Hi cordob; Re: "Philips makes no dram (stopped years ago), so they would not be doing this to support the ddr sales per se."
Philips is making DDR support chips because they perceive it as a way to make money. The whole industry has known that DDR was the next mainstream memory since late 1999. The only people who believe that RDRAM is going to be around 2 years from now are Rambus cultists like you. The industry believes that Intel's move to RDRAM was a mistake, and that Intel will resolve the situation as soon as they can. Therefore, industry predictions are that RDRAM production will peak in late 2000 or early 2001 and then decline. Even Samsung, Rambus' big producer, thinks that next year DDR will equal RDRAM in production. (See #reply-15905449 for the recent quote. Note that Samsung counts market share percentages in $s instead of bits, so RDRAM production numbers in bits are considerably lower.) The rest of the industry expects DDR to crush RDRAM in 2002.
You're a smart guy, and you have experience in engineering. I think the main problem you have comprehending this issue is that your engineering skills are out of date. Back in your day, high pin counts really increased package and PCB costs. This was because in order to connect a "pad" on a "die" to a "pin" on a "package", a robot (or a worker with a microscope) had to individually connect a "bonding wire" from each pad to each pin. This was slow and expensive, but with glass topped ICs it created things of great beauty, stunning under the microscope.
Rambus' big claim to fame is that it allows you to "save pins", but the fact of the matter is that there are no longer any "pins" to save. Instead, dies and chips are connected through "balls", and these are a lot cheaper than "pins". Here's an article from 4 years ago that you may find interesting:
High-density BGA package design requires tools to analyze parasitics
Charles H. Small, Computer-Design, December 1997
Ball grid array (BGA) packaging is quickly emerging as thetechnology of choice for packaging high-I/0-count ICs.
...
Providing a high ratio of I/0-to-package size, BGAs deliver higher density and yields than traditional peripheral-lead packages without requiring fine-pitch processing or new assembly equipment. Packages other than BGAs, such as QFPs (quad flat packs) and PGAs (pin grid arrays), top out when I/Os increase beyond 400 pins. In addition, high pin-count QFPs are difficult to assemble on to PC boards because they are plagued with coplanarity and solderability problems.
...
ProLinx is offering these substrates to assembly houses, allowing BGA packages to be manufactured at a projected cost that meets the market's "magic" pricing threshold of one cent per ball.
...
computer-design.com
At a cost of 1 cent per ball (a figure which has dropped even lower in the past 4 years), the price of adding an extra DDR channel to a chipset is going to be a bit above $1. But Rambus wants 2% royalties, LOL!!! That means that even if you ignore the fact that RDRAM chips use more silicon, and are therefore more expensive than DDR chips it is cheaper to build a DDR system as soon as the price of the chipset and memory exceed $1.00 / 2% = $50.
Got it? It's not about how elegant the Rambus solution is. It's a simple matter of engineers minimizing total system costs for the same performance.
Yes, I've left off a bunch of other differences between the cost of RDRAM and DDR, but most of the differences I've left off are to the disadvantage of Rambus. For example, DDR has lower latencies, RDRAM requires heat sinks, DDR is compatible with industry standard FPGAs, RDRAM is available from fewer suppliers, DDR is available on the spot market, RDRAM memory chips use a more expensive package (and one that has more pins, errrrr I mean "balls") than DDR memory chips do, RDRAM requires more expensive, and difficult to design PCB manufacturing, DDR uses less PCB area, RDRAM always provides exactly 128 bits per access and so is inefficient if what you need is some other size, DDR allows byte or word-wide write enables, RDRAM has problems with ECC coding due to its multiplexed data lines, DDR uses less power. (Did I leave anything off???)
The only advantage that RDRAM has is that the higher pin count that DDR may require (in the memory controller chip) can force an increase in the number of layers in the PCB. (This is due to "escape" PCB routing requirements, but as PCB technology improves, this difference will go away.)
The "granularity" advantage of RDRAM is illusory. Granularity is only an issue with very small memories, and the worst case for granularity is memories of a single chip. But the bandwidth out of a single chip of DDR SDRAM (for graphics applications) is higher than from even the new 1066 MHz RDRAM chips:
RDRAM: 16-bits x 1066MHz = 2.1GBytes/sec
DDR: 32-bits x 600MHz = 2.4GBytes/sec
hynix.com
Yes, I know that the above DDR chip is not a cheap mainstream DRAM chip, but hey, neither is RDRAM, LOL!!!
-- Carl
P.S. Super cool link, the Taiwan IC Exchange, shows spot pricing for various (mainstream) memory modules:
tice.com.tw |
