Carl, in general terms, the cost differential of producing a SDRAM and RDRAM (as well as DDR) depends on the three major elements, the cost of SI real estate (about 15% premium for RDRAM vs SDRAM and possibly DDR DRAM, the number of processing steps and the "learning curve" (the learning curve includes things like yields improvement, feature reductions etc.) The learning curve is a function of the total number of pieces produced cumulatively and that is where the "rub" is, unless the producers are "motivated" to enter mass production, they do not go down the learning curve and costs (and ASP stay up).
The SI real estate differential is really trivial since the cost of this part is in the range of 2% to 4% of total selling price of chips, but it is also the currently, the most immutable cost differential, and while the the cost might be low, the real estate differential still determines the number of dice per wafer starts, so you got to have this 15% differential there what ever production rate are accomplished.
The number of processing steps should be roughly the same (MU was known to have the least number of processing steps, but I am not sure if this is still the case), but some of the final processing steps, particularly testing of the end dices, is more expensive with RMBS (you need very costly test equipment), that cost differential has more to do with the signals "speed" (pulse rise and decay times are extremely short because of the high frequency), but that should be the same problem with DDR, so their should not be a differential between RDRAM and DDR in that respect.
So eventually (down the learning curve and after each company has under its belt 100 MM chips cum production or so), the costs differential should not really be much larger than the 15% to 30%. The ASP however is determined by supply and demand, and right now there is much more demand than supply, thus the current high price.
I hear rambling that Samsung is going to convert some DRAM production to flash because of the acute shortage (and very high ASP) in that segment. That might slow down conversion of DRAM to RDRAM and keep prices higher for longer, eventually, things do balance out and when MU see the unholy profits Samsung is making on their RDRAM, they will want a piece fo that action as well, IMHO.
I think that what you are missing is that DDR is an evolutionary development and thus has an upside limit (yes it can be pushed out marginally from time to time) on "transfer rate", the bus, because it is a "revolutionary" development needs some infrastructural elements (like chip sets, RIMMs controllers etc.) to be in place but these are being put in place. Eventually (and here is the rub, what is "eventually") these will fall in place and cost will tend toward the intrinsic differential. By the way, that intrinsic differential is not that "intrinsic" either, and I am sure a lot of outfits are working at cutting into this differential as we argue here.
One of the questions we will need to examine is what is the upper limit of the bus technology (and Si based CPU as well), since we may not be too far from reaching these limits here at the GHz range.
Zeev
PS, my GTC is at $122 not $133, we already had $133 at the April lows, and almost doubled from there to $249 or so. |
