Hi John Walliker; Re: Large numbers of pins inherently mean large capacitances to charge and discharge, resulting in high power consumption and electrical noise generation.
This statement is true with respect to pins on a PCB, but for pins on the new direct chip to chip packages, we are talking about a whole new ball game, much lower capacitances, inductances &c.
Re: When data transfers are taking place at multi-Gbit per second per pin rates the module will be designed according to microwave practice.
This is true only if the wire length were held constant. Since the wire length is being shortened at the same time as the frequencies are increased, the design will be simple. The link I gave to the Triquint clock termination application notes explain how the design of terminations work. It is all about scaling.
Re: There will be new packages that bring signals on and off each chip through constant impedance transmission line structures.
What I am saying is that essentially all the high bandwidth signals are going to remain in the package, and never see a PCB. This has been the trend in high frequency parts for some time. Example, instead of having to route a 1GHz clock around, for instance, you route 100MHz, and let each chip frequency multiply. Look at the FMM4006KC part from Fujitsu, for instance. (The spec is here on my desk.) The reason for this tendency is to avoid having to build complicated, delicate, power-hungry circuits to send signals around at high speed. Even with Rambus's "solution", this is difficult.
Re: As data rates increase, lumped component methodology is doomed to fail, however much you would like to believe otherwise.
Take a look at the vast majority of signals inside current integrated circuits. They are lumped components, no? The whole secret to building inexpensive, high speed circuitry in the future will be the avoidance of the need for the technology that Rambus provides. That technology is expensive for a lot more reasons than just the license fees.
Rambus designs take more parts, use more power, are not as robust (i.e. Camino), and will take up more board space than solutions that bring the chips together in one place. The SDRAM solutions we see today take up more board space, but they are not the solution that Rambus must compete with. The place that PC technology is going, is to package the DRAM with the CPU, and that technology is likely to have such huge numbers of extraordinarily short wires, that there is no need for fancy transmission line connections.
By the way, Rambus knows that Direct Rambus only has something like 4 years left in it. The future of the company is being spent on getting patents on other ideas for sending signals around at high speed. With the stock at such a high P/E (even prospective earnings for the next four years added together), the fact that the company has to force itself onto unwilling customers should give you a clue as to what is going to happen over the next 3 years or so.
At the same time as Rambus works on those new technologies, all those companies that didn't want to pay Rambus royalties for this technology are also, quietly, getting their own patents. In fact, the only reason why Rambus got a foot into the door with its current technology is cause Intel held it open for them. With this fiasco, it is not all that likely that Intel is going to do this again. All the people who said that Rambus wasn't a good idea look real smart at Intel, Dell, Micron, and Compaq right now. Who do you think will be listened to for the decisions in the future?
The high bandwidths are available with less expensive technology, such as DDR, but the memory makers had Rambus shoved down their throats. Now the box makers are puking up those RDRAM chips, you guys should be puking up your shares, you'd feel better if you just got it over with. It is an unfortunate fact that someone has to hold the bag. But you might as well make sure that it isn't you.
People are buying this company on a very speculative multiple of its ability to extract cash from an industry that does not want to pay it. The Direct Rambus technology is falling apart, it is now obvious that the drop in RDRAM pricing (against all predictions by industry pundits) was caused by box makers like Dell canceling orders. If Dell thought this was going to be an easy problem to fix, they would be still taking RDRAM, and stockpiling it for the product release. Instead, they are probably sending product back, and nobody is taking them. If you think RDRAM is going to start shipping in volume soon, you should consider buying some RDRAM chips. They are selling below the costs of production right now, that is why Samsung stopped production.
By the way, the big advantage that Rambus provides is a decrease in the number of pins. Did you look through the list of articles I gave links to last night that gave costs per pin for various current and new technologies? If you did, then you would know that the cost of pins on a Rambus chip's data bus is more than the cost of a standard pin. Why? Do you suppose that having a pin that can drive DC 90mA without eventually failing would be cheaper than a standard CMOS output? Engineering is all about trade offs, and most of those trade offs involve cost. Rambus is just too expensive compared to the alternatives.
-- Carl |
