by Anna Filatova
11/10/2002 | 12:33 PM
In the past 10 years the phenomenon of the Rambus Company has grown too significant to be ignored, but at the same time too complex to be comprehended at large. So, we will only concern ourselves with one aspect of the company's activities. The one that directly affected the PC industry and made the whole computer and computer-related world talk about it. It is the timeline of the passed events that we offer, but the dates conceal a thrilling story of high hopes and bitter disappointments. Let's live it through.
Let's live through the past ten years or even more. The company was founded in March 1990 by two quite respectable persons from the well-known universities: Dr. Mike Farmwald from Illinois and Dr. Mark Horowitz from Stanford. Rather an ordinary microelectronics start-up, to tell the truth. The industry has seen a number of such companies, 99% of which are either dying off or being bought up by big sharks. But the way of Rambus was very violent from the beginning, although not suggesting the storm that was about to break loose.
A month after its foundation, in April 1990, the company applies for a patent for a technology developed by Rambus. The application was not satisfied, but the peculiarities of the US patent system are such that all subsequent additions and applications of Rambus in the wake of the original one were considered as if filed in 1990. Let's remember this fact as it turned to be of great importance in the events that were to happen.
In the events that were to happen, Rambus, as any memory maker, couldn't pass round JEDEC - the organization for collaborative development of electronics components. One of its 48 committees (the 42nd, to be exact) deals with the development of specifications for new DRAM types. The process was slow, though. Rambus representatives were first present at a JC-42 meeting in the end of 1990 and the official entering into JEDEC happened only in July 1992. By that time Rambus had already gained some fame: in March its new Rambus DRAM was licensed by Fujitsu, Toshiba and NEC.
Well, Rambus was interested not only in its own standard. That year the company votes against the ratified SDRAM standards on four JEDEC meetings (the overall climate was quite favorable: Nintendo announced that it was going to use RDRAM in its new game console). It was getting even more interesting: as the JEDEC statutes required the voting members to disclose all their patents concerning the ratified standards, September 1995 Rambus refused to vote on SyncLink and RamLink technologies without producing any reasons. The reason was quite simple, though: the company had already patented those technologies.
More to that, JEDEC's statute, Rambus's business plan and its chief officers' talk contradicted one another. The statute required showing the cards up while the company executives said that it had already been done, but only with respect to RDRAM technologies. As for the patents dealing with SDRAM, Rambus kept silent: it hadn't got them yet. And now a quote from the company's business plan, dating back to June 12, 1992:
"[We believe in] Sync DRAMs infringing on some claims in our filed patents; and there are additional claims we can file for our patents that cover features of Sync DRAMs. Then we will be in position to request patent licensing [fees and royalties] from any manufacturer of Sync DRAMs. Our action plan is to determine the exact claims and file the additional claims by the end of Q3 '92. Then to advise Sync DRAM manufacturers in Q4 '92".
So, during 1992-1995 the company attends JC-42 meetings, although it wouldn't normally be allowed to do so, and has every opportunity of using the ideas and concepts voiced at the gatherings. Rambus sought to have a ratified mass standard that would be based upon the patents belonging to the company. But sh-h-h-h! Don't tell anyone!
Don't tell anyone until the right time! Here starts the most thrilling part of our story. In June 1996, Rambus sends JEDEC a letter, which tells of the company's quitting the organization. The form of the letter was like "we quit and won't pay dues as our business plan doesn't agree with JEDEC's policy in respect to patents. By the way, we've got a few here: number such and such". Not a word that all this concerns SDRAM. And that's exactly when the PC66/100 specs had been polished off and the manufacturers started producing chips.
For now, Rambus had secured its rear and got to promoting RDRAM as a surer way to profits. We might never know the details, but anyway, in November 1996, after a few years of negotiations, Intel signed an agreement with Rambus to the purpose that it would promote Direct Rambus DRAM specification to become a widely accepted standard. This memory type was collaboratively developed by both companies. In particular, Intel was to develop a PC chipset with DR DRAM support in the following two years, that is, by the end of 1998.
Of course, Intel had a significant financial interest in the deal, but there is one more thing worth mentioning. Intel regarded and still regards Direct DRAM as the most perfect memory type existing and still thinks that it would be quite nice if it won the market due to its real competitive advantages and not by compulsion. More to that, the transition to the 0.13micron technological process has made this memory type widely available… But back to the past.
But back to the past, back to the revolution. In the processor market Intel puts off open Socket-7 in favor of proprietary Slot-1. In the graphics market Intel replaces PCI with AGP, although the latter was as open as the former. The obvious complement would be a transition to a new memory type. It's much easier to change the whole, than parts of it, especially as the new CPUs and graphics bus required something more solid than PC133 SDRAM.
But this something was already there! A year before, in 1996, two variants of further SDRAM evolution had shaped themselves: Double Data Rate DRAM and SyncLink DRAM. Big memory makers started developing designs of DDR DRAM chips. The SLDRAM consortium was established, Micron was actively promoting this memory specification. Memory makers knew what they were capable of better than Intel and their memory evolution plans were focused on performance growth and, unlike the Intel/Rambus suggestions, were based on the existing developments and infrastructure. The future proved to be on the DDR side as it won the market, while SLDRAM-relating technologies were used when developing DDR-II. Anyway, let's not rush ahead of time.
To cut it short, in November 1996 Intel took DR RDAM's side and announced that this memory type was going to come in place of PC100 SDRAM. If we recall the overall situation then - there was nothing besides PC100 in the market and Intel's shadow behind DR DRAM looked really impressive - we should agree that this memory type had all chances to become the No.1. Everyone then still remembered the coordinative and leading role Intel played in the PC100 development. That's why Rambus' enter into the stock exchange turned to be the success of the year: 98% growth on the very first day. Well, no one could have doubted it, actually.
There was a period of calm, which lasted several months, and then we witnessed the beginning of public DR DRAM advance. In the middle of October 1997, at Microprocessor Forum, Intel and Rambus announced their intention to launch the technology into the market in 1999. This was connected only with 800MHz DR DRAM with 1.6GB/sec bandwidth. They promised the RIMM modules with the capacity from 32MB to 1GB, with physical and thermal characteristics equal to those of SDRAM. It was also proudly mentioned that most memory makers had licensed the technology already.
Complete success. And it was a success, really. LG Semicon, Samsung, Mitsubishi - all and sundry - were licensing DR DRAM. Well, they were accustomed to it. Production diversification is quite familiar thing in the field and it's common practice when a company produces a few memory types at a time, sometimes even targeted at the same market sector. Of course, no one was going to miss out a perspective technology, which might win the market. And of course, no one was going to wholly devote itself to DR DRAM: the companies that sang praises to DR DRAM were at the same time promoting either DDR DRAM or SLDRAM or even both.
The following half a year - from June 1998 to January 1999 - became a real golden age of DR DRAM. Praise, optimistic forecasts and all that. Intel started testing memory chips and modules. Dell and Compaq announced that they were to ship PCs with DR DRAM the beginning of 1999. That is, the key point was reached, the technology was demanded by end-buyers. Toshiba forecast that DR DRAM would have 50% of the memory market by 2001.
A few months later, in October, the long-time Intel's foe, AMD, announced the licensing of the interface to be used in K7 chipsets in 1999. The company's vice-president said that it was a required OEM choice. But at the same time, AMD didn't want to tie K7's future with that memory type. After that announcement the Rambus' vice-president stated that this made DR DRAM the world memory standard. By that time, such giants as IBM and Compaq had entered the list of licensees.
A month later, at Comdex Fall'98, Intel proudly demonstrated a working PC with DR DRAM. However, the application running was an ordinary game called "Forsaken", as there were no programs requiring the increased bandwidth. But, there were problems with the latencies. Although, Rambus acknowledged the problem, they nevertheless claimed that it was solved and the wait period between the clock cycles was 10ns in DR DRAM, which was faster than in any available or planned technology.
The contract-bound Intel, which also was strategically interested in the success of the new technology, supported Rambus at its best and stimulated the memory makers financially. Micron received $500 million, Samsung - $100 million, negotiations with NEC and Toshiba were in progress. Rambus promised the launch of DR DRAM mass production in the first half of 1999, so that the arrival of i820 in the second half would be immediately followed by mass production of i820 based computers.
These were the last days on the Olympus. The hail of problems could wipe out a much stronger technology. All the time there had been an aggressive anti-campaign of VIA-headed group of companies that put their stake on PC133 SDRAM. And against this background there were permanent delays of the i820 launch, huge problems with valid DR DRAM chips output, and no certainty with most large manufacturers if they want to produce the new memory at all. I think you hardly remember S-RIMM, for instance, one of Intel's suggestions how to reduce anticipated DR DRAM chip shortages. It's a RIMM module with the power converter to 3.3V, which allowed using PC100 SDRAM chips on the PCB.
Well, the starting-gun fired and all rushed to the long-waited finish. Module makers were amassing production volumes: "The demand is sure to be at the end of the year. We don't know how high it will be, but want to be ready". They invested money in module testers trying to overcome the competitors. Connector and frequency generator makers were all getting ready for the June start. Graphics chips and graphics cards makers were planning to roll out their new AGP 4x products after the June launch of i820. All were investing money in production keeping the month of June in mind. Rambus' stocks reached their maximum: $109-15/16.
January. But the golden age had its problems, too. In particular, DR DRAM design was far from perfection in the beginning of the year. We should bear in mind that DR DRAM is a technology, an interface rather than a chip design. So, the manufacturers had to learn almost on the fly and introduce all the re-designs on the fly, too. DR DRAM was confronted by obvious rejection from workstation and server makers, as no chipset maker supported this memory. Not a single one! Although they were more numerous there, than in the desktop market. Even Intel had to develop a server chipset with DDR DRAM support! Expensive memory with questionable advantage in performance was of no use here and Intel's word wasn't that weighing: they couldn't force their position in this sector. i840? A server chipset that can only support 2GB of memory? You must be kidding!
February. But the real problems emerged a little later when the fortune turned its back on Intel/Rambus as if punishing for their excessive initial optimism. It all started back in February, with the rumors that only some Camino Lite with 600MHz DR DRAM support is scheduled for June. Samsung said it had been planned beforehand, Hitachi - "we know nothing about a 600MHz DR DRAM". Well, they were not quite frank here, as we remember well that 600MHz DR DRAM first appeared in November and was supposed to become a mass product for PCs.
And there was one more Intel Developer Forum coming. A year before, at IDF'97, the promotion of DR DRAM was announced. A heap of work had been done in the past year, but… The launch of i820 was delayed by three months, until September, because of the technical problems and low production capacities by the memory makers. Only one of the four offered frequency generators complied with the specs, only five of eight memory makers who claimed their support of the new memory type, showcased ready memory modules, but no module of all had been validated. Moreover, the only chipset supporting DR DRAM didn't feel that well. In addition, there emerged i815 with 133MHz system bus, although supporting only PC100 SDRAM. The escape way, no doubt.
And the hell broke loose: i820/DR DRAM was battered from all sides. They chastised all: heat production, unreliability, high cost compared to SDRAM and the licensing fees. Voices from the audience: DR DRAM brings no advantages to uni-processor systems while its cost is too high. The competitors raised their heads. February, VIA announced officially the establishment of a developer group working on PC133, which started its work in the middle of January.
March. Another blow followed. In the end of March the public learned about the bug Intel found in DR DRAM designs: CMOS truncated bit, an error that can stop data read from DRAM registers. Yes, the Rambus' approach implies that the logic is directly integrated into the memory chips and this may cause these things. When questioned, DRAM makers had no definite answer and pointed at the item in the DR DRAM licensing contract that forbade them to criticize the technology.
April. Intel started losing its ground fast. Firstly, an intermediate 700MHz DR DRAM specification was ratified as an attempt to increase valid chips output. A reasonable step, as 600MHz DR DRAM was not interesting to anyone in the industry while 800MHz might be too much for some manufacturers. Well, when there are such changes half a year before the final product release and after a few years of development, we can't possibly call it a stable design. But Intel and Rambus still claimed that whatever the difficulties, i820/DR DRAM was coming out in September. DR DRAM success was questioned more then. Samsung heroically covered the flank for Intel/Rambus and told everybody and everyone that PC133 SDRAM was a baby toy while 128bit DR DRAM chip market was expected to reach 50 million pieces in 1999. But at the same time the company still supported PC133.
Secondly, as people were concerned that there will be no opportunity to use the Coppermine with 133MHz FSB, Intel started giving away DR DRAM architecture, too. Paul Otellini, senior manager of Intel architecture business group, said: "I wouldn't consider it [Rambus] a must. 133MHz bus can be used apart of Rambus". And if it can be used, why not use it? But Intel will make this step a few months later.
Late April - Early May. VIA rolled out its Apollo Pro133, in spite of Intel's brandishing GTL+ license and claiming it didn't cover the 133MHz system bus. Well, there should be the last warning, so one of Intel's advocate offices sued VIA for all possible and impossible sins. The suit was taken back the same day, as it was merely a threat. Meanwhile SiS announced SiS630, an integrated chipset with built-in AGP 4x, UltraDMA/66,133MHz system bus and PC133 SDRAM support.
May. Intel rolls out an error-ridden i810. VIA continues shipping Apollo Pro133. Memory makers are polishing off the technological process, reduce DR DRAM chip size and produce PC133 SDRAM chips. By September, by the arrival of i820, most manufacturers will use 0.20micron technology already. There are a lot of AGP 4x graphics cards with no systems to use them on in this mode. The manufacturers of RIMM module connectors are increasing their production volumes. They have already got facilities to produce 800 thousand a month, are going to hit 1.5 million a month by September and are perfectly sure about the mass demand.
Seems familiar, doesn't it? Yeah, that's exactly what we saw in December-January. The same hurry, the same last preparations, the same wish to be the first to offer products for i820. Only then the companies invested their money relying upon Intel's promise to roll out i820 in June, and now - in September.
Intel already had the A1 version of i820 chipset. Mainboard makers referred to it as a flawed, slow, and too raw chipset. What's more, it required expensive equipment to design and test mainboards, which most manufacturers just didn't have. Taiwan was far from being delightful. As the situation with Rambus was not quite clear, memory makers continued working on DDR DRAM and developing 128, 256, 512MB and even 1GB chips. Moreover, 128MB samples were already available. The situation with i820/DR DRAM was just the opposite: memory makers were ready to start mass production of DDR DRAM and were only waiting for chipsets to support it.
June. AMD put DR DRAM off for a year and announced it was going to start with PC100 and then - PC133. On the other hand, IBM said that they were not going to use DR DRAM in its computers that year which slumped Rambus' and Intel's stocks. The next day IBM spokesman blamed journalists for telling lies and assured that IBM would use DR DRAM. At least, in high-end PCs.
Two days later, at Computex'99, VIA showcased its Apollo Pro133 - the first PC chipset to feature 133MHz system bus. There were also a lot of Apollo Pro133 based mainboards: from ASUS, Gigabyte, MSI. SiS, ALi and Reliance were also expected to roll out their PC133 chipsets soon. More analysts agreed that PC133 was a reliable bridge to DDR DRAM rather than Direct Rambus DRAM.
Intel started dropping hints that it had always liked PC133 SDRAM and that it wasn't against its use if necessary. If that standard becomes popular, why not? It took Intel five months to come to such a conclusion. The final decision on the issue will be voiced in September at IDF. For now, Intel decided to occupy itself with VIA.
First, all customers received a letter telling that VIA's chipset supporting 133MHz system bus infringed the license VIA was given in November of the previous year. Intel agreed to disregard the fact that VIA had already sent out chipset samples to the customers, but nothing of the kind should ever repeat. Seemed like the warning didn't help, as in about ten days Intel called back VIA's license and sued VIA accusing it of breach of contract, patents, negative advertisement and unfair competition. And gave the license to ALi instead. VIA was ready to defend its rights. We already know in what way: the head of the company didn't waste time in Silicon Valley and the negotiations with National Semiconductor were about to lead to actual results.
Meanwhile Intel was still working on i820. The B0 version was ready, although it was just a little better than the previous one. B1 again aroused negative comments. At the same time, although the chipset was still under development, Dell announced a new workstation family supporting DR DRAM, Precision 220, 420, 620, which was due in September.
July. Intel got a few reasons for joy: it could show to the customers several i820 based mainboards from big Taiwanese companies. They even showed an ASUS mainboard working. While Intel continues polishing off the chipset, VIA, guarded by National, officially announces Apollo Pro133 and begins shipments.
August was rather calm. And then September came.
September. At the Intel Developer Forum taking place on the first days of the month, Intel once again showed a system with 800MHz DR DRAM and announced that they were going to roll out a chipset with PC133 support in the beginning of the year 2000. It was there that Dell demonstrated the test results of DR DRAM vs. PC100 SDRAM that showed no great difference between the two memory types in available applications. Two weeks before the i820 launch, the test results appeared on various sites. And as the drivers got improved with the time, the results grew noticeably better. ASUS, AOpen, ABIT, Chaintech made public their i820 based mainboard specs.
Like the bolt from the blue: a weak before the i820 launch Micron said it preferred Apollo Pro133A to i820 for its PCs. They said it was the more attractive price-to-performance ratio. According to Micron, the difference was only 2-3% in typical business applications. Micron was the only big PC maker to listen to common sense. All the others listened to Intel.
And they paid dearly. A few days passed and Intel delayed i820 launch due to an uncovered design error: a problem with the third RIMM slot that led to system malfunctioning even when the slot wasn't used (memory bit error). By some estimates, it resulted in mainboard makers losing about a million of i820 products. Assume that the average price of a mainboard like that is around $100, the total loss can be evaluated as $100 million. At that time the memory makers were already not very eager to work with DR DRAM because of the SDRAM price growth (we are talking about the autumn of 1999!) and now they lost even the slightest interest in this product. They switched the production lines to the manufacturing of 64MB SDRAM for an indefinite period of time. Well, they had to cover the losses as every silicon wafer that was dedicated to DR DRAM turned to be lost profit.
That's when Rambus needed the spare solution. DR DRAM couldn't bring any money, could it? The industry preferred PC133 SDRAM and produced millions of these chips? Ok, but let them pay license fees then. So, in the beginning of the year 2000, Rambus entered the market with an unexpected announcement: PC100/133 and DDR SDRAM were based on its, Rambus', patents that is why all the manufacturers had to pay fees. The sum of the fee, according to some officials, was shockingly high.
Seems like Rambus guys turned out better businessmen than engineers as their actions in juridical and economic fields were immaculate. They easily took advantage of the weaknesses of the companies that confronted them. Toshiba was the first to give in and on June 16, 2000 officially licensed SDRAM from Rambus. Well, the company produced RDRAM for the new Sony PlayStation 2 and wouldn't live through a license withdrawal.
Hitachi was the second to give in a week later. It was most surprising as the company was a stronghold against Rambus and proclaimed fight to the victory end. Well, it was Dreamcast here instead of PlayStation 2 and Sega instead of Sony. And Rambus demanding to stop console shipments. Rambus stock quotes were jumping up and down by dozens percents a day. There were other victims as well. NEC was the biggest of all, but that didn't change the overall picture. An interesting fact: SDRAM was licensed by all big and many small Japanese memory makers, but outside Japan there were few who succumbed.
Samsung remained neutral as it could afford to do so. Hyndai, Micron and Infineon decided to stand their ground to the last. So, the four world's biggest memory makers appeared against Rambus. Intel was also losing the nerve and talked about PC133 SDRAM and DDR SDRAM. In the middle of October 2000 Rambus reported company's record of $14.5 million profit compared to $3.4 million in the same quarter of the previous year. However this announcement had no real effect: the next day the stocks slumped more than 10% down. That was the end.
It was the year of complete defeat for Rambus, its Waterloo. In March, Judge Robert Pain ruled Rambus vs. Infineon suit in the way that the former cannot produce claims for allegedly infringed SDRAM patents. It was all clear since then: Anglo-Saxon law is greatly based upon precedents, so the result of the first Rambus process could later determine the ongoing ones.
That's exactly what it did. Rambus still sues Hyndai (now - Hynix) and Infineon, as well as they sue Rambus, the appellations can be considered for years, but it's all clear from the juridical point of view: Rambus lost.
Rambus lost, but it also won. It all depends on the viewpoint.
What about the economical aspect? A clear win, although not that big as might be expected. Intel wanted to get the opportunity of producing regular Pentium 4 chipsets (with DDR support) at any rate and just paid Rambus off. In September 2001, they signed a contract by which Intel had to pay Rambus $5-8 million a quarter for five years!
What about the PC market? A clear failure. i820 couldn't make it on top. The situation with i850/i850E is not that clear yet, but by the arrival of Pentium 4 CPUs with FSB533 where this high bandwidth could be demanded, and by the time DR DRAM price got smaller due to the introduction of 0.13micron manufacturing technology, Intel had already started producing chipsets with PC2700 DDR support. They can be compared with dual-channel PC800 DR DRAM in terms of bandwidth.
On the other hand, there is DR DRAM PC1066, officially certified by Intel for i850E, there are 32-bit RIMM memory modules, there is a newly made SiS658 and all this gives some hope to Rambus. As the technology is improved, the price of DR DRAM is going down. So, it's not quite clear what we will have in the future. Yes, the next year is the year of PC2700, but then? DDR-II? Seems likely. But maybe it will be DR DRAM that can compete with DDR-II in the specs and featuring some, though scarce, infrastructure.
This story is an instructive one and helps to understand the laws of the computer market. The idea is that open architectures are on top here due to much higher economic effectiveness. Intel and Rambus wanted to lead the market their way and were beaten. Intel isn't an immaculate leader in technology market. The company may be the best in technology, but the market may not accept their "price-to-performance" ratio. And there is no power to force unattractive products into the market.
You thought it was all about Rambus? Yes, it was. But the main point here is the factual proof of the thesis that competition is good. It may not allow the most perfect technologies into the market, but always takes up the most optimal ones.