Since GT300 will have about 500mm^2 die, which is similar to GT200, and assuming similar SP group defect rate (ie @65/55nm, 24SP occupy similar area as 32 GT300 SP @ 40nm).
GT200 launch (10 SP groups of 24):
GTX280 - 240 SP
GTX260 old - 192 SP (80%)
GTX260 new - 216 SP (90%)
RV870 (20 SP groups of 16, each with 5 execution units):
5870 - 1600
5850 - 1440 (90%)
Trend seems to be 80-90%.
For GT300, where SP are grouped by 32. Possible combinations:
512 <- difficult to get enough chips for "hard launch"
480 93.75% <- "1 less". Sounds a lot like PS3 Cell or G92 doesnt it?
448 87.5%
416 81.25% <- most likely choice for initial GT. Later can make 448SP.
384 75% <- if yields really bad. Still 60% more SP than GTX280.
I doubt nVidia will ship full out "Ultra" at launch. Trend with GF7, GF8, GF9 and GT200 is to have "half-refresh". So better to initially launch 416SP GT, and then 6month later "new" 448SP version.
However, perhaps engineers have anticipated yield issues. Perhaps whole SP group can be "salvaged" if 1 SP affected. Perhaps each group has 1 or 2 extra (So that even if 1/33 SP is bad, there will be 32 good remaining). Maybe can make groups of 30 instead of 32. Also, notice that AMD didn't change the grouping of SP. They are still sets of 16.
Ofcourse, for marketing purposes, its possible that Fermi actually has 544 or 576 SP, and that "512" versions will be sold as "perfect" product. And why not? Hard drive manufacturers have been selling hard drives with extra sectors and firmware that transparently replaces sectors with "defects" for years.
PRICE:
Putting down money on $699. All nVidia high end since GF7 have launched at $600 or higher.





Reply With Quote
Bookmarks