up the Teramac stuff and it's very interesting. HP claims that it performs much faster than a normal workstation even though it has around 200,000 hardware defects. Cool stuff. So only 1 of these was ever built?

 

[dkozloski]

IBM has technology to handle materials one molecule at a time and can actually repair junctions and traces in semiconductor IC's. This is used in prototyping to make small modifications or repairs as part of a developement process. It is much too expensive for repairing production parts. For demonstration purposes they have positioned individual molecules to form the IBM logo. Scientific American Magazine has had some good articles on this work.

 

[pm]

<< For processors, is redunancy only used on the cache since it takes up the most room (at least I think it does) or because it's tougher to have redundancy for the rest of the CPU? >>

Both. Because caches take up a lot of room and are thus big enough to possibly have a problem, and also because it's relatively easy to do on memory arrays, but vastly harder for datapath.


The atomic positioning work that IBM has done uses STM (Scanning Tunneling Microscopy) which is capable of moving individual atoms around. Follow this link for more details.

Rewiring completely manufactured chips is usually accomplished with millions of atoms using FIB (Focused Ion Beam) machines in coordination with SEM (Scanning Electron Microscope). Follow this link for more details.