It seems as though 40nm is here to stay for a long time. The 40nm first graphics card introduced was the ATI Radeon HD 4770, way back in April 2009. Unfortunately, the 40nm TSMC process was plagued with a myriad of issues, most notable a “chamber mismatch” problem, they led to poor yields for the best part of 2009. The Radeon HD 5800 series released in September 2009, but experienced massively crippled supply till December 2009.

The worst affected by TSMC’s troubled 40nm process has to be NVIDIA’s GF100. Over 6 months late, far too hot and well short of target specifications/performance. These problems with 40nm led TSMC to cancel the 32nm process, which was to be ready in 2010. The 28nm process, first planned for Q4 2010, kept being pushed back, and now Fudzilla reports that both TSMC and Globalfoundries will only be ready for mass production of 28nm GPUs in late 2011.

More next page…


It seems as though 40nm is here to stay for a long time. The 40nm first
graphics card introduced was the ATI Radeon HD 4770, way back in April
2009. Unfortunately, the 40nm TSMC process was plagued with a myriad of
issues, most notable a “chamber mismatch” problem, they led to poor
yields for the best part of 2009. The Radeon HD 5800 series released in
September 2009, but experienced massively crippled supply till December
2009. The worst affected by TSMC’s troubled 40nm process has to be
Nvidia’s GF100. Over 6 months late, far too hot and well short of target
specifications/performance. These problems with 40nm led TSMC to cancel
the 32nm process, which was to be ready in 2010. The 28nm process,
first planned for Q4 2010, kept being pushed back, and now Fudzilla
reports that both TSMC and Globalfoundries will only be ready for mass
production of 28nm GPUs in late 2011.

AMD GPG realized this problem early on, and designed an intermediate generation to replace the next-gen N. Islands, initially planned at 32nm. The solution is S. Islands, a transitional hybrid between N. Islands and Evergreen (possibly more Evergreen than N. Islands). N. Islands now has been pushed back to 28nm. Considering S. Islands / HD 6000 releases in Q4 2010, AMD have got the delay well covered.

Things are a bit more complicated for Nvidia. Like GT200, GF100 screams for a die shrink. However, unlike GT200, it doesn’t have a superior process waiting. It is more than a year away. Nvidia have done the right thing with GF104 – make a die that aims for a better performance/die-size/watt ratio suitable for 40nm. Nvidia will continue filling up the Geforce 400 series for the rest of 2010. The problem is that the Nvidia’s 2010 products are only good enough to compete with AMD’s 2009 products. Once S. Islands releases, Nvidia will have to compete with the effectively one generation behind Geforce 400 products for a long time. However, it is possible to make massive advancements using the same process. HD 3800 to HD 4800 was by all means a shocking success – with both on 55nm. AMD managed to fit 250% more shaders in 30% more die space.

Certain rumours have also appeared that AMD might try out Globalfoundries’ 32nm SOI process. While GPUs are manufactured on Bulk process, AMD have had experience putting a GPU on 32nm SOI – as in Llano. Rumours also suggest that AMD is likely to move to Globalfoundries in favour of TSMC if they have an equivalent process ready.

Clearly, TMSC’s problematic 40nm Bulk process has greatly affected the development of GPUs. We may have to wait for more than a year for the next great GPU, but we hope AMD and Nvidia continue to innovate despite the constraints.

Reference: Fudzilla