By using this site, you agree to our Privacy Policy and our Terms of Use. Close
Squilliam on 08 August 2010
Edit Thread
Soleron said:
Squilliam said:
...

The GTX 460 is a ~330mm^2 chip. Charlie was wrong btw. Since it came out later you can assume it'll be better tuned to the 40nm process than Cypress so it may infact have better yields. The reason why they are all salvage parts is because Nvidia doesn't want to kill their own GTX 470 and 480 line like they did when they released the 8800GT. In any case you could say that the pricing it launched at is already a response to Southern Islands, thats a pretty good explanation in itself.

TBH Im not sure what Southern/Northern Islands will be. From what I think I know, it'll have a revamped 4P shader architecture and adjusted ROPs etc.

Charlie had it professionally measured to a tenth of a mm on a side.

And it is not yielding better, since it doesn't use double-vias to compensate for 40nm's leakage. Even Anandtech agrees with Charlie that Nvidia should have done that.

SI will be tweaked Evergreen shaders with a new frontend that will hugely increase efficiency. Notice how poor the 3870 -> 4870 -> 5870 scaling was with shader count. Nowhere near 5x, more like 3x. SI will fix some of that.

Nvidia will release a full GF104 or dual GF104 when they run out of GF100 supply in a few months. Because they only ever produced a limited number of wafers and are working through packaging them.

But how do you know it doesn't have double vias? Im certain that the RV740 didn't have them and yet lessons learnt were applied to Evergreen which featured them. From what I am aware, (correct me if im wrong) as I have been following this topic as well is that Charlies sources were off on their measurements because I saw a cut down GT 460 die measured to roughly 330mm^2 using accurate instruments. The rough part comes from the fact that the die package was damaged slightly when they cut the heat spreader. However this is relatively inconsequential.

I do have to agree that Southern Islands could be quite powerful. Especially as they can take advantage of faster and larger 2Gbit GDDR5 32bit wide ram chips. This would solve their bandwidth and especially frame buffer issues whilst at the same time allow them to take advantage of better shader utilization and front end architecture.

I also have to agree that they are only really producing Fermi chips for the professional market. Its pretty much the same deal as the EOL of their last range of high end cards as well.



Tease.