gg,
The marketing goons at AMD didn't realize that the way to press
K8's advantage was to price Opteron 8xxx the same as 2xxx and
press motherboard manufacturers to build cheap 4S infrastructure.
That would have been disruptive and 90% of the server market
would have moved over to 4 die AMD servers. Instead, it was Intel
that first moved the market to low-price 4 die servers. How ironic! AMD
had a far superior infrastructure to connect 4 dies together.
Yet they stupidly chose to price the dies in such a way that
AMD 4 die servers remained a small niche.
I have to agree here. Selling 2x the number of server dies in a server is enough of a premium for the user to pay, IMO.
OTOH that is only a temporary step to what I think an ideal server chip would be, which is a one (huge) die server with on die memory controller(s), several memory channels, multiple cores, L3, and most importantly, no memory coherency trafic overhead.
The cost of the server would shift from expensive mobos and chipset to an expensive silicon. IMO, a dollar spend on a silicon die area (cost-wise) has to have bang for the buck of one or more orders of magnitude over cost of designing complex motherboards.
Adding cores is now just a cookie cutter type of thing now. AMD has the inside the chip communication worked out very well, well positioned to adding cores. Intel's Conroe approach to shared L2 does not scale well, but my guess is that Intel will revise it for Nehalem, probably closer to Barcelona approach.
Anyway, next thing to work out the redundancy of cores, meaning you would make say 10 core die to get 8 core chip, or 5 core die to make 4 core chip.
Intel is already capable of making large chips (with Itanium), so I think this whole thing of one die servers is very possible.
The problem is the economic one. Looking at who would benefit from this, clearly not server OEMs, since it would not take a whole lot of expertise to make these, so they would be commodity that today's 1S servers are. CPU manufacturers would be able to sell more expensive chips, but fewer of them, so they would not necessarily benefit. Customers are the only ones who would benefit, being able to get a lot more bang for the buck, which unfortunately puts the whole thing in doubt...
Joe |
