After canceling the 32nm process, TSMC is now rumoured to skip 22nm and move straight to 20nm in H2 2012. Despite the poor 40nm process, TSMC is over-burdened with orders, including Nvidia and ATI GPUs. Initially, TSMC was expected to move to 32nm some time in 2010, thus enabling refresh/next-generation GPUs in late 201-. However, TSMC’s move to cancel 32nm has meant the next advancement will be the 28nm, which is now rumoured to have slipped to H1 2011.

This has caused rescheduling of product releases for AMD especially, and perhaps Nvidia.

More next page…


After canceling the 32nm process, TSMC is now rumoured to skip 22nm and
move straight to 20nm in H2 2012. Despite the poor 40nm process, TSMC is
over-burdened with orders, including Nvidia and ATI GPUs. Initially,
TSMC was expected to move to 32nm some time in 2010, thus enabling
refresh/next-generation GPUs in late 201-. However, TSMC’s move to
cancel 32nm has meant the next advancement will be the 28nm, which is
now rumoured to have slipped to H1 2011.

This has caused rescheduling of product releases for AMD especially, and
perhaps Nvidia.

It is clear that neither AMD nor Nvidia are satisfied with the troubled 40nm process. AMD especially are unlikely to be content with the 32nm cancellation as the next-generation Northern Islands GPUs were set for release on 32nm in H2 2010. The current rumours suggest that AMD will have to do a half-generation Southern Islands in 40nm, before it can move to 28nm in 2011 for the “proper next-gen” N. Islands. For Nvidia, this is bad news as well, as the GF100 is desperately screaming for a
die shrink – one which it cannot get before 2011. So, Nvidia may be
forced to do a base layer respin at 40nm instead.

Now, it seems TSMC is canceling a process yet again, perhaps further disturbing release schedules. 22nm, which was last scheduled for late 2011, will not see the light of the day, as TSMC will move straight to 20nm. With 20nm due in late 2012, this would once again mean a massive gap between process transitions with no half-node. Effectively, this could kill the popular policy of die shrink refreshes we have seen in the past. 40nm was first introduced in 2009. 28nm comes next some time in 2011, and 20nm late 2012. Clearly, the gap is beyond one year – maybe even larger than a generational gap. So, GPU makers might have to change their approach to revisions on the same process rather than die shrink refreshes – as AMD are doing with Southern Islands. One could argue AMD started this approach with the HD 4000 series, where only one product (4770) got the die-shrink treatment.

Of course, with Globalfoundries starting bulk production of 28nm in 2011, it is likely that at least AMD will never get to order any significant quantity of 20nm TSMC wafers. If TSMC fails to deliver with 28nm it is likely Nvidia would also be considering jumping ship to Globalfoundries.

Reference: Xbitlabs