Scumbria, and all RMBS doubters, here is something for you to chew on, posts about Rambus from PTNewell on the Yahoo! board
messages.yahoo.com
messages.yahoo.com
messages.yahoo.com
messages.yahoo.com
It?s easy to sympathize with supporters of DDR RAM such as Mr. Yao, author of the 32bitsonline article "Will Rambus Go Bust?" For years the JEDEC committee labored to extend the life of SDRAM until the technically superior Direct Rambus (RDRAM) arrived. With its cost saving lower pin count, greater effective bandwidth, lower latency, and lower power consumption (hence lower cooling costs) the future clearly belongs to RDRAM. However DDR hoped to be quicker to develop and quicker to the market, and thus carve out a niche, at least in the short term. With such hopes rapidly evaporating as RDRAM ships and DDR stalls, Mr. Yao and like minded DDR believers grow more strident in their denunciation of Rambus. However in his entertaining but counterfactual article, Mr. Yao does much to undercut the position he advocates:
(1) Rambus and RDRAM obviously did not "peak in 1999". Production of RDRAM in 1999 amounted to a few hundred thousand 128 megabit dram units a month; production this month (April 2000) has hit 8 million units 128 megabit units per month. By the end of 2000 (according to Samsung, Toshiba, and NEC), production should hit 20 million units/month. If a two order magnitude increase in production in a year is "going bust", another such "bust" would leave Rambus as one of the most valuable companies on the planet.
nikkeibp.asiabiztech.com
(2) OEMs have not given a "cold" reception to Rambus; on the contrary dozens of OEMs have introduced scores of products in the first quarter of 2000 alone. A comparison of OEMs choosing RDRAM versus those choosing DDR is quite revealing:
dramreview.com
The four entries for DDR contrast poorly with the scores of entries for RDRAM; however this overstates DDRs marketplace presence. Only Nvida is using DDR, and that only for graphics cards, in a non-standard version unusable for PC or server memory.
Mr. Yao claims that "Dell does not recommend using Rambus for servers". On the contrary, Dell has just begun using Rambus for servers, such as the Dell Precision Workstation 620. Dell -- both the company and the man -- seem highly committed to RDRAM as the next memory standard (page 176 of Dell?s book), and has by all impartial accounts been aggressive in incorporating RDRAM steadily downward, starting with the top models. Of course supply limitations still constrain how far this can be done. Plainly Dell loves RDRAM:
dell.com
(3) Mr. Yao skirts the latency issue, implying but not quite stating that this is a problem for RDRAM. Despite frequent claims to the contrary, RDRAM has lower latency than either SDRAM or DDR, as even JEDEC committee members and former DDR advocates Samsung and Toshiba clearly state:
toshiba.com
dell.com
Apparently official white papers by leading DRAM manufacturers do not carry the same weight with Mr. Yao as does the unsupported claims of physician turned hobbyist Thomas Pabst. Mr. Yao cites the latter approvingly, although Pabst, who has a long history of declaring Intel projects "failures" (e.g., the Pentium II), began his review by confessing a personal loathing for Rambus ? BEFORE he "tested."
Incidentally, PC133 SDRAM actually has higher latency than PC100 SDRAM, because of the timing mismatch (refer to the Samsung white paper). This is largely irrelevant, because for nearly all PC applications it is bandwidth that really matters, and there RDRAM is unmatched ? and will further quadruple by year?s end.
(5) Not content with repeating older disproven myths, Mr. Yao manages to generate a few new ones. He claims, without a source, that RDRAM must be noisier because of its high frequency. Maxwell?s equations require that induction currents be proportional to dE/dT. RDRAM uses a voltage swing of a fraction of a volt; while SDRAM is 5 volts rail-to-rail. You do the math. PC100 or PC133 (100 or 133 MHz) should generate more noise than does 400 MHz (PC800) RDRAM. Mr. Yao provides no evidence that new laws of physics are needed.
(6) The additional cost of manufacturing RDRAM in practice is far smaller than Mr. Yao?s theorizing would suggest. Intel believes the price difference will be 20% by the end of the year; while Samsung believes 50% and still declining is more realistic. In any case, RDRAM is much cheaper to test, because of it?s high speed.
members.home.com
biz.yahoo.com
ee.asiansources.com
(7) The enormous design cost savings of the lower pin count of RDRAM are pooh-poohed by Mr. Yao, who declares that RDRAM still requires about 75% of the same motherboard territory. First, even by his account, this would be a large cost savings. Second, DDR requires a HIGHER pin count than even SDRAM. Third, the pin count savings are very real. Sony has announced that they saved $40 /Playstation II by using RDRAM, despite the higher initial costs of RDRAM (which has already been dropping rapidly). Samsung has announced (but is not yet shipping) a "disposable" $200 PC taking advantage of RDRAM?s granularity and low pin count. These issues are much more important to most OEMs than to Mr. Yao.
(8) In February, at the IDF, Intel dealt what appeared to be a death blow to the hopes of the DDR advocates, when Desktop Vice President Pat Gelsinger proclaimed that "DDR is too little, too late, and does not work in the desktop." By too little, he meant the bandwidth advantage over SDRAM would soon be insufficient as processor speeds rapidly advance beyond 1 GHz. Too late should be obvious; early DDR supporter Samsung recently stated that their will be no volume shipment of DDR in calendar year 2000. "Doesn?t work in the desktop" might not be literally true, but probably refers to the high costs of designing with DDR?s exorbitant pin count.
Intel also invested large sums of money with Micron and Infineon in return for the latter supporting RDRAM production. They did everything but draw a picture to show that the desktop environment was transitioning from SDRAM to RDRAM, with no place for DDR. No, pardon, me, they did draw the picture:
developer.intel.com
With DDR literally out of the picture, Mr. Yao resorts to a trick. He states that Intel will "support DDR and PC133" with the 815 chipset. On the contrary, the 815 supports only SDRAM, consistent with Intel?s stated intention of a gradual transition. Obviously no one can wave a wand and make RDRAM supplies instantly large enough to supply all the world?s PCs. Mr. Yao also appears incorrect in stating that the forthcoming Williamette will support anything but RDRAM.
(9) DDR advocates are left pinning their hopes on VIA and AMD. AMD will support DDR in a server environment, but has straddled the fence at the desktop level. However hints are appearing. They have licensed Rambus technology, and numerous advertisements for engineers with RDRAM experience have recently appeared on the AMD corporate web site. This seems to leave VIA as a DDR advocate?s last, best hope. Alas, three years after announcing the imminent arrival of a DDR chipset, VIA has reiterated the announcement in a way that makes DDR appear more distant than ever. Reading the 1997 announcement is entertaining:
viatech.com
There are numerous further counterfactual statements in Mr. Yao?s article, but listing all would tax anyone?s patience. As mainstream reviewers grow increasingly positive toward RDRAM:
hardwarecentral.com
DDR advocates are reduced to relying on professed Intel and RDRAM hater, hobbyist Tom Pabst. It is easy to see why their desperation grows. Of course they have every right to argue vigorously for the continued viability of their product. But it only undermines credibility to proclaim that Rambus, which has just experienced a 100-fold increase in product shipment in one year, with scores of OEMs introducing computers using the product just in the last few months, is "going bust." Mr. Yao is simply positing the untenable and defending the indefensible.
Patrick T. Newell
The Johns Hopkins University
Applied Physics Laboratory |
