← Back to context

Comment by DanielHB

6 hours ago

> Bear in mind thats this 4-7% loss only counts dies that have just one broken CPU unit. There are many other failure modes as well. That just seems very very high.

Is it? I thought the average for lastest-architecture chips was around 5%.

Sorry I was unclear about what "very high" meant.

From what I can see, one can expect about 80-90% yield per wafer, the bit that that doesn't make sense is that the "binned" narrative implies that of those broken parts of the wafer, 25-50% are usable with just one GPU disabled.

To me that sounds wrong, and far too high.

  • I would expect 80% of the failures would have only one core not pass QA.

    I remember back in the day it wasn't that unusual for intel to sell quad core CPUs and dual core CPUs that exactly the same hardware-wise, but the dual-core ones didn't pass the QA to be sold as a quad-core.

    In fact they sold many functional quad-core CPUs as dual-cores with 2 cores disabled and you could unlock the extra cores with some magic if you got lucky and got one that passed the quad-core QA.