Semiconductor fabrication [Split from: Do objects move faster...]

[StringJunky]

15 minutes ago, Strange said:

As I have spent a large part of my career designing CPUs I know this not true. But you carry on with your irrational beliefs. 

Considering what it costs to set up to manufacture a specific CPU, I doubt there's much trial and error.

[Strange]

55 minutes ago, StringJunky said:

Considering what it costs to set up to manufacture a specific CPU, I doubt there's much trial and error.

Absolutely not. When you’re signing off millions of dollars to start manufacturing you need to be pretty certain it will work. 

[DrP]

5 minutes ago, Strange said:

Absolutely not. When you’re signing off millions of dollars to start manufacturing you need to be pretty certain it will work. 

I saw a program/documentary about chip manufacture about 25 years ago.  On every Si wafer they made, lets say, 50 CPUs (can't remember how many they get per wafer)....  and they expected about 2 or 3 to work.  I am pretty sure with the development of better resists for microlithography that the yields have improved over and over as the years have gone by, lol.

Do you know what kind of yield they get these days? I might look it up when I get home tonight if I get the chance.  I would 'guess' that the yield have been getting better and better....  but as the writing has been getting smaller and smaller this could still throw up errors.

 

 

[DrP]

1 hour ago, DrP said:

Do you know what kind of yield they get these days?

Wiki suggests the die yields are pretty secretive between manufacturers but can be 'as low as 30%'.    (sorry - off topic I know - it seemed relevant to the conversation at the time).

Edited November 1, 2018 by DrP

[Strange]

8 hours ago, DrP said:

I saw a program/documentary about chip manufacture about 25 years ago.  On every Si wafer they made, lets say, 50 CPUs (can't remember how many they get per wafer)....  and they expected about 2 or 3 to work. 

I would imagine the number would normally have been in the hundreds, rather than 50-ish. But I have worked on a couple of projects where they pushed the limits of how large a die could be and the numbers probably weren't far off that.

8 hours ago, DrP said:

Do you know what kind of yield they get these days?

It varies enormously, depending on the technology, the type of circuit, the size of the die, etc. Defects are pretty much randomly distributed across the wafer so you can work out a curve for the probability of a fault in a die versus the die size. But it isn't quite as simple as that because there are areas of the die that are more susceptible to faults than others (for example, it might have no effect if it is near the edge of the die where there is no circuitry).

And then it depends on what the device is. A memory chip will have large amounts of redundancy so if there is a fault that stops one row working, that row can just be marked as invalid and we still end up with a working device. (In the case of flash memories, where there are regular failures of memory cells during their lifetime they constantly map out new defective cells as they fail to keep the memory working at full capacity for longer.) 

In the case of a CPU a defect anywhere in the circuit will probably render it useless. Although, modern feature sizes mean that CPU dies contain large amounts of memory, so you can get better yield because of that.

6 hours ago, DrP said:

Wiki suggests the die yields are pretty secretive between manufacturers but can be 'as low as 30%'. 

I would say this is the most confidential manufacturing data there is. Companies will often present their latest (and even up-and-coming) technologies at conferences, etc. But they will never talk about yields.

Yield is very variable. It normally starts out quite low with a new process (possibly worse than the 30% figure) and then will improve dramatically as the process is tuned. For high volume memory chips (DRAM) it can be up in the high 90s (I think. Even I was not privy to accurate yield figures).

 

[Enthalpy]

Manufacturers don't speak about yields. When starting a process, yield is catastrophic, and 0% is normal - or rather, a new factory and process get tuned before any marketable product runs on the chain. But when production is near its maximum volume, yields are closer to 1 than to 30%.

Several other factors influence the yield. The chip's size, because the same factory can make 3cm² GPU and 0.3cm² southbridges. Redundancy, where possible, with highly uniform chips having 1 or 2 units more than the minimum guaranteed to the customer.

Sometimes defective chips are sold as good chips of an other series, typically for GPU: the MegaSuper XXX has 20 zlimmers while the Super XX has 18. The company tries to make all 20 work, but if 1 or 2 zlimmers are defective, they sell the chip with 2 zlimmers turned off. And when the market wants more Super XX, the company turns good zlimmers off. Then come smart customers who unlock the zlimmers and have a MegaSuper XXX at the Super XX's price.

This operation works more often than not, telling that yields are near to 1, not to 30%.