| Pemalite said: Well, technically, the design is already "done" with AMD's 28nm Oland core, unfortunatly it still doesn't get as cheap as the 40nm Cedar and Caicos Chips, but that will change eventually... A possible respin with the removal of the GDDR5 memory controller might be a possibility? Then again, Cedar and Caicos has DDR2, DDR3 and GDDR5 controllers, so it can't be that expensive transister wise. AMD tends to only shop at Global Foundries and TSMC. - Should be interesting to see the ramifications of Apple buying up capacity from TSMC in regards to nVidia's and AMD's 20nm GPU lines. |
Although the design maybe finished you still need to port it to a newer process node ... (That costs money.)
Actually I don't believe that cost reductions will come from manufacturing the chips with smaller feature sizes as we're forced to use double patterning which exponentially increases the wafer cost. The only way I can see a cost reduction for an XB1 unit is that with a smaller chip comes with a smaller power consumption therefore we can use a cheaper power supply and fan to drive down the costs but these gains are marginal at best and could be counteracted by the fact that it can be potentially more expensive to produce the chip at a newer process node ...
I think Apple will have to eventually switch to Samsung cause they have better chip scaling with their 14nm FinFETs compared to TSMC's 16nm FinFETs so I believe Nvidia can get a lot of TSMC's fab capacity easily and the same goes for AMD too but I think they might consider Global Foundries for their GPUs in the future since these guys licensed Samsung's 14nm node which will give AMD a possible leverage on Nvidia in chip scaling department ...
