At the time the sweet spot strategy was a fig leaf to cover that ATI no longer had the dollars, to develop a really large gpu. It caused problems at first as it allowed their competitor free reign and quite good profits from temporariy having no competition.
No, the strategy was in place before R600 arrived, hence RV670 was small and the rumours of the X2 version were part and parcel of the expectation of its arrival. It was only ~190mm² (perhaps <180mm² if you allow 0.5mm of packaging margin). G92, its theoretical competitor, is 256mm² on the same process (240mm² if allowing for packaging margin?). You could say RV670 was actually competing with G94, not sure what die size that ended-up at on 55nm.
Back in 2004 R420 was 281mm², half GT200 size in 2008. No wonder the prices of the halo cards have gone up.
since then, like sand in the desert, the millions of customers shifted so that today the landscape looks completely different to the one of a few years ago. Now there is just not enough grains left to cover the cost of developing the really huge gpus.
Huge GPUs also predicate a significant testing period on a process if you want to be assured of a successful launch - whoops R520 and hence NVidia's process reticence all these years.
So if you want to adopt a new process rapidly with a huge GPU that launches first then the effort naturally hinders the rest of the chips. Just cutting out the 400mm²+ chip saves a lot of engineering time and cost.
R5xx GPUs were supposed to be a top-to-bottom launch over ~5-6 months (~May-October). That was 3 GPUs.
This would be quite bleak, if not for another development over the last year or two. The design process appears to have got much more automated such that can scale the different parts of their design up and down very rapidly to cover all parts of the market.
AMD simply not doing a 400mm²+ GPU leaves resources for the other parts. In theory the ring bus was part of making it easier to scale a GPU, since it's not just the count of units that needs scaling, but the connectivity.
(I can't help thinking that a ring bus might return... Larrabee has one.)
Last year AMD produced 3 chips and a dual card within 3 months to overhaul their complete lineup. Similarly nvidia was pursuing the same course with 4 chips top to bottom: GT212->GT218. Nvidia had some issues with the top 2, some combination of lack of market, price/performance over existing offerings or technical difficulties(ie excessive leakage in their larger parts).
GT200b took 3 revisions on a "mature" 55nm process.
NVidia's cancelled G100, GT206 (though did that ever exist?), GT212 and GT214. I'm now wondering if NVidia has cancelled the original GT300 chip, because the conflicting tape-out rumours coupled with a chip of <<500mm² conflict with a chip that's just taped-out and is >500mm². Maybe the GT300 that gets launched is another re-think like G100->GT200?
Jawed
