Hidden Consequences: Rambus And DDR SDRAM Prices

Jimmie writes: "Looks like Rambus, Inc intends to use licensing fees arising from its recent patent settlements to force the price of DDR SDRAM to be high enough that Direct RDRAM (which we know is ridiculously expensive right now) can compete. When asked that very question, the VP of worldwide marketing at Rambus replied 'I wouldn't argue with that conclusion.' Story at ebnews.com." Sometimes the computer industry's oldest saying seems to be "If at first you don't succeed ... squeeze out some competition."

Comments

Whatever happened to ESDRAM and SLDRAM?

[Guppy]

There used to be several other types of memory under development for the next industry standard, some of which showed a great deal of promise. Two of these were ESDRAM and SLDRAM.

ESDRAM was a type of SDRAM-like memory that included a small amount of SRAM cache on each chip which lowered latency, allowed for greater utilization, and also could boost bandwidth by allowing wider buses and moderately higher speeds. Here's an ESDRAM article at Lost Circuits.

SLDRAM was, like RDRAM, a protocol based memory. Unlike Rambus, it was developed by an industry consortium, and was to be royalty free. It allowed for a faster bus, and could also be operated at a double data rate. Supposedly, in some situations it might have actually been faster than RDRAM. Here's a link to the SLDRAM Corporation.

Re:doesn't make sense

[smack_attack]

Isn't this a monolopoly of the SDRAM market?

No, it's not a monopoly, they haven't built any hotels yet. :)

Re:Where's the DOJ now?

[rodgerd]

This is, of course, factually incorrect on the part of the gentleman from Intel. Clean room reimplementations of a technology are a great way of evading copyright and trade secret laws, but they do nothing for patents - reinventing patented technology will still see you paying the patent owner, even if you can prove that you'd never heard of the patent or consulted it during the invention phase.

The only way around a patent is to design a different mechanism that has the same outcome; not so much reinventing the wheel as discarding the wheel in favour of the catipillar track. The problem with this is that in the modern era of loosely written patents, it's almost impossible, since patents usually cover enough ground to prevent alternate mechanisms, or patent the outcome itself.

Moreover, you'd probably end up with a technology different enough it wouldn't be business competitive - requiring different motherboards, etc.

Can Rambus possibly be that cocky?

[Anonymous Coward]

For more info on Rambus' past business practices, check Tom Pabst's article on "Why We Don't Trust Rambus"

However, I am puzzled by this strategy. For one thing, despite TI's success at defending their patents on fundamental aspects of DRAM (one of many legitimate IC patents they obtained in the early days of IC manufacturing), RAMBUS must know that there is prior art that weakens their claim. While exisiting licensees would arguably tend to support (or stand idly by) as they pressed others to license their technology, cockily taking on an entire industry in this manner is just begging for a barrage of assualts on the patent itself.

I can't believe they would be so blind, even in the heady glee of being gran such ridiculously broad patents like #6,067,592 System having a synchronous memory device [May 23, 3000] and #6,049,846 Integrated circuit having memory which synchronously samples information with respect to external clock signals

In fact, I think we have a duty to use the accumulated experience of our older (and much underappreciated) readers to start picking off the more outrageous patent claims one by one.

The easier we make it for the remaining memory companies to see their options, the better for *us*

To save you some work, here's a partial list of active Rambus patents (and linked to each of the patents themselves). It's not a complete list (IANAPA), but it should save everyone some work getting started. See anything that looks familiar from "way back when"?


1. 6,075,743. Method and apparatus for sharing sense amplifiers between memory banks

2. 6,075,730. High performance cost optimized memory with delayed memory writes

3. 6,070,222. Synchronous memory device having identification register

4. 6,067,592. System having a synchronous memory device

5. 6,049,846. Integrated circuit having memory which synchronously samples information with respect to external clock signals

6. 6,044,426. Memory system having memory devices each including a programmable internal register

7. 6,038,195. Synchronous memory device having a delay time register and method of operating same

8. 6,035,369. Method and apparatus for providing a memory with write enable information

9. 6,035,365. Dual clocked synchronous memory device having a delay time register and method of operating same

10. 6,034,918. Method of operating a memory having a variable data output length and a programmable register

11. 6,032,215. Synchronous memory device utilizing two external clocks

12. 6,032,214. Method of operating a synchronous memory device having a variable data output length

13. 6,021,076. Apparatus and method for thermal regulation in memory subsystems

14. 5,995,443. Synchronous memory device

15. 5,966,731. Protocol for communication with dynamic memory

16. 5,956,284. Method and apparatus for writing to memory components

17. 5,954,804. Synchronous memory device having an internal register

18. 5,953,263. Synchronous memory device having a programmable register and method of controlling same

19. 5,940,340. Method and apparatus for writing to memory components

20. 5,928,343. Memory module having memory devices containing internal device ID registers and method of initializing same

21. 5,913,046. Protocol for communication with dynamic memory

22. 5,896,545. Transmitting memory requests for multiple block format memory operations the requests comprising count information, a mask, and a second mask

23. 5,872,996. Method and apparatus for transmitting memory requests by transmitting portions of count data in adjacent words of a packet

24. 5,844,855. Method and apparatus for writing to memory components

25. 5,748,914. Protocol for communication with dynamic memory

26. 5,748,554. Memory and method for sensing sub-groups of memory elements

27. 5,680,361. Method and apparatus for writing to memory components

28. 5,657,481. Memory device with a phase locked loop circuitry

29. 5,606,717. Memory circuitry having bus interface for receiving information in packets and access time registers

30. 5,511,024. Dynamic random access memory system

31. 5,499,385. Method for accessing and transmitting data to/from a memory in packets

32. 5,499,355. Prefetching into a cache to minimize main memory access time and cache size in a computer system

33. 5,434,817. Dynamic random access memory system

34. 5,430,676. Dynamic random access memory system

35. 5,390,308. Method and apparatus for address mapping of dynamic random access memory

Oh dear

[seizer]

This isn't good for anyone, not even for Rambus. Cyclically, companies who behave like this will have a brief period (or no not so brief - look at IBM, they managed for decades) of dominance, verging on monopoly - but their nemesis will come. Rambus is quite happily shooting itself in the foot for short-term gains. *sigh*

--Remove SPAM from my address to mail me

Re:Oh dear

[arivanov]

Exactly.

RDRAM, SDRAM, DDRAM are not the only technologies around. There is a whole lot os technoilogies circulating around Big Blue, Big Q and other not so big companies for their Big and not so Big boxen. For now these technologies cost Big Bucks. But this is for now as they are not manufactured in big enough quantities.

Rambus with its "milk the cow with a vacuum pump until it bleeds" behaviour is asking for these technologies to be brought to the PC market. And it does not have any IP there whatsoever.

The only problem is that this will take time and cost money that will come out of the consumer pocket. And all this because Hitachi did the favourite "japanese corp versus american in court" and chikened out. I hate japs when they do this. It had a perfect case (IANAL) with a list of violations of all kinds of regulations by RAMBUS.

There is some hope, though as price-fixing is a serious offence in EU for example and the EU comission has been amazingly fast recently on quite a lot of cases (like MCI/Sprint merger).

Maybe this cloud has a shinny lining

[maraist]

Though this situation really eeks me, and I am appauled that Rambus can get away with this sort of patent ( though not knowing all the details, perhaps the patent has legal merit ), there might be a bright side.

Case 1: Rambus doesn't do an Apple and raise the price of licenses to shut out competition. SDRAM will remain the memory of choice for a while. Intel will deminish support for it ( more mishaps with i820, etc ) SDRAM begins to go the way of EDO. DDR-SDRAM is expensive because it doesn't have a wide support base. The rift in memory market-share allows RAMBUS to market RDRAM as server memory: low-volume, high price. Thus the consumer is faced with either cheap, yet antiquated memory or expensive memory. New memory technologies ( which have been trying to emerge ) do not get a chance because the rift in memory markets and chipset support will be hurting.

Case 2: Intel does as is currently projected. DDR-SDRAM becomes comparible to RDRAM in total value. RDRAM is going to win out, as far as I can tell. Intel is most likely only going to support RDRAM, so the market for DDR will be too small to really hit critical mass. I speculate that RDRAM is actually faster then DDR ( especially under heavy concurrent access, such as truely utilized AGP and SMP ). To my knowledge, DDR only ups the speed of the interface, the underlying technology is not significantely different than that of SDRAM ( much like ATA66 or SCSI 100). With this RDRAM will become mainstream, espeically as CPU's break the 1GHZ barrier ( more speculation on my part, based on the starvation of CPU on both memory latency and bandwidth. Of which RDRAM addresses BW. DDR mildly addresses latency ( wrt RDRAM ) and provides BW ( though only superficially ) ). In the medium run, cheaper RDRAM is going to help a lot of power-hungry people ( though probably not as much as it will hurt intro and intermediate-level system purchasers )

Case 3: RAMBUS blatently prohibitively prices DDR-SDRAM. Now SDRAM and RDRAM are the only real players for PCs. What happens here is that RDRAM production can really begin ramping up to critical mass quickly because there is less uncertainty about the future. Prices will drop quickly over time ( though no where near SDRAM prices ). This is exactly what Intel would want. Their low-cost Celeron systems are perfectly suited for 66MHZ SDRAM. If you're an intro system, why would you bother with high-perf memory. Previously the blur between 66 and 100 allowed people to over-clock the external celeron bus. Now there is a world of difference between celeron and their "workstation-class" systems which come at a significant premium. SDRAM will either become cheap or expensive in the medium run ( lower volume production might mean higher prices if demand is sufficiently high. A typical sinario would be over-stocking of ?SDRAM causes prices to plummet, which prompts massive volume reduction, which later causes prices to go through the roof ( where it will stay ) ). Thankfully I don't think SDRAM demand will tank, so it should stay relatively normal for a while. Now here is the good part. There will now be a massive rift between the expensive ( though significant less than today ) RDRAM and the SDRAM. Hard core gamers, most servers and many desk-tops are going to opt for RDRAM. Most value-PCs are going to opt for SDRAM, but those value-conscious hard-core people ( like myself ) are still going to have demand for something inbetween. And perhaps there is a better compromise between latency and BW in a newer technology ( I read memory technologies a while ago about multi-row caching via SRAM which seems an intelligent approach which is compatible with most interfaces )

Basically I'm saying that those who demand performance are going to benifit from the demise of DDR because it'll ultimately be cheaper for them. Those that are on a budget are going to see minimal increases in their low-end memory ( assuming things don't gyrate too wildly ). And those of us looking for an overall better solution might find solace in a new technology that fills this important niche.

In the long run, I think this has good consequences ( if you ignore the whole moral imperative of stifling competition, which we can't really effect here ). In the short run it doesn't really even affect me ( being an AMD person ). It is only the medium run that has issues ( I wonder if I can hold off purchasing another computer for 2 or 3 years ).

-Michael

Re:Maybe this cloud has a shinny lining

[Upsilon]

You're whole argument is based on the simply untrue assumption that RAMBUS is ultimately a better solution than DDR. Well, I've got news for you: it ain't. The goal behind RAMBUS was to increase bandwidth and lower the pin count. They did both of these things, creating a 800 mhz memory module (actually 400 DDR) that uses a 16 bit bus. This delivers the same bandwidth as 200 mhz SDRAM would, since SDRAM uses a 64 bit bus. However, RAMBUS ran into a LOT of unexpected problems. Let's go over them all:

First of all, there's the whole latency issue. Basically, RAMBUS will always have higher latencies then SDRAM. This is a big part of the reason that RAMBUS can't seem to significantly outperform regular PC133 even though it should have 50% more bandwidth.

Then there's the memory limit issue. As one adds more and more RAMBUS memory, because of signalling issues (which I'll be the first to admit I don't completely understand) it becomes extremely difficult to keep it all working. The end result is that there is currently no way to have more than 512 megs of RAMBUS memory. This simply isn't enough for servers, and because of this even Intel, the big RAMBUS supporter, plans on using DDR memory for the server version of Willamette. It won't be long until 512 isn't enough for the desktop either. Do you really want to be stuck with this limit?

There's also the price issue. RAMBUS was supposed to be cheaper than SDRAM because it uses fewer pins, but things don't always work out the way they're supposed to. RAMBUS yeilds are horrible, and it is unlikely they will ever be as good as SDRAM yeilds. RAMBUS also requires a more expensive packaging and needs its own heatsink to deal with heat issues. The idea of ever putting RAMBUS in a laptop horrifies me. Also, because RAMBUS is yeilding so terribly 800 mhz modules are actually quite rare. Most RAMBUS being made is either 600 or 700 mhz, which compares quite pourly with PC133 considering that the 800 mhz modules can barely outperform it. There are also concerns that RAMBUS might not have much room to grow. Some video cards already use 200 mhz (non-DDR) SDRAM, and although these speeds are not mass-produced yet, the biggest limiting factor is motherboards which support it. SDRAM still has a lot of room to increase its frequency. RAMBUS doesn't.

Of course, during all this I was comparing RAMBUS to regular SDRAM, and RAMBUS doesn't compare very well at all. It wins a few points, but not many. Now, let's compare it to DDR SDRAM. DDR SDRAM is estimated to cost about 3% more to manufacture than regular SDRAM. 3%! RAMBUS costs at least 4-5 times as much! Now, PC2100 (133 mhz DDR, they rate it by the bandwidth and not the frequency for marketing reasons. Probably because they don't want the average consumer to be choosing between PC266 and PC800 not realizing what it means.) will offer more bandwidth than RAMBUS, still have the lower latencies of SDRAM, cost significantly less, and not be limited to 512 megs. Do you really want RAMBUS to win?

Re:Maybe this cloud has a shinny lining

[maraist]

I definately agree that RDRAM is not worth the money. But I know companies that shell out mega bucks on name brand server equipment because they can't afford to do any less, so it's not really an issue.

I'm not fully understanding why there is a memory size limit. I know that you can do 1, 2 and 4 channel RAMBUS configurations. I don't know if it's possible to do more ( seems likely that you should, at least in custom boards ). It would make sence to me that additional channels should be completely independant and therefore there would be no theoretical limit to the size. Thinking in terms of a SUN server where they design them from scratch, only customizing the back-plane, etc. You should be able to make a rack-based server that could have attached memory channels. This would help dessipate the heat ( by seperating them ) and if done in a sufficiently modualar and heirarchical approach, you should be able to have as much memory as you can afford ( at this point, the memory would be the least of your costs ). From what I've read ( mainly on Toms Hardware, and the like ), the timing limitiations are because each chip has to ( at boot time ) figure out the longest possible probagation path and sync itself accordingly. I believe the RDRAMS ( I can't remember from RIMMS, modules, etc ) are kind of daisy chainned, which means that the signal has to probagate through each device on a channel, and therefore the more memory chips you add ( RIMMS I believe ), the slower that channel. There's a max probagation delay that you can achieve, so depending on the setup of the memory modules, there's a max memory size that a channel can hold. I am not aware of any inter-channel dependancy, so you should be able to go as wide as necessary, though this adds incredible expense to the system.
High performance systems, however, tend to spend a lot of money on the memory subsytem. I remember studying the SGI, HP and SUN setups, where you'd have many interleaved channels. It's really just a matter of what becomes the mainstream. And it looks like 1, 2 and 4 will be it ( just like 1, 2 and 4 CPU configurations ). I can't imagine that you'd have the same sort of problems with many independent channels as you would many independent CPU's. The only bandwidth you're really fighting over is the inside of the chipset, and you can cheaply splurge there( especially since your wire-count is drastically reduced ). 4 16bit RAMBUS channels shouldn't be many more wires than a single SDRAM channel. I'd be interested in learning more about the limitations though.

Onto some of your other points. The I believe that the RAMBUS guys knew that they'd have latency problems from the start ( not that you made any inference on the matter ). Intel has been working that problem for years. Their solution is making latency tolerant devices. Latency from a slow-responding memory device running at clock speed is no different than fast responding SRAM that's clocked 1/12 of your CPU on average ( since on average you'll have to wait 6 cycles before it can hear you, then you'll have to wait an additional 6 cycles before it can transmit ( even after it has a response ) ). DDR-SDRAM alleviates the problem at least, but as far as I know, it's still pipelined and therefore does not provide the best possible latency. Still, I belive it's significantly less than that of RDRAM so DDR wins this battle.

I must admit a 16bit bus could almost only justify itself if it were cheaper or easier to manufacture. You're needlessly adding delay ( albeit at 800MHZ on an 800MHZ machine, this should be negligable ). The fact that it's so many times more expensive is just an insult to everyone ( though I honestly believe this is an economy of scale, and as with my article, reaching a critical mass would mean signifcantly more affordable system ). As with the above, the real benifit is probably in the ability to multiplex many channels. 4 168pin devices, for example, would be a nightmare to multiplex.

If nothing else, the excessive heat is unbelievable. The design spends great length discussing the idle and sleep modes which are used to keep the devices from overheating!! Those sleep modes wind up causes significant latencies, especially since I've gotten the impression that a given channel can only have 1 active RIMM(?) at a time. Therefore a significantly loaded server is going to be spending almost all it's time RIMM thrashing.

Finally, in response to "do I really want RAMBUS to win". Of course not; my silver-lining was really that we _might_ get another [superior] technology to win out that would not otherwise have gotten a chance. Will RAMBUS be shoved down our throats? Of course. Will people buy into it. Well, they're already doing so ( either by ignorance, or because their particular app is well suited to RAMBUS ). Will I personally ever buy RAMBUS, or would I ever recommend a friend to do so? Not unless all other options are exhasted.. which I guess means, maybe.

-Michael

At least micron hasnt caved yet..

[puppet10]

Micron hasn't decided to fall like the dominoes of Tosiba and Hitachi. And Intel doesn't think this will stop independant memory development either. This offers some hope that someone will stand up to these overbroad patents. Its not quite over yet.

A larger problem

[Veteran]

Rambus is symptomatic of a much larger problem which is coming in industry; the problem of intellectual parasites. In the past patents have not played that large a part in shaping how industries produce goods; most patent disputes were handled by cross licensing. If producer 'A' had a patent it tried to enforce against producer 'B', usually 'B' had patents it could enforce against 'A'. The results were cross licensing agreements which allowed both companies to continue in production.

Of course sometimes companies had no Patent portfolio to cross license, and were forced to pay royalties. However, being producers themselves, the patent holders couldn't charge too much; corporate 'karma' prevented it - after all if they charged predatory royalties - somebody could do the same things to them.

Rambus on the other hand produces NOTHING; they exist only as a shell firm with a PR department, a legal department, a portfolio of patents, and little else. Because they produce nothing, companies like Rambus are not subject to cross licensing, and corporate 'karma' has no effect on them. The result is that a Rambus style firm is free to gouge on their royalty demands. The only thing which limits them is the threat of a counter suits by deep pocketed memory producers with the object of invalidating the Patents.

In effect companies like Rambus have the moral stature of email spam; both are a parasitic drain on a system which tends to hamper the productive.

I think that we are only seeing the tip of the Rambus future. Because Rambus makes nothing, their 'cost of goods sold' is zero; at least Microsoft has to pay for the blank CD's that they press. Thus, minus their expenses, everything that Rambus takes in is gravy. One of the things that I expect them to do in the future is use their money to acquire additional patents to suck even more blood out of the economy.

It is no accident that both spam and productionless, patent holding companies were dreamed up by lawyers. Neither is an approach that productive people would think of.

Is there a solution to parasitic patent behavior? I think that there is. In the United States patents can only be granted to individuals, NO COMPANY HAS EVER BEEN GRANTED A PATENT ON ANYTHING. Companies obtain patents by having them assigned to them. This is typically done by means of employment contracts which force inventors to assign their patent rights to the corporation. This puts into effect the first layer of parasitic behavior. Most abuse of the patent system occurs because of the assignment process; if the law were changed so that only an INDIVIDUAL could own a patent - as well as be granted one - most of the parasitic and bad consequences of the patent system would disappear.

Some clarifications

[jeroenb]

First off, I'm not a big supporter of Rambus memory, but I do think it's a pretty decent solution.

The fact that RDRAM has a higher latency than SDRAM will become less and less important as CPU caches increase in size - this will mean larger chunks of memory will be copied into the cache at a time, making bandwidth the bottleneck instead of the latency.

The signallig problem is called bus skew and basically has to do with synchronizing at high speed. It might sound ridiculous, but when you have signals moving across a bus at 800+MHz, pin 0 could already be sending its n+1th bit when pin 15 is still waiting for bit n. This means you have to be absolutely certain your timing is perfect, because adding circuits to fix these problems will only slow things down again (defeating the point of a high speed bus.) This is something SDRAM will run into sooner or later as well though and the problem is much worse with 64 pins.

The yields are currently very bad because you have to assemble a RIMM entirely before you can test them. This means that if you have 1 chip that doesn't work, you can throw away the entire RIMM. Some calculating shows that in a situation like that, 90% yields in chips will only mean about 50% yield in RIMMs. This is probably just a matter of time though, before someone figures out a way to test them without assembling the entire RIMM first.

In the long run, I think Rambus could have a really interesting technology, but the way they try to force it onto everybody and demanding licensing fees and royalties, it will be hard to convince people... The company sucks and will probably drag down the technology with it.

Re:what?

[maraist]

Fist, and foremost. Moderation has nothing to do with agreeing with content. I therefore disaree with your supposition that I ( or comments like mine ) should not have been moderated so highly. The goal is to weed through uninteresting, or flaming remarks. If someone like myself is erroneous in my comments and it is highly marked, then someone like yourself will be able to pay attention to it and make your rebuttle. Those that see my comment will at least be given the chance to see yours as well, and all will be good. Additionaly, if I was truely incorrect, then by having someone like you find my mistake, then I would actually be given the chance to learn the err in my ways. This is called information exchange. I like this system.. It works; It has for me at least; I have been incorrect in the past about an original post or two ( meaning I would not have otherwise had the opportunity to learn ).

As a second point, I would like to thank you for not producing a flame, but instead being insightful.

With that said, I'll have to disagree with you and stick with my original comments. I do not believe that your agument holds.

First, my choice of the word speculation was not an admission of ignorance, but more a disclaimer saying that I do not have definitive evidence. Nor am I trying to prove a point based on the speculation, so it really matters little. Notice that my point was not that we're all going to be better off if RDRAM is our only choice, so long as it gets cheaper. But that, of the many ways that this situation can go, RDRAM will most likely become cheaper. Which will, in turn, help period. There is a possibility that this will backfire on RAMBUS / Intel in the creation of a nich for someone else to fulfill ( and ideally free us of memory woes ) which was the whole "shiny lining" that the article was about.

As for your comment about the "PROVEN" superiority of SDRAM. I would like to challenge you or ANYONE to verify this. I have a personal pet pieve about the word proven.. Anyone that uses it is put into a catagory in my mind; nothing in life is for certain or proven. Evidence is not proof of something, but only support of it's concept. I get tired of using the phrase, "the world was once proven to be flat because all _known_ evidence supported it".

As for SDRAM and RDRAM, I read Toms Hardware, Anandtech and Sharky Extreme. Yes, there have been benchmarks that showed SDRAM in the lead, but I have ALSO seen benchmarks by them that say the opposite. I really don't feel like digging it up, you can either take my word for it, or go research it yourself. I'm sure their articles are well marked. I'd be curious to learn of your findings. Additionally, almost all their benchmarks are on single CPU configurations, which is NOT the theoretical target of RDRAM. The stated design goals of RDRAM revolve around multiple simultaneous access; something not very well supported by existing single CPU systems. Intel's Italium is completely based on massively concurrent access. AGP is theoretically bassed around this as well ( though I do not think much aside from the failed intel 740 card ( or possibly even the i810 chipset ) makes full use of the concurrent bandwidth ( by among other things masking the latency ) ). My use of the word speculation is because I have not yet seen ANY benchmarks that test this theory. Provide a fully speced out AGP and SMP configuration then directly compare SDRAM ( and DDR for that matter ) with RDRAM, then one will be able to comment for or against my claim. I speculate primarily on the theory ( which I know can not be used as a basises for an argument, but again, that wasn't my main point in the argument ). Once again, I HAVE actually seen numbers that put RDRAM based systems ahead of others ( they were topping out the scores ). Thus, if you are paying top-dollar for a system, then I believe that your best bet at the moment is an RDRAM system.

As for DDR-SDRAM, I admit ignorance on the details. I've only read pices here and there, and don't even remember if I've seen any benchmarks on it. I do however remember that it's main call-to-fame is the double-transmittion ( rising/falling-edge transmission ) which we've come to know and love for the past two or three years. The main advancement here is NOT that they've created faster memory, but that they've learned how to transmit higher frequencies over longer distances. High freq. over the length of a motherboard will introduce lots of noise, and will drain a lot of power. By taking the same signal and getting twice as much data out of it, you're essentially getting something for nothing. A true advancement in technology. I made the comparison to ATA66 which allows maximum data-transfer rates to be un-hindered, but there is NO harddrive today that I have ever heard of that can transmit 66Meg/second continuously. This is primarily for use with cache bursts. Write 2 meg to the drive quickly, so that you can allow the second device to read. It stays in the drive's cache until it can be committed to disk at a more realistic 8 Meg / second. Upping the bandwidth allows improved latency by getting info to and from the device quicker. It works best under heavily loaded situations ( moreso in SCSI than in IDE however ). Likewise, DDR-SDRAM will allow addresses and writes to be transmitted more quickly, and thereby freeing up the bus for additional operations. However, this does not say anything about the memory's ability to sequentially process it. To my knowledge, we're still dealing with plain old, old-school memory access, though in a slightly more pipelined fashion. Concurrent access is not addressed by this technology. Theoretically, the mem-controller could rearrange memory accesses to maximize locality ( as would a SCSI controller ), but this does not garuntee throughput. RDRAM's use of independent channels can ( note: this is speculation based on the theory ) allow, under most circumstances, greater bandwidth provided critical mass of concurrent access. As we've seen, this critical mass is not zero. There are obviously situations where SDRAM _has_ outperformed RDRAM. The race, however, is closer with slower CPU speeds than with faster. And in case I somehow didn't make my point strong enough, the _theory_ only take hold after a critical mass of concurrent access is achieved. RDRAM has a massive drawback in terms of latency, so it's curve starts off with a handicap. The fact that it can outperform under extremely heavy-loads "suggests" ( here's that whole using evidence to prove, or shoud I say support a case ), that the trend should continue with ever faster machines.
The theory suggests that only server-class machines ( having multiple concurrent processes ) will fully take advantage of RDRAM. This is the whole Pentium Pro argument of 32 v.s. 16 bit code. The situation back then was not well suited for 32 bit code, and they just happened to suck at 16 bit code so benchmarks could really ream the PPro. All in all, however, it was a superior architecture. The concept of RDRAM is more advanced than SDRAM ( I think this can go unquestioned ). The real question is whether the drawbacks from the more complex technology outweight the advances IN A SERVER. I highly doubt Intel will not market RDRAM for laptops or value-PCs for this very reason. I am a very theory / principle based person, so I believe that the evidence supports my case, but I can not say this with any real amount of certainty because I have not seen benchmarks that test the theory properly.

Lastly, I fully admit that my theory biases me; I am not an authority on the matter. I am reading into some benchmarks. But again, this has nothing to do with the point in my previous article.

Re:Is RAM really all that expensive?

[astrophysics]

So far Rambus's agreements only give them ~0.5-3% royalties, depending on what type of memory and which manufactuer. So even if the royalties were charged off the retail price (I don't htink this is the case.), and they got 3% royalties on everything (this is an approximate upper bound), then the increase in memory costs due to Rambus enforcing it's patents would only be ~0.03$/MB.

Presently, both RDRAM and DDRRAM are very expensive. This isn't because of rambus taking lots of royalties. It's because the products are still new, people who buy early are willing to pay extra, they haven't become comodities yet, memory manufactuers still haven't matured their manufacturing and testing to the same extend they have for SDRAM, etc..

So I'm not worried about memory prices soaring due to Rambus. I'm confident that with time prices on RDRAM and DDRRAM will come down.

Big deal

[Reality Master 101]

Even if this manages to survive, all it means is that manufacturers will rush the next generation of memory without patent problems into computers. It's happened before, and it'll happen again. I seem to recall IBM tried to squeeze the industry with Microchannel and Apple tried to squeeze with Firewire. The industry will survive.

--

Is RAM really all that expensive?

[dorzak]

I can remember when a friend of mine purchased 1 meg of ram for $150 dollars. This was in 1991.

I can remember when it broke the $100/meg barrier.

I can remember when it broke the $50.meg barrier. I bought four megs of RAM.

Now that it is in the $1/meg range, I bought 128 megs

Thing about what it costs now, and will cost if Rambus succeeds. They will have their day in the spotlight, and then they will be replaced with the next technology.

If they succeed I see DDR staying in th $3-5/meg range for a while, but eventually it will have to come down to compete.

Where's the DOJ now?

[len(*jameson);]

If this isn't the clearest example of anti-comptetitive practices I've seen in recent years, then maybe I'm missing the point.

No doubt the DOJ will step in 5 years from now after Rambus barely exists and we're on to some new, non sucky, memory type and say "Hey - remember when you were screwing people, well now we're going to get you."

Thanks DOJ. For nuthin'.

Patents are e-vile.

[Black Parrot]

> 'I wouldn't argue with that conclusion.'

Ah, another case of patents promoting the Progress of Science and the useful Arts, eh?

--

Rambus, Patents, and Antitrust in the US

[doogieh]

Unfortunately, if the Rambus patent portfolio is valid, they can do whatever they want with "immunity" to US antitrut laws.

About a year ago, the US Federal Circuit (the US court with jurisdiction over patent appeals, answerable only to the US Supreme Court) decided that enforcement of a valid patent can never be an antitrust violation.

The decision was Intergraph v. Intel.

This means the best way to resolve this is finding prior art; one bit of prior art has been found over at technocrat.net.

-doug

Re:Capitalism and its bastard offspring

[malkavian]

Strictly speaking, this is all no longer capitalism. Capitalism is where the market forces prevail. Good, well marketed products succeed, where shoddy products don't grab the consumer heart, and so fail. Capitalism isn't such a bad thing really. It forces innovation (real innovation) and evolution. This, however is a corruption of capitalism. Law has been applied to stifle the growth aspect of the capitalist system. No longer do companies have to fight their competition for the best product, or the best angle on the product. They simply pay lawyers to 'nobble' the competition, and prevent them from competing. This nullifies the effect of market forces, as customers are no longer allowed to buy the best product from the person who finds the way to produce it at the most efficient cost. It all comes down to who has the most expensive lawyers that allows the public to buy what they think they want the public to buy. Other options become illegal (read 'unlicenced applications of patents'). Personally, I think capitalism is ok.. I think this bastardised hybrid is far from ok. A small dash of common sense and looking beyond the end of their noses and past sacrificing all for immediate gain would serve these companies in good stead. I can but keep my fingers crossed that they start to look this direction. Malk