Originally Posted by 2010rig
True, but with GM204 being 398mm^2, they can pack another 25% worth of SM's... We shall wait and see
They will start out conservative. They have to leave room for a future product and the faster these perform, the less room for the future. I could see these chips being 300mm to 350mm2. If they are stuck another 4 years at this node again, they will be their biggest competition I have a feeling. Unless there is a remarkable improvement in AMD finances, they are unlikely to be able to afford a double architecture launch like Nvidia and I think we will see rebrands for AMD 16nm chips as well. Tape out are simply too expensive and nvidia too will try to milk the next node.
The simplification of the memory controller should allow nvidia to allocate another 20% more shaders to other components like shaders, alu and ROPs. AMD was able to shave what could have been a double sized or 700mm chip based on the double the component transition from fiji to tonga( 360mm chip size) down to 600mm2. This means AMD was able to allocate or shrink their die 16.7% from the HBM memory controller simplification. Since Nvidia's memory controller is larger than AMD's generally so if these transition over to nvidia chip, I could see about a 20% savings in die area. With the double packing of the transistor density from going to 16nm, a 300mm2 chip could fit 20% more components and a 350mm2 die could fit about 40% more components. Thus for a gm204 part, 400mm2 die would hardly be necessary. Such a die i am guessing could fit about 60% more components. At this point, it could very well turn into be gp210.
A 400mm2 die with HBM2 at 16nm is atleast the same cost as a 600mm 28nm die and is likely more expensive to make. I don't think Nvidia is going to release 600mm2 any time soon(next 2 years). AMD is struggling with a very mature 28nm at making such chips. Although Nvidia has more experience designing monolithic chips, it doesn't make sense to make such monster dies when the cost for 16nm finfet chips are so prohibitive due to the wafer cost doubling.