Hi all; The latest EE-Times news:
The big guys are getting ready (through the unpleasant mechanism of competition) to take the profit out of the server arena. Read the following two articles closely:
Intel Tips Eight-Way Server Plans
Intel's plans for segmenting its Pentium II line for various retail to large-server markets is a fundamental part of its strategy to maintain average selling prices on its microprocessors, despite steep price cuts in desktop markets
techweb.com
Bus war breaks out in PC server realm
At stake: a strategic advantage in the PC server market, which has become one of the last bastions of significant profits and technical differentiation in mainstream computing.
techweb.com
Intel's decision to use rambus is now coming back to haunt them. The single advantage that rambus has over other DRAM is the high bandwidth in a small number of pins. That advantage has been largely eliminated, for small systems, by the combination of the BGA package, (a cheap IC package that allows incredibly huge numbers of pins) and on-chip DRAM, (which elminates the DRAM interface pins compeletey, and provides massive bandwidth and short access times, as well as reducing system costs by increasing integration.) The first provides a much cheaper alternative in designs that are core-limited. The second provides a much cheaper alternative in designs that are pad-limited. Between these two technologies, rambus will be squeezed out of competitive markets. Here's the cost add-on for choosing rambus over SDRAM:
... supply crunch as they race toward transition to Rambus DRAM
"We've been talking about the die-size adder and packaging and testing costs, and building the RIMMs with a heat sink, plus tighter tolerances for pc boards. It does look like a 40 to 60 percent cost adder by the time all is said and done," Mailloux said.
techweb.com
Incidentally, when Intel announced their intention to support only rambus on future architectures, I posted the following to SI (August 1997):
Unless you want to take the chance of making a chip that will only work with a memory type that is unavailable, you never, ever, ever design a processor that runs with only a single type of RAM. ... Designing only RAMBUS compatible processors would be an incredible gift to the competition, and isn't going to happen.
Message 2082162
The funny thing is that it did happen. I was wrong. (About a technical subject, no less!) Intel did design processors compatible only with rambus. This is going to haunt them big time over the next couple years, as their competitors eat their lunch in the server processor market, as well as the low-end market where they have already had their lunch eaten. Maybe Dell shouldn't tie itself too closely to the Intel boat...
-- Carl
P.S. About "pad-limited" and "core-limited." Engineers try to minimize the silicon die area that a design requires. There are generally two requirements. First, that the die have a large enough area to do all the core logic. Second, that the perimeter of the chip be long enough to connect to all the pads (or pins). For any given design, one or the other of these constraints typically define the smallest die size that will work. For a given process, the cost of chips is typically somewhat more than proportional to the die size. By "pad-limited" I mean a design where it is relatively cheap to add on more core logic. By "core-limited" I mean a design where it is relatively cheap to add on more pads. |
