At the end of last week, we saw the first pictures of the future 3D high end GPU of AMD code named R600 on a Chinese forum. We checked the source and found out that they were actually genuine:
The chip is very big but much smaller than the NVIDIA G80 (used for the GeForce 8800). According to the photos its surface would be of 430 mm² instead of 350 mm² for the R580 (X1900) and 495 mm² for the G80. Unlike those two chips which are already in shops, the R600 is using the 80 nm fabrication process to improve the density of the transistors and of the chip by more or less 20%.
This transistors' density is also variable according to the optimisations of units, quantity of cache memory (the memory transistors can be easily positioned to improve the density), etc. In consequence, it is impossible to guess precisely the number of transistors from these pictures. Nevertheless, our estimation is from 600 to 700 millions (this is approximately equivalent to the NVIDIA G80).
The picture also shows the chip position at 45 degrees and the very high number of pins. There is probably a reason for that and it could be to facilitate the connections with the pins. These practical details aren't really important. The number of pins however confirms the rumours of a memory bus of 512 bits.
It is too early to go into the details of the chip architecture, but it is possible to suppose that it will be a logical evolution of the X1000 (even if it is unified) and of the Xbox 360 chip. With the G80 and mainly the "scalar architecture" that maximises the rendering of the 64 units of texture filtering, NVIDIA stroke a decisive blow at ATI… the confrontation of the R600/G80 will probably be highly interesting…