The site explains TSMC's 40nm process for GPUs will allow them to put 2.35x as much transistors per square millimeter than on the 65nm process. Furthermore TSMC claims a performance improvement of 60 percent, if you add this up you get an increase of 3.76x and the site argues that with some additional performance improvements by the GPU makers you might get GPUs that pack four times as much performance than the current RV770 and GT200 generation.
The site put together some specifications of what's theoretically possible with the 40nm process but says it's very likely the next-generation GPUs will have totally different specifications than what they've come up with. Theoretically, the 40nm process may enable AMD to create a RV870 with 2.2 billion transistors, 1.2GHz clockspeed, 1800 shaders, 90 texture units and 32 ROPs while NVIDIA could use it to make a GT300 with 3.2 billion transistors, 960/2100MHz clockspeeds, 500 shaders, 180 texture units and 64 ROPs.
Alternatively, the GPU makers could take advantage of the 40nm process to create much smaller chips. The first 40nm GPUs are expected in Q1 2009 and by the end of 2009 TSMC is expected to offer a 32nm process for GPUs.