55nm G92b die-shot @ PCPer

The author should say that it's approximately a 30% reduction in size instead. And yeah, that would actually be extremely close to an ideal scaling factor in both dimensions assuming linear scaling from 65 to 55nm. They had to have done some optimizing...
 
Now I really want to see 55nm GT200b and what they are gonna do to save die space and transistors. :D
 
Very impressive. One of the few times that you get the mathematically predicted result. 324 * (55/65)^2 = 231. I guess this is one of the reasons that R600->RV670 was such a drastic shrink.

Although we don't know the speed of the 4870 yet, it definately gets NVidia close in perf/mm2. The question now is how high ATI can push the clock with its next gen.
 
As for G200, I have some doubt mainly because of the bulky memory interface.
Does 55nm shrinks IO logic at all, from 65nm?
 
65->55nm is actually a 19% shrink in theory, and this was just about exactly the case for RV610 and RV630 who shrank about 17-18% to 55nm iirc, but added DX10.1 and DisplayPort. So yeah, this is a 12% optimization from G92 if true, and it actually makes G94 rather pointless. It would be very interesting to have a ~160mm² 192-bit G94-like SKU, but AFAIK they have no such thing on their roadmap.
And yes, 55nm shrinks everything, including I/O. This is unlike 80nm and 110nm actually, IIRC. And what's interesting with this is... what does it mean for GT200b?
 
That would be nice, but honestly I'm skeptical it will happen. While GT200's PCB is indeed very expensive, so would be 1024MiB of GDDR5 at this point and the PCB can't be made incredibly cheaper anyway because of the high power consumption. So I don't know if they will bother, we'll see...

BTW, I just had an interesting thought. GT200 ~1.4B transistors. If I optimize that by ~12%, I get about 1.2-1.25B, which is a number I've seen before: http://www.anandtech.com/weblog/showpost.aspx?i=453 ;)
 
Uhhh, the probability that they added GDDR5 support to G92b is approximately -inf. The die size cost just isn't worth it for that product unless maybe you made it 192-bit (and then GDDR3 would be too slow and you'd need 768MiB of memory, not viable in that segment of the market...)
 
Are their fotochoped comparison shots right? Maybe it seems the border heat spreader support should be scaled to fit each other?
 
Are their fotochoped comparison shots right? Maybe it seems the border heat spreader support should be scaled to fit each other?

It's skewed. Try this:

75636818zg7.jpg
 
Quickly checking in paint I'd say about 30%, the shape of the die seems to have changed a bit, from square to a bit rectangular.
 
Back
Top