Hi jim kelley; Re: INTC packaged the Pentium on a PC card to reduce quality problems associated with having a lot of OEM's building their own motherboards with varying degrees of quality.
I disagree for a subtle reason. Packaging the Pentium on a card was mostly done in order to remove the necessity of having the communication to cache be done through a socket. It is true that socketing the PC was difficult for the motherboard makers, but RIMMs are difficult too, and that didn't stop Intel from pushing them onto the MB makers. What Intel did was a form of integration, but using a PCB instead of a die.
Sockets are a pain at high frequencies; putting the cache on the same board as the CPU probably takes a clock out of the time latency of the stuff. I remember having to buy cache cars when I built PCs from parts, way back when. Then the cache went onto a package with the CPU. Now the cache goes into the CPU. Does anybody see a pattern here?
As I've posted before, the trend towards integration is all devouring. Long, long before Rambus will have earned its current stock price main memory will have been pulled onto the same card (or Multi-Chip-Module) as the CPU. This will largely obviate the necessity of having complicated mechanisms for dealing with the long wires and connectors needed for user installable memory. The final step will be placing system memory onto the CPU chip itself, but that is way in the future. (Of course with stocks trading at 500 times earnings, seeing way into the future is a useful thing for the long term buy and hold types.)
Even the Dataquest article says that DDR has won the graphics market. And the reason is that the graphics market is one that does not involve user installable memory, (though it once did). Consequently the designers are able to achieve very high frequencies on the DDR bus, and DDR becomes very efficient. That same efficiency is where mainstream PC design is headed, slowly but steadily. The elimination of cache cards is one step in that process.
By the way, there has been some speculations on this thread that my motivations for posting here is that Rambus takes all the design work out of a design. I've ignored this, but I really shouldn't ignore useless drivel like this. A lot of people have no idea what it is like to design a memory interface, and they might be inclined to believe that with the simplification of RDRAM, any moron could design a memory interface.
The RDRAM interface is considerably more complicated than SDRAM or DDR. This fact can be readily seen by simply placing the manuals for the two types of chips side by side and weighing them. SDRAM and DDR are relatively simple to control. The previous generation of graphics chips, VRAM, were a mess to design in, SDRAM was an amazing simplification.
In terms of the design of the external analog circuitry, the termination etc., it is generally fairly simple with either RDRAM or SDRAM. Of the two, RDRAM is probably more complicated, but it really isn't a big deal.
The more complicated part of memory system design is the part where you figure out state sequences that move data efficiently in and out of the memory.
As long as I'm on the subject, I should mention that both Altera and Xilinx, the two leading FPGA manufacturers, have reference designs for SDRAM and DDR memory interfaces. But neither of them have designs available for RDRAM. This is likely due to a combination of the high speeds of RDRAM and the unusual voltage/current specifications. Links:
altera.com
xilinx.com
One of the primary uses of FPGAs is the prototyping and early mass production of designs that will later be converted to ASICs. Quite a number of ASIC design houses will take XLNX or ALTR netlists and convert them into an ASIC more or less automatically. The fact that DDR is supported by the lion's share of FPGA makers, and RDRAM is supported by no FPGA makers at all, is significant, in terms of future design wins.
Ah but I forget. This is fantasy land, where no one actually has to build anything, and the customer will pay a premium for the product just as long as it says RDRAM on the side. Funny thing though, customers seem to be paying for a lot of those Nvidia DDR graphics card, and they are not at all shy about mentioning that they run on DDR. Maybe marketing types are taking notice...
-- Carl |
