Here is additional info for interested laymen in understanding the critical yield per wafer data...
A new report says TSMC's 3 nm node has yields of just 55%. On the other hand, Samsung Foundry is faring much better at 60%. AMD is rumoured to use Samsung Foundry for some of its upcoming parts.
www.notebookcheck.net
A previous report stated
Samsung Foundry was doing quite well with its 3 nm node with 60% yields.
Mapple_Gold, an industry insider, said it went
as high as 75% in one instance.
This is normal. In an industry standard batch of 25 wafers, in a development process, there is always one wafer that will produce the highest yield of functional dies. So in this case, Samsung experienced a 3nm wafer with %75 functional dies. It is a plus for Samsung but in terms of development, every manufacturer have this accomplishment with every design.
The Taiwan-based chipmaker's yields currently stand at 55%. While it doesn't mention the N3 variant in question, it is likely N3B, the first iteration. TSMC's decision to stick with a FinFET design may have something to do with the sub-par yields.
On the other hand, Samsung's 3 nm node uses the newer GAA FET technology,...
This means that while TSMC is capable of producing dies at 3nm, TSMC have a lower yield per wafer than Samsung does, and it also means that TSMC do not have the same level of manufacturing expertise as Samsung does. If customers have only the choices of Samsung and TSMC at 3nm, they will buy from Samsung over TSMC. This is what the yield per wafer data produces at the market level. A truly matured manufacturing process would produce at %95+ yield per wafer. If all manufacturers encounters the same yield problems, at 10nm or 7nm or 2nm or whatever node, then that yield per wafer data will become industry standard just for that node level and industry expectations for previous nodes will
NOT change.
Despite higher yields, Samsung Foundry is yet to attract customers.
This is normal. The 3nm node is still majority experimental in terms of scaling down and in terms of equipment required to descend to that level. The yield per wafer data is still too scant for the industry to tell customers that what is acceptable. This does not mean no customers will buy from Samsung at that yield per wafer data. Any customer willing to take the chance, at less than %95, will perform their own testing which is likely to include even more strenuous criteria. If AMD buy 1000 3nm dies from Samsung, by the time AMD is done their own testing, most likely the final functional dies remains will be 1/2. AMD has to make sure because once they packaged and sell, their reputation is at stake. Whatever remains has to work.
So going back to the original article...
One of the research firms said it expected Huawei to use SMIC's N+1 manufacturing process, though with a forecast yield rate of usable chips below 50%, 5G shipments would be limited to around 2 million to 4 million units.
The word 'forecast' cannot be simply dismissed as something the West will do to disparage Chinese manufacturers, in this case SMIC. Industry research firms made their predictions based on known technologies, manufacturing outputs, and technical data provided by the manufacturers themselves. Even if we generously add %10 to that yield forecast, that would still be bad news for SMIC and Huawei because other 5G technology manufacturers who are not burdened by US tech sanctions
WILL have a higher yield per wafer than SMIC can produce, and customers will bid for those other manufacturers.
