Originally Posted by HalfAHertz
But the GTS450 didn't have any disabled units last time i checked. So they'll try to get over 35% more performance from the same chip on the same process aaand keep the same power envelope?
I'd sure like to see that.
And who says it's the same chip? It has never been stated it's the same chip afaik and I don't think it will be the same, because GF106, the way it was, didn't make a lot of sense. I know, from our perspective, it does make sense to assume it's the same chip "fixed", because GF110 and GF104 were just that, but GF116 is quite different IMO, read below.
IMO GF106 was an experiment. I mean, the chip itself had 192 SP (and related stuff like TMU, SFU...) and a 192 bit memory interface (1:1 ratio), that didn't make any sense to me (especially since they disabled 64 bits on the actual cards), except for Nvidia trying to find out more about their Fermi architecture. To try out different combinations of SP/ROP etc. It's just a theory, but I think that once they got first GF100 silicon back, they found out something was wrong (obviously) and as an alternate solution to the problem, they might have decided to experiment a little with the cut down versions. On GF104 they increased the SP to ROP ratio compared to GF100 and in GF106 they decreased it (GF104 = 3:2 , GF100 = 4:3 , GF106 = 1:1). This way they could know which combination works best for which tasks and so they can better design future chips (i.e Kepler), which no doubt will be based on Fermi's base architecture.
What does all this mean for GF116? Well according to my own theory and assumptions, GF116 may very well be a very different animal than GF106... The higher SP-to-ROP ratio on GF104/114 clearly won in regards to gaming performance and so it makes more sense to apply a high ratio on GF116. GTX550 is going to have 128 bit again and this time around it sure doesn't have any disabled, they will just not be there to egin with. Instead they could have added another SP cluster or two* in there which would make up for the performance improvement they are talking about. ROP partitions take up more silicon space than SPs after all.
* In fact, I'm going to make a wild guess/speculation and I'm going to bet that GF116 is still some sort of experimental chip and will sport 64 SP wide clusters (4x16 SP) instead of 48SP (3x16) and 32SP (2x16) on GF104/114/106 and GF100/110, respectively. 64 SPs always made sense in my head, more than 48 anyway. Every cluster or SM has dual schedulers after all, so with 3 SIMDs, one scheduler is idling every odd cycle.
And the chip on the GTX550 has nothing to do with the one in GTX460, that's why newtekie1 said that...