Return to Answer

The initial versions of the CPU and GPU above were over 200 mm², which is quite large. Conjecture: making them initially a single chip would have resulted in substantially diminished yield.

Being already at the upper end what can be done

A Pentium III (Coppermine) of that time had about 10M transistors and ~100 mm², so a die with more than 200 mm² was most definite at the upper end of what could be done with reasonable expectation of success. Doubling this might have been past what could have done at that time.

Increasing absolute yield rate

Chip fault rate goes not linearly, but exponentially with size. But already when assuming a constant defect rate per wafer, doubling the size will double the fault rate as faults are now be distributed among half the dies.

Example:

• Let's assume two dies of 250 mm² each or one die of 500 mm²
• If a wafer offers 25.000 mm², it gives 100/50 dies (*1)
• If the process has an average of 10 defects, then 10 dies are dead (*2)
• For the 250 mm² die this gives 90% yield
• For the 500 mm² die this gives 80% yield

Doing two wafers, one with each, will result in 90+90 good chips enabling the build of 90 machines, while doing the same two with double sized chips gives 40+40 good chips for 80 machines. It's easy to see that doing multiple chips will increase yield and decrease cost per chip.

Was that the reason, or was there another factor?

Reducing development complexity

It resources in development to do multiple chips, as that not only reduces complexity a lot but also allows independent schedules for both. This may sound contra intuitive at first, as they both have to finish before the machine can be build. But it allows each project to advance at their own pace for interim milestones while a common die would need to interlink every step and iteration. A classic problem of complex developments. When looking at development like production, it's much like before: Yield increase by reduced dependence.

*1 - Simplified by ignoring cut areas and alike

*2 - Simplified by assuming even distribution

The initial versions of the CPU and GPU above were over 200 mm², which is quite large. Conjecture: making them initially a single chip would have resulted in substantially diminished yield.

Being already at the upper end what can be done

A Pentium III (Coppermine) of that time had about 10M transistors and ~100 mm², so a die with more than 200 mm² was most definite at the upper end of what could be done with reasonable expectation of success. Doubling this might have been past what could have done at that time.

Increasing absolute yield rate

Chip fault rate goes not linearly, but exponentially with size. But already when assuming a constant defect rate per wafer, doubling the size will double the fault rate as faults are now be distributed among half the dies.

Example:

• Let's assume two dies of 250 mm² each or one die of 500 mm²
• If a wafer offers 25.000 mm², it gives 100/50 dies (*1)
• If the process has an average of 10 defects, then 10 dies are dead (*2)
• For the 250 mm² die this gives 90% yield
• For the 500 mm² die this gives 80% yield

Doing two wafers, one with each, will result in 90+90 good chips enabling the build of 90 machines, while doing the same two with double sized chips gives 40+40 good chips for 80 machines. It's easy to see that doing multiple chips will increase yield and decrease cost per chip.

Was that the reason, or was there another factor?

Reducing development complexity

It saves resources in development to do multiple chips, as that not only reduces complexity a lot but also allows independent schedules for both. This may sound counter-intuitive at first, as they both have to finish before the machine can be built. But it allows each project to advance at their own pace for interim milestones while a common die would need to interlink every step and iteration. A classic problem of complex developments. When looking at development like production, it's much like before: yield is increased by reducing dependence.

*1 - Simplified by ignoring cut areas and the like.

*2 - Simplified by assuming even distribution.

The initial versions of the CPU and GPU above were over 200 mm², which is quite large. Conjecture: making them initially a single chip would have resulted in substantially diminished yield.

##Being already at the upper end what can be done

A Pentium III (Coppermine) of that time had about 10M transistors and ~100 mm², so a die with more than 200 mm² was most definite at the upper end of what could be done with reasonable expectation of success. Doubling this might have been past what could have done at that time.

##Increasing absolute yield rate

Chip fault rate goes not linearly, but exponentially with size. But already when assuming a constant defect rate per wafer, doubling the size will double the fault rate as faults are now be distributed among half the dies.

Example:

• Let's assume two dies of 250 mm² each or one die of 500 mm²
• If a wafer offers 25.000 mm², it gives 100/50 dies (*1)
• If the process has an average of 10 defects, then 10 dies are dead (*2)
• For the 250 mm² die this gives 90% yield
• For the 500 mm² die this gives 80% yield

Doing two wafers, one with each, will result in 90+90 good chips enabling the build of 90 machines, while doing the same two with double sized chips gives 40+40 good chips for 80 machines. It's easy to see that doing multiple chips will increase yield and decrease cost per chip.

Was that the reason, or was there another factor?

##Reducing development complexity

It resources in development to do multiple chips, as that not only reduces complexity a lot but also allows independent schedules for both. This may sound contra intuitive at first, as they both have to finish before the machine can be build. But it allows each project to advance at their own pace for interim milestones while a common die would need to interlink every step and iteration. A classic problem of complex developments. When looking at development like production, it's much like before: Yield increase by reduced dependence.

*1 - Simplified by ignoring cut areas and alike

*2 - Simplified by assuming even distribution

The initial versions of the CPU and GPU above were over 200 mm², which is quite large. Conjecture: making them initially a single chip would have resulted in substantially diminished yield.

Being already at the upper end what can be done

A Pentium III (Coppermine) of that time had about 10M transistors and ~100 mm², so a die with more than 200 mm² was most definite at the upper end of what could be done with reasonable expectation of success. Doubling this might have been past what could have done at that time.

Increasing absolute yield rate

Chip fault rate goes not linearly, but exponentially with size. But already when assuming a constant defect rate per wafer, doubling the size will double the fault rate as faults are now be distributed among half the dies.

Example:

• Let's assume two dies of 250 mm² each or one die of 500 mm²
• If a wafer offers 25.000 mm², it gives 100/50 dies (*1)
• If the process has an average of 10 defects, then 10 dies are dead (*2)
• For the 250 mm² die this gives 90% yield
• For the 500 mm² die this gives 80% yield

Doing two wafers, one with each, will result in 90+90 good chips enabling the build of 90 machines, while doing the same two with double sized chips gives 40+40 good chips for 80 machines. It's easy to see that doing multiple chips will increase yield and decrease cost per chip.

Was that the reason, or was there another factor?

Reducing development complexity

It resources in development to do multiple chips, as that not only reduces complexity a lot but also allows independent schedules for both. This may sound contra intuitive at first, as they both have to finish before the machine can be build. But it allows each project to advance at their own pace for interim milestones while a common die would need to interlink every step and iteration. A classic problem of complex developments. When looking at development like production, it's much like before: Yield increase by reduced dependence.

*1 - Simplified by ignoring cut areas and alike

*2 - Simplified by assuming even distribution

The initial versions of the CPU and GPU above were over 200 mm^2, which is quite large. Conjecture: making them initially a single chip would have resulted in substantially diminished yield.

##Being already at the upper end what can be done

A Pentium III (Coppermine) of that time had about 10M transistors and ~100 mm^2, so a die with more than 200 mm^2 was most definite at the upper end of what could be done with reasonable expectation of success. Doubling this might have been past what could have done at that time.

##Increasing absolute yield rate

Chip fault rate goes not linear, but exponential with size. But already when assuming a constant defect rate per wafer, doubling the size will double the fault rate as faults are now be distributed among halve the dies.

Example:

• lets assume two dies of 250 mm^2 each or one die of 500 mm^2
• If a wafer offers 25.000 mm^2, it gives 100/50 dies (*1)
• If the process has an average of 10 defects, then 10 dies are dead (*1)
• For the 250 mm^2 die this gives 90% yield
• For the 500 mm^2 die this gives 80% yield

Doing two wavers, one with each will result in 90+90 good chips enabling the build of 90 machines, while doing the same two with double sized chips gives 40+40 good chips for 80 machiens. Easy to see that doing multiple chips will increase yield and decrease cost per chip.

Was that the reason, or was there another factor?

##Reducing development complexity

It as well saves resources in development to do multiple chips, as that reduces not only complexity a lot but as well allowing independent schedules for both. This may sound contra intuitive at first, as they both have to finish before the machine can be build. But it allows each project to advance at their own pace for inbetween milestones while a common die would need to interlink every step and iteration. A classic problem of complex developments. When looking at development like production, it's much like before: Yield increase by reduced dependence.

*1 - Simplified by ignoring cut areas and alike

*2 - Simplified by assuming even distribution

The initial versions of the CPU and GPU above were over 200 mm², which is quite large. Conjecture: making them initially a single chip would have resulted in substantially diminished yield.

##Being already at the upper end what can be done

A Pentium III (Coppermine) of that time had about 10M transistors and ~100 mm², so a die with more than 200 mm² was most definite at the upper end of what could be done with reasonable expectation of success. Doubling this might have been past what could have done at that time.

##Increasing absolute yield rate

Chip fault rate goes not linearly, but exponentially with size. But already when assuming a constant defect rate per wafer, doubling the size will double the fault rate as faults are now be distributed among half the dies.

Example:

• Let's assume two dies of 250 mm² each or one die of 500 mm²
• If a wafer offers 25.000 mm², it gives 100/50 dies (*1)
• If the process has an average of 10 defects, then 10 dies are dead (*2)
• For the 250 mm² die this gives 90% yield
• For the 500 mm² die this gives 80% yield

Doing two wafers, one with each, will result in 90+90 good chips enabling the build of 90 machines, while doing the same two with double sized chips gives 40+40 good chips for 80 machines. It's easy to see that doing multiple chips will increase yield and decrease cost per chip.

Was that the reason, or was there another factor?

##Reducing development complexity

It resources in development to do multiple chips, as that not only reduces complexity a lot but also allows independent schedules for both. This may sound contra intuitive at first, as they both have to finish before the machine can be build. But it allows each project to advance at their own pace for interim milestones while a common die would need to interlink every step and iteration. A classic problem of complex developments. When looking at development like production, it's much like before: Yield increase by reduced dependence.

*1 - Simplified by ignoring cut areas and alike

*2 - Simplified by assuming even distribution