5 ms·

I spent some time working in a fab. Nothing is cheap, competition is tough, yield is everything. 99.8% yield on each step of a 500 step process is good for 38%

I spent some time working in a fab. Nothing is cheap, competition is tough, yield is everything. 99.8% yield on each step of a 500 step process is good for 38% overall recovery. These events have been shown to be very rare, and the costs would greatly outweigh the benefits (for the time being). The cheapest piece of equipment was the stand the FOUPS come down and rest on. From memory, that was easily 5 figures, perhaps low 6. Every other piece of processing equipment was insanely more expensive. Queuing theory is an active field of research with great applications in how wafer processing. The total process has to be looked at. Wafer priority, wafer value, planning preventative maintenance on the machines, etc. could all be taken into consideration when scheduling when wafers will run. Retrofitting every "robot" that grabs FOUPS for a much heavier load, upgrading the tracks if needed, replacing every single FOUP, adjusting factory flow and inventory holding areas, doing qualifying trials on the FOUPS (to make sure they're functioning correctly and not leaking), etc. are considerable expense. Prioritization of wafers and timing the preceding step so the required machine is certain to be ready in the window would yield better results from a strict business ROI perspective.

toast_coder 8y ago
>99.8% yield on each step of a 500 step process is good for 38% overall recovery. What does that mean? Is 'yield' not the end usable product from the batch?
- mirimir 8y ago
  Well, Calc tells me that 0.998^500 is 0.3675 :) It's a familiar concept from organic chemistry. Edit: If yield per step averages 99.8%, and there are 500 steps, overall yield is 36.8%.
  - oh_sigh 8y ago
    'yield' is a really impossible measure for this kind of process, because it presupposes full end to end knowledge of fabrication processes.
    - froindt 8y ago
      Not sure I follow. The 99.8% is just an example number. If one machine messed up really bad and had just 50% yield, all other machines would need to do far better in terms of overall quality to achieve 36% yield. Of course every single machine isn't at the same quality level. At the same time, when wafers get scrapped, they do justify why they were scrapped. There are metrology processes and tests performed as you go to ensure you don't run ruined wafers through additional manufacturing steps.
- froindt 8y ago
  Yes it is. The 99.8% hypothetical number is on a per operation basis. Depending on the layers and complexity of the chip, the ballpark number of operations might be 300-1500 (lots of hand waiving here). The gist of that part is meant to show how important and impressive the entire process is. In "normal" manufacturing, 99.8% good parts is a pretty darn good process. Many of the easy wins are already implemented. Most normal things are manufactured in far fewer steps, so even if you're only making 95% good parts, it doesn't absolutely kill your total recovery (start to finish).
mirimir 8y ago
> These events have been shown to be very rare, and the costs would greatly outweigh the benefits (for the time being). Yes. And I was responding to georgeburdell's comment about long-term losses from this virus infection. In that maybe some crude hack could ameliorate them. > For those of you thinking that TSMC only gets set back by the time it took to recover the equipment from the virus (1-2 days), realize that there are some critical steps in the manufacturing process that require completion within X time or the wafer is trash. Also realize that wafers take several weeks of fab processing, and this may be a substantial hit to their output for a while.
- avs733 8y ago
  It's not a bad argument just an one lacking some information on how this all works. And sincerely that isn't meant as an insult...it's just reality. Ask me about surgery and I'll shrug and say 'i got nothing yo' Everything in semiconductor is basically incomprehensible in scale. A scanner (i.e., the photo lithography projection tool) takes about 2 seconds to fully expose a wafer. that wafer probably has 100 individual chips on it each of which have around 2 billion transistors. That means that tool is helping create 100 billion transistors a second. The great grand parent comment is right...but the lose of that is probably on the range of one-two weeks production. retrofitting the fab like you suggest, and to be clear it isn't a fundamentally bad idea, would cost billions and take months if not years. What you propose is technically correct but business irrelevant. Long term here means a week or so...it's expensive but rare. They don't play with 'crude hacks' because you don't know the implications of that hack 10 years from now when the chip is running part of the stock market. It's just an industry that is hard to process, whether you work in it or not, because of the scale and the precision.
- froindt 8y ago
  Sorry if that came off abrasively. avs733 mentioned everything is on a different scale in semiconductors and I forgot how long it took me to grasp that. In school (industrial engineering) we talked about length and thickness tolerances being quite tight at 1/10,000 inch. The units used in semiconductors are nanometers, angstroms, etc. You might do a process and lay down just a couple atom thick layer of material...I had no idea that was physically possible.

toast_coder 8y ago

>99.8% yield on each step of a 500 step process is good for 38% overall recovery. What does that mean? Is 'yield' not the end usable product from the batch?

mirimir 8y ago
Well, Calc tells me that 0.998^500 is 0.3675 :) It's a familiar concept from organic chemistry. Edit: If yield per step averages 99.8%, and there are 500 steps, overall yield is 36.8%.
- oh_sigh 8y ago
  'yield' is a really impossible measure for this kind of process, because it presupposes full end to end knowledge of fabrication processes.
  - froindt 8y ago
    Not sure I follow. The 99.8% is just an example number. If one machine messed up really bad and had just 50% yield, all other machines would need to do far better in terms of overall quality to achieve 36% yield. Of course every single machine isn't at the same quality level. At the same time, when wafers get scrapped, they do justify why they were scrapped. There are metrology processes and tests performed as you go to ensure you don't run ruined wafers through additional manufacturing steps.
froindt 8y ago
Yes it is. The 99.8% hypothetical number is on a per operation basis. Depending on the layers and complexity of the chip, the ballpark number of operations might be 300-1500 (lots of hand waiving here). The gist of that part is meant to show how important and impressive the entire process is. In "normal" manufacturing, 99.8% good parts is a pretty darn good process. Many of the easy wins are already implemented. Most normal things are manufactured in far fewer steps, so even if you're only making 95% good parts, it doesn't absolutely kill your total recovery (start to finish).

mirimir 8y ago

> These events have been shown to be very rare, and the costs would greatly outweigh the benefits (for the time being). Yes. And I was responding to georgeburdell's comment about long-term losses from this virus infection. In that maybe some crude hack could ameliorate them. > For those of you thinking that TSMC only gets set back by the time it took to recover the equipment from the virus (1-2 days), realize that there are some critical steps in the manufacturing process that require completion within X time or the wafer is trash. Also realize that wafers take several weeks of fab processing, and this may be a substantial hit to their output for a while.

avs733 8y ago
It's not a bad argument just an one lacking some information on how this all works. And sincerely that isn't meant as an insult...it's just reality. Ask me about surgery and I'll shrug and say 'i got nothing yo' Everything in semiconductor is basically incomprehensible in scale. A scanner (i.e., the photo lithography projection tool) takes about 2 seconds to fully expose a wafer. that wafer probably has 100 individual chips on it each of which have around 2 billion transistors. That means that tool is helping create 100 billion transistors a second. The great grand parent comment is right...but the lose of that is probably on the range of one-two weeks production. retrofitting the fab like you suggest, and to be clear it isn't a fundamentally bad idea, would cost billions and take months if not years. What you propose is technically correct but business irrelevant. Long term here means a week or so...it's expensive but rare. They don't play with 'crude hacks' because you don't know the implications of that hack 10 years from now when the chip is running part of the stock market. It's just an industry that is hard to process, whether you work in it or not, because of the scale and the precision.
froindt 8y ago
Sorry if that came off abrasively. avs733 mentioned everything is on a different scale in semiconductors and I forgot how long it took me to grasp that. In school (industrial engineering) we talked about length and thickness tolerances being quite tight at 1/10,000 inch. The units used in semiconductors are nanometers, angstroms, etc. You might do a process and lay down just a couple atom thick layer of material...I had no idea that was physically possible.