few days ago i was reading a validation report that started "How To Repeat the Error". we are so used to the often quoted dogma "never repeat the same mistake twice". but in the world of Reliability engineering, repeatability of experiments is a mantra. an experiment should be designed in such a way that it should produce the same result consistently when repeated multiple times under the same conditions. be it PASS or FAIL, correct or wrong, the result has to be consistent and stable. else, the design of the experiment is questionable. the assumptions involved in designing the experiment requires careful analysis and in the worst case, the validity of the experiment does not hold its ground at all.
all said, there is one question that has been bothering me since i had encountered it. when you transfer a semiconductor manufacturing process from one fab to another, with all the process parameters and fab set up maintained at the same values (meaning everything DITTOed), won't we be able to produce an exactly similar piece of silicon? i was told NO. but i think, YES.
all said, there is one question that has been bothering me since i had encountered it. when you transfer a semiconductor manufacturing process from one fab to another, with all the process parameters and fab set up maintained at the same values (meaning everything DITTOed), won't we be able to produce an exactly similar piece of silicon? i was told NO. but i think, YES.
3 comments:
I would have let it go like the previous blogs but this was so offensive to all the work that goes into process transfer between fabs. Believe me, the fab is the best controlled running systems on earth.. and hope you have a understanding of the scale and precision here. Its just not fun working in parallel on umpteen short loops, being the bad guy for process engineers, having the VP breathing down ur neck with unrealistic deadlines... Yeah, you will get some Si, and with the yield, it will at best be a mirror in the cube. But, I guess you will understand better as you get deeper into this.
A typical fabrication sequence consists of atleast 700 processes and it may go upto 2000 or even more. Several process variables such as process chemicals, metrology techniques, humidity,temperature control, some residues left in the tool etc. etc..can impact one set of wafers differently than the other. Even though there are metrology techniques set up after every process step, only one in 5 wafers or one in 10 wafers are sampled. The probability of being successful is axbxc...where a,b,c are probabilities of success in each process step. But the final success is measured only at the probe where functional testing of the die is done after packaging. As Raghu mentioned the dimensions for which these processes are set to are extremely challenging. As the famous saying goes, metal interconnects are currently at the dimensions of 1/10000th of human hair. Being sarcastic Raghu, he failed to elucidate on why there is variability. Hence this lengthy comment.
Sometimes there’s this tendency of mine to channel some disappointment in the wrong direction and worse still, hope the damage dies off over time. What I put before wasn’t sarcastic, but rude and I apologize for that. But let me see if I can put the point across right this time. FOTK (Fingerson...) has a given a very good explanation of the loophole and I would like to add my 2 cents from the little of what I learnt.
The word DITTOED means a lot here. When there’s a process transfer between fabs, the other factors that come into play are the newer machines and tools with higher capability (since the destination fab a few years ahead of the origin fab), the budget that the 'division' allots for the process (sometimes you get a nth hand tool for the Fab2 since there is one available in eBay) and more importantly the possibility that the process can go wrong in the n number of parameters that are set in each tool. What is not acceptable can be as simple as that of an overetch of 10 microns which was very much within the spec limit of the origin fab. There were cases wherein the destination fab gets a lot better yield than the origin just because of the newer tools there.
From what I understand, the biggest chipmaker dittoes all its recipes across all its fabs. While its competitor takes another route and allows its fabs to develop individuality and not stressing on a exact copy and has been successful to some extent with that. But I just think it all boils down to the ingrained culture.
Post a Comment