Yields..
vary a lot depending on what libraries are used, logic or SRAM, etc.
Pure specualtion, but we can infer somethings:
- RV670 didn't have a "4830". - this suggests yields were very good.
- RV770 both 4850 and 4870 at launch had 800 SIMD. Few "4830" and company instead opted for seperate 40nm die. Later revision clocks all the way to 1Ghz. - this suggests yields were spectacular
- RV870.. notice for first time in many years the "050" part has cut down shaders etc. Even though die shrink, clockspeed about same. - this suggests yields are poor
Now lets look at nVidia:
- G80. Right away we see GTS variants. Suggests many defects on ROP or MEMCTRL.
- G92. Launced as 112 shader 8800GT. Suggests yields quite poor initially. Couldn't even launch "full" part. But, later we get GTX, and GTX+ and only GTS250 on market. No cut down versions. This indicates yields improved a lot.
- G200. Once again, cut down GTX260 and GTX275 versions. Notice how they "improved" the shaders on GTX260 to 216... yields gradually improving, but because of huge die, many sub-perfect parts.
GT300.
We know 3B transistors. We know 384bit. We know 512 CUDA cores. Thats it. Rumours about yields are just rumours.
But, we do know 3B transistor chip will be bigger than RV870. And AMD launch suggests they are having yield problems. This suggests that *if* and *when* GT300 launches, it might be like G92 launch.
To harvest more chips with defects, 'GTX' could launch with 480 SP, and the GT with 416 SP. Later, after several months, a "Ultra" can be launched with full 512 SP.
Dont be surprised if this happens. They did it with 7800GTX/7800GT. 8800GTX/8800GTS, and ofcourse GTX280/GT260.





Reply With Quote

Bookmarks