there is no NV47, but G70 is on the way

[nutball]

Multi-chip graphics solutions exist already, you can buy them today. It doesn't take a genius to work out that they're an inevitable next step over the next couple of generations for more mainstream configurations.

Inevitable? For mainstream configurations?
If we just look at high-end gfx chips and extrapolate forward, then multi-chip look fairly logical.
However, that bird flies against the overall trend towards integration, continuosly lower ASPs for desktop systems, and the trend towards mobile computers (not only are half of all computers sold laptops, but among private consumers the proportion is even higher).

So going to multi-chip solutions may be logical, but only within an ever narrowing niche, hence my questioning that "mainstream" statement. Even now, among people who play games, only a low percentage goes for the highest performing parts. Multi-chip will never be a cheap solution due to packaging and interconnect costs. If costs are driven even higher than today, how large is the market? Mainstream?

Hehe, I knew when I typed that that someone would pick me up on it ... so I put a "more" in there. "Mainstream high-end" maybe I should have said. I mean the £300-400 market, rather than the £700-800 market.

Out of interest, how big to dies have to get before the cross-over in cost between one large die and two half the size with extra packaging & interconnect?

 

[Entropy]

Multi-chip graphics solutions exist already, you can buy them today. It doesn't take a genius to work out that they're an inevitable next step over the next couple of generations for more mainstream configurations.

Inevitable? For mainstream configurations?
If we just look at high-end gfx chips and extrapolate forward, then multi-chip look fairly logical.
However, that bird flies against the overall trend towards integration, continuosly lower ASPs for desktop systems, and the trend towards mobile computers (not only are half of all computers sold laptops, but among private consumers the proportion is even higher).

So going to multi-chip solutions may be logical, but only within an ever narrowing niche, hence my questioning that "mainstream" statement. Even now, among people who play games, only a low percentage goes for the highest performing parts. Multi-chip will never be a cheap solution due to packaging and interconnect costs. If costs are driven even higher than today, how large is the market? Mainstream?

Hehe, I knew when I typed that that someone would pick me up on it ... so I put a "more" in there. "Mainstream high-end" maybe I should have said. I mean the £300-400 market, rather than the £700-800 market.

Out of interest, how big to dies have to get before the cross-over in cost between one large die and two half the size with extra packaging & interconnect?

Hmm, I don't think that question has a straighforward answer. It seems most of the industry avoids dies over 400mm2. Yields of two 200mm2 dies will be higher, and offset some of the cost, but designing the large die to achieve good yields is possible (redundancy et cetera) and the increase in yield has to offset both increased costs associated with two complex chips and the loss of potential high-speed on-chip communication, (i.e. the functionality of the chip has to be easily splittable in two reasonably like-sized parts, that do not need the kind of communication that is best achieved on-chip).

At the end of the day, if you can fit your functionality onto one die, that is typically what you'd prefer to do, multi-die approaches probably makes most sense where the chip otherwise had become too big. Given the size of the G70, I'd say that 400mm2 is probably still a good ballpark figure.

This is not my professional field though, I'm just trying to use reason.

 

[KimB]

Out of interest, how big to dies have to get before the cross-over in cost between one large die and two half the size with extra packaging & interconnect?

Depends upon the yields and the cost of the interconnect, things which are going to be highly dependent upon manufacturing process and the architecture in question.

 

[Megadrive1988]

If I am to trust the above it sounds more to me like the "spring part" turned out to be a "summer part" (G70) and the "fall part" to turn out to be a "holiday part" (G80?). I still don't see anything beyond SM3.0 up to that point. In that case something like late summer 2006 for a WGF2.0 part could actually make sense.

ATI's timeline sounds somewhat "tight" though if I am to think R520/R580/R600 all within less than a year?

R520 = August/September

R580 = Spring 2006?

R600 = Q4 2006 at the soonest

If so, that means R520, R580, R600 introductions over a span of *more* than 1 year.