Rambus, Inc.
Various [Thoughts]

Format for Printing

Format for printing

Request Reprints


By ptnewell
November 8, 2001

Posts selected for this feature rarely stand alone. They are usually a part of an ongoing thread, and are out of context when presented here. The material should be read in that light. How are these posts selected? Click here to find out and nominate a post yourself!

elixe writes:

[ptnewell wrote:]
"The failure of both RDRAM and DDR to do much for the P3 should have shown how pointless it is to try to outstrip the CPU with unnecessarily fast memory (of course the lousy FSB on the i820, which was most suitable to PC133, didn't exactly help).............

Just as RDRAM did not clearly surpass PC133 until 2 GHz class processors appeared, RDRAM probably will not pull well ahead of DDR (i.e., by more than a few measly per cent) until the Tehema-E, (sic, with 533MHz FSB), and 3 GHz processors."

Do you see the irony in those statements? For those who failed to see the irony:

We should have known RDRAM would fail with PIII/i820 because of the lousy FSB.
Wait till you see how the P4 with RDRAM at 3.2GB/s Bandwidth clobbers P4/PC133 at 1GB/s. Oh, the P4 with RDRAM doesn't clearly surpass PC133 unless the P4 is 2GHz.
Wait till you see P4 3GHz with 4.3GB/s bandwidth. That will pull ahead of DDR at 2.1GB/s.

The question that really needs to be answered, "Why didn't the P4/RDRAM at 3.2 GB/s memory bandwidth clobber a P4/PC133 at 1 GB/s bandwidth?"

Out of respect for elixe's ongoing useful contributions to the board (I seem to have a higher opinion of him than either the pro or anti-Rambus supporters here) I will answer this comment, although otherwise it would not be worth the while.

First, elixe, look up "irony" in a dictionary. An ironic sentence is one where the literal meaning is the opposite of the intended one. For example: "You must have thought long and hard before posting." Of course many people nowadays loosely use "irony" to mean just about anything. In this case, it means elixe did not understand what I meant, and did not think too clearly.

If anyone should feel sheepish about the i820, it is surely Intel, not me. In fact I think people are overlooking the fact that when Intel designed the i820, they never anticipated PC133. In fact compared to the PC100 machines at the time, the i820 was pretty good (in fact, when Tom's Hardware first went after the i820, he used an overclocked 440 chipset, to get a PC133 FSB and chipset). However neither the CPU nor the FSB could make much use of the advantage of PC800 over PC133. That is, even at about 800 MHz CPU speeds or a bit less, PC800 was better than PC100, but not better than PC133.

The fact that the i820 did not improve a P3 over PC133 platform led many to claim that RDRAM was no better than PC133. In fact, look back at any message boards to early or mid 2000, and you will find this argument debated endlessly. The difference between the i820 and most P3/PC133 platforms was typically only 1% or less overall. (It did not help either that AMD passed Intel early in 2000, so that the fastest processors were Athlons plus any kind of memory at all).

Longs, however, claimed that at higher CPU speeds, and with the availability of faster FSBs, RDRAM would open up a gap over SDRAM. This has happened. It is true at 1.5 GHz, it is true at 2 GHz, and the gap will obviously only grow. I do not think that there is anyone today who does not admit that the i850/PC800 is faster than i845/PC133. Clearly this is a vindication of the position of many long time Rambus longs.

Likewise, according to testing at Tom's Hardware, or Annand, the gap between i850/PC800 and Via/P4/266 MHz DDR is quite small. On some tests it is zero, on others more than 10%. On overall measures it is about 4%. (This is using the expensive CAS latency 2 DDR, by the way. No one is willing to test anymore with the CAS latency 2.5 DDR people actually are mostly buying. For many purposes, a latency of 2.5 cycles is really the same as 3, since nothing can happen halfway through).

Anyway, my claim is simply that next year when 3 GHz processors are available, and 533 MHz FSBs are available, PC1066 will have a wider lead than 4% over DDR 266 MHz on the P4. No one will be able to dispute the RDRAM advantage, as a few still do. Incidentally, I emphasize again that PC1066 is not just faster in bandwidth, it has lower latency. The sharp drop in latency (by almost 1/3) will also improve performance.
Of course there will always be applications that do not need or even benefit from higher memory bandwidth. Those applications will never gain anything from faster memory, no matter what the CPU speed. Other applications already see an enormous gain however (i.e., some applications see 30% and even more gain).

Finally, the argument that most users simply do not need the faster memory overlooks a fundamental fact about the way technology always evolves. Most people do not need a P4, or even a P3. In fact, they would do just fine with a 486. If MSFT did not keep upping the ante to match the hardware, a 286 would be fine. However the truth is that it is just as cheap for Intel to manufacture, say, a 600 MHz processor as a 100 MHz. In other words, the bulk of the population is simply dragged along by the minority. The 1.5 GHz P4 is CHEAPER than the 5 MHz 286 used to be.

Technology evolves in this way. An innovation is introduced (let us take color television). A few people only are willing to pay the high price, besides the benefits are limited at first (most stations are not broadcasting in color anyway). But instability exists. If the market expands even a little, costs go down. Then more people buy. Then costs go down further. And more people buy....

RDRAM used to boost system costs by an obscene amount, and produce trifling gains, on a very few applications. Now it has a modest system premium cost, and it produces recognizable and hard to dispute gains. In the future, it will have negligible cost premium and produce still larger gains.

The valid skeptical argument about RDRAM is not whether most people will ever pay for it (ultimately everyone would save money, from DRAM makers to OEMs if they made only one kind of memory). The question is how long the process will take (and if it takes too long, whether DDR can be improved to stay "good enough", just as the introduction of PC133 was "good enough" for the P3). It took color television about 15 years to reach half the market. During the early years, it naturally seemed to most people that it would never become the majority, since the costs were just too high. (The intrinsic costs of color television, involving 3 electron guns, were MUCH higher than black and white, far more in proportion to RDRAM/SDRAM systems). Of course the inexorable logic worked its magic, and you literally cannot buy a black and white television any more. Most "luxury" features introduced on expensive cars sooner or later becomes common, as the technology cycle works.

Of course I hope and believe it will not take decades for RDRAM to take over. Samsung says that the relative manufacturing cost of 32 bank RDRAM is only around 20% more than SDRAM, and that 4i RDRAM will have a smaller die size than SDRAM. Meanwhile, the advantages of RDRAM will grow, with the lower latency of PC1066, along with its higher bandwidth, and the availability of faster FSBs and busses. But these things simply speed up the inexorable logic of technological evolution: costs drop, more people buy, costs drop further...


TMF Money Advisor
Got money questions? Your answers are just a phone call away! TMF Money Advisor puts you in touch with an objective Financial Planner whenever you need it.