3870 to 4870 was literally a quantity increase. Cypress to Northern Islands is much more. There's obviously going to be more shaders, probably a higher clock speed too, and then an improved "rest of the core". As I keep saying, they're reworking how the rest of the core works.
THe 3870 to 4870 went up 30% in die space, yet got what, 2.5 times the shaders, the performance improvement came from that, and it happened because they essentially cut out a huge amount of waste, they essentially cut the 3870 back to an incredibly slimline chip that would be 60% of so of the original size, then bumped up the shader count 2.5times, it is essentially a doubling in die size without any of the waste.
I'm telling you right now, Nvidia still hasn't provided a single 512sp chip to retail coming up on a year after it was supposed to launch, AMD are not making a 4billion transistor chip, they just aren't.
GF100 to GF104, essentially, cut the die size by almost 33%, yet lost next to no performance, thats essentially what the 3870 did before it then increased the shader count.
THe problem is theres very little waste in a 5870, its incredibly well made and efficient, thats how a 5850 is up to 30% faster than a 460gtx, yet is over 10% smaller.
While the 5870-6870 is a large increase in the style of the uncore(rops, tmu etc) theres not much more to shave off, its more a balance/efficiency situation. 3870 to 4870 was a FAR bigger increase in architecture, ring bus back to a crossbar is one of the biggest structural changes in a chip there can be, it effected how EVERY single part of the core is connected, thats transistors on every single shader, every rop, every mem controller, every step of the way it was increased transistor count.
A rop is pretty much a rop, and a tmu is pretty much a tmu, moving stuff around could increase transistor count but increase efficiency to make it worthwhile, it might decrease transistor count, but its unlikely to do it much.
THeres simply not 25% of the core to be shaved off this time around, to help bump the shader count/rops/tmu's up like 3870-4870.
The GF100 had a lot of waste to be shaved off, the 5870, just doesn't.
As to why would they release it, they are a company that makes cards, no new cards, no company. Prescott wasn't a massive leap forward for Intel, Phenom 1/2 weren't a massive leap forward, Penryn wasn't a massive leap forward, Bulldozer and Sandybridge won't bump performance by 100%, won't stop people buying them or upgrading.
Also an overclocked Cypress, well, a 1Ghz cypress with 1.3v would, as now use a lot of power, be hotter, need better cooling to be quiet and isn't really a good design.
The thing is, if they can produce something 30% faster, at the same clock speeds, that still overclocks to the same clocks as a 5870 can, then it will end up being about 40-45% faster.
Not everyone has a 5870 anyway, infact about 4billion people don't, of those that don't, several think the 5870 isn't fast enough, but a 5870 +40% would be fast enough to upgrade to. Faster = better, even more so if its at the same price point as before.
What's so different now, compared to the last 4 generations of cards?
