Become a fan of Slashdot on Facebook

 


Forgot your password?
Close
typodupeerror
×
Technology

Samsung Claims World's First 288Mb Rambus DRAM 123

Posted by Roblimo from the more-memory-per-square-millimeter dept.
Hugo writes "Samsung Electronics announced the completion of development of a 288Mb Direct Rambus® DRAM (RDRAM) component and 576MB Rambus In-line Memory Module (RIMM?) Module. A design rule of 0.17-micron is used. The data processing speed at each pin has been improved to 800Mb per second, so the device can process the equivalent of 6,550 newspaper pages of information per second." Samsung does a *lot* of predicting about future RAM, but seems to be moving right along. Check this story from June '98.
This discussion has been archived. No new comments can be posted.

Samsung Claims World's First 288Mb Rambus DRAM More

Samsung Claims World's First 288Mb Rambus DRAM

Comments Filter:

  • YAFYAFSCWF2RDSS (Score:1)

    by Anonymous Coward
    Yet
    Another
    Fucking
    YAFSCWF2RDS
    Story
  • Not small, just small-minded...
  • San Francisco (TP) The Newspaper Industry, already beset by rising costs and an increasingly educated population, was rocked today by the announcement of a new chip which could put thousands out of work. The new Samsung "Rambus" chip will purportedly process in a single second what it takes a 12-year-old newsboy an entire morning to deliver.

    "I was hoping to go to MIT or Stanford with my earnings, but now it looks like I'll have to deliver twice as many newspapers every 18 months just to survive" said Tommy Poorhouse, a $12.97-a-day carrier for the St. Louis Dickensian. The RAMBUS chip, although expensive, is already expected to nearly halve the cost of moving newspapers, compared to wages for carriers like Tommy.

    "We've exempted ourselves from child labor laws; we've cut our journalistic standards to the bone, but we're still struggling" says Ken Garcia, anger columnist for the San Francisco Comical. He says many like himself are considering moving into the toilet tissue industry, which has so far resisted competition from both traditional newspapers and the burgeoning technology industry. "I think I can give my readers what they really need, probably in a two-ply" says Garcia of the move.


  • This is cute:

    Moderation Totals:Funny=4, Overrated=1, Total=5.

    Overrated ADDS to the score or something? word.
  • I've read the data sheets. There is a 256Mb non-parity part, and the 288Mb is the parity version. I guess the reason they don't mention is because a) most systems today - servers - that want rambus would want parity or b) 288 is bigger than 256, and hence sounds better in press release.

    Though I'm still unimpressed with rambus. If only it was as easy to engineer as Rambus Inc. says, that would at least be something.
  • The RAMBUS channel is 16 bits wide. Thus the maximum throughput is 1.6GB a second, twice normal SDRAM and the same as DDR.

    You are right, though, in that the latency can be much greater than normal SDRAM. The creators (Rambus Inc) apparently decided to tell people that latency wasn't important, bandwidth is all that mattered. Intel agreed. Too bad benchmarks disagree.

    Another thing to remember is that there is latency between the chipset and the processor in addition to the latency of the rambus channel. This gets worse in an SMP environment, where all the processors share the Front Side Bus (FSB) and access to memory.

    Motherboards have proven much more difficult to build than expected, again due to falacious reasoning on the part of Rambus Inc. They said that a 16-bit bus (27 total signals for the entire RAMBUS channel) is easier to build than 128 (+ whatever control/address signals there are for SDRAM). This is correct. Unfortunately, a 400-MHz double-pumped bus - with pico seconds of margin and a very high sensitivity to ISI, ringback, and coupling - is much harder to build than a 100-MHz bus.

    Another problem with rambus that I haven't seen publicized (probably because it doesn't matter much for desktop systems) is that there is a limit on size. One channel can have, as a hard limit of the architecture, 32 RDRAM devices. This means that 32*256Mb = 1GB max(optionaly with parity).

    To get more than this, you need a repeater that translates the information from a single channel onto two channels, for a max of 2GB. Two repeaters on a channel gives you 4GB max. This adds cost, even more latency, and many troubles for the poor engineers having to make motherboards.

    I've heard (but can't confirm) that the Alpha 21364 is going to get around some of these problems by having as many as 4 RAMBUS Asic Cells (RAC - the chipset-side controller for a RAMBUS channel) on-chip. This eliminates the latency of the chipset and gives each processor direct access to 4 GB of RAM. Notice that having 4 SDRAM channels on-chip would be infeasible with a 128-bit bus, which I guess makes this a reasonable use for RAMBUS.



  • Is that, like, meta-fucking?

    I had a girlfriend who was into that -- and boy was I glad I meta!
  • 288M / 9 = 32M

    Silly idea, should be using SECDED instead.
  • i had an ex who was late once and it scared [...] me.

    I have a Mrs who was late once and now we have a happy little baby boy. It still scared me, though. I hear it gets better around the fifth or sixth child.
  • 16 bits x 288Mb = 4608 megaBITs (Mb).

    4608Mb / 8 bits-per-byte = 576 megaBYTEs (MB).
  • I see motherboard support for this 800mhz rdrams, but pricewatch doesnt even list prices for it!
    Searching the net for a prices has been harder than finding sblive linux driver sourcecode. ;)
    Also cant find benchmarks for motherboards running normal pc133 vs. 800mhz rdrams.
  • Just an observation, but it's this interest in minute details and incremental updates that even some geeks find unattractive in other geeks. Repeat after me: Optical computing achieved - newsworthy, another 50MHz added onto CPU speed - dull. (Optical computing achieved - newswort... ) Oh never mind :)
  • If you assemble 8 bytes (with parity) into a 64 bit word, you can use SECDED (single error correct, double error detect) ECC on the 64 bit word. 8 check bits are needed for ECC on a 64 bit word.
  • 8 x 288Mb = 288MB.
    (Eight bits in a byte).

    Therefore, 16 x 288Mb = 576MB.

    Hope this helps.
  • ...but in the US, a newspaper page's worth of information fits in about 1k. :)
  • The REAL question is if it's KOREAN newspaper pages, or American! That tricky Korean alphabet stacks characters vertically in one space to 'overload' sounds and increase information density.



  • Another difference.

    That 1970's vette, while it may not be quite as fast as some of the newer Vette's, with some tweaking can get some good runs at the track. With computers, in 5-10 years, you may as well scrap them, unless you give the interior of the case an enima (sp?)....

    Also, the 1970 'vette is more valuable to people than a 1970's computer (okay, well most people...) Also, the 1970's vette can still do what people need it to do -- get them from point a to point b in a reasonable/acceptable amount of time.

    Try computing with a 1970's computer with today's expectations. Most everyone would not really like that. But, going 120+mph in a car can be done whether it's a 1970's vette or a 2000 vette.

    Just my offtopic .02$
  • My A1200 w/ 68030 + 4MB "fast" ram was 60ns. Standard A500 (Kickstart -1.2) was 120ns. A600 and newer A500+'s had 80ns RAM.

    To answer your question: some RAM has ultra fast latency (graphics card RAM is 6ns or so), SDRAM though probably has high latency times like you said.

    Don't ask me what, try a RAM manufactures website for details.
  • now we need a quad 1GHz Athlon system with dual channel DDR PC133. 4GB/s.
  • Nate,
    The reason his post was funny is that DRDRAM is so darn expensive. Get it? You'd have to give an arm and a leg in order to get any?

    Ha ha?

    He he?

    Anything?

    Is this thing on?
  • Overrated=1 means that one moderator thought your comment was overrated. The "Overrated" rating adds nothing to the comment's score. I would guess that your comment was rated a 1 initially just as the comment I replied to was.

  • 8b = 1B ... It's true that sometimes, mostly in non technical articles, you're not totally sure whether they mean bit or byte...

    Except for us seasoned veterans of the game-console wars, who are all-too-familiar with the heyday of cartridges whose sizes were listed in megabits (Mb) instead of megabytes (MB) to make them seem more impressive. Bleah.

  • Thaz pretty cute. Caracter is spelled 'character'.

    And that sort of indicates you are prawly not american. That would also explain the comment 'these us newspapers must be very small'.

    I'm not american either, but Imperator says that the information fits in 1K. That does not mean that the actual articles occupy 1K.

    A little example to explain the difference: When I say 'the earth is a globe' one could say that that occupied 20 characters (including whitespace). But really it's no information to you because you allready knew that (presumably ;o))

    Only reading (viewing!) an actual american newspaper makes this perfectly clear. You should only feel blessed that you didn't have to experience this first hand. :)

    Breace.

  • 0.17 micro is small? RAM chips are usually at least 1 process generation ahead of CPUs. You should be asking why they aren't using 0.13 micron for this.

    Samsung is currently sampling Alphas in 0.18 micron.



  • For each 9-bit chunk of parity memory, one bit is the parity of the other eight. That means it'll be true if there are an even number of 1s among the other 8 bits, and false if there are an odd number of 1s.

    The advantage to storing the parity is if any of the 8 bits is unreadable, its value can be recovered given the 7 other bits. Simple error correction.

  • Sixteen COMPONENTS not MODULES.

    I presume the 288 guy has 8 such "components".
  • I don't know the full details on how Rambus works, but I believe that what happens is:

    A signal is sent down the memory bus.
    This signal travels through ALL of the chips sitting in RIMM slots.
    The signal bounces back.
    The RIMM modules do something to send a signal back.

    Now, if my understanding of this is correct (which I am not certain of, caveat emper) only one "pin" on each chip can transmit at a time...

    At least, that's my guess...
  • Oh please. They don't sell these things directly to Joe Public, and even if they did, if you bought one for your non-Intel, non-parity based system, it probably wouldn't even fit in the slot.
  • The point is, it may look better to you, but not to anyone who a) knows what they are talking about and b) makes purchasing decisions in this area. RAM is always quoted in bits. What are the poor guys to do, quote it as if it only had 8 bits per byte, to level the playing field? I don't think so. Give them a break...
  • newspaper pages per second is about the most ass-backward bandwidth unit I have ever heard...
  • "most people actually used their real names as their user names. none of that gay Signal 11, foogle, effuegas crap"

    Gee, then would you posting as an AC make you both gay and unoriginal? Or just gay and scared to write under your own name?

  • I thought that .25 was the basic limit of standard optical lithography, and that by adding new high-tech focusing lenses, we could get the beam down to to .18 micron. After that, X-Ray lithography was necessary. Anyone know what kind of fabrication they used?
  • Actually, you can't store 288 Mb, you can store a total of 256 Mb. The other 32Mb is being used for error checking/correcting.
  • he said "given the 7 other bits"... in that case you could do it. But you are never given which bit is broken, so you are right too.
    -
  • Firstly, the size. As many people have pointed out, the 288Mb unit uses parity, thus using a ninth parity bit per byte. So the effective size is 256M. Of course, Samsung will try to push it as a 288M chip to seem like it's bigger than 256M... stick 8 of them together and you could get a 512M DIMM. Nice.

    As for the speed, the stated bandwidth is 800Mb/s *on each pin*. Can someone clarify what this means? Does that mean total bandwidth is about 800M * 130? or what?

    My last question regards the frequency - how does this chip relate to the 200Mhz/400Mhz or higher rambuses we're seeing? Would the Athlon have better performance than the PIII as a result?

    Inquiring minds want to know!

    Fross
  • At least they also included the real information. I'm really sick of every numerical quantity being converted to football fields per novel.

    I'm a little surprised at how large their newspaper pages are. But then, the figure is probably entirely fictitious anyway.
  • You are crazy! Windows 2000 takes AT LEASY 1 gig! Once you got that thing booted up you coulden't do anything with it. Just look at your ugly desktop, mayby use the calculator. Then wait for the blue screen. And you would be happy because you did it on a fast file system.. Sounds like a plan to me. Can I come over and help you time the how long it takes for the Blue Screen of death to pop up? I will bring my stop watch (It measures time in 1000ths of a second, and I have fast reflexes.. We may be able to get a fairly accurate reading)..

    OK, enough of that.

  • Maybe I am mistaken, but I think the expansion ram for my old Amiga 500 had 60ns RAM. If not 60, then 80. Has latency not improved at all? Or is it a result of these special burst/high bandwidth designs such as SDRAM and RDRAM, that latencies are high, which the L1/L2 making up for it?
  • Thanks for the info. I wasn't sure about it, but I thought I had heard the figure 60ns tossed around in an Amiga context. I actually owned a 1200 aswell later on...
    The point though is how little progress has been made in RAM latency. Density goes up, but not the response time. Obviously this is why caches dominate modern CPUs, but perhaps there is a great bottleneck there which will worsen in the future?
  • Whatever happened to 256 and 512? Why are they not basing memory sizes on powers of 2 anymore like they have for as long as I can remeber?
  • Simple: 8b = 1B, so 16x288Mb = 576MB.

    It's true that sometimes, mostly in non technical articles, you're not totally sure whether they mean bit or byte...
  • Because they are ECC

    No - they aren't ECC! ECC can self correct one bit errors, single bit parity RAM (which this is) can't.
  • If I had moderator on this article you'd be on your way to +4. =) Nice.


    Pablo Nevares, "the freshmaker".
  • mb would be millibits.
  • Actually, 0.17um for a memory isn't that odd. Usually the feature size (like 0.18um or 0.12um) is defined as the minimum width of a transistor in a CMOS process, which is what is usually used as a benchmark. Most likely this odd size has to do with optimizations they were able to do because they are building something akin to a DRAM, which is not exactly CMOS. Robert
  • hmmm get a couple of these together, and set up a RAMDrive then you have a new definition of performance filesystem.
    I bet with 1GB of this RAM Win2K would still take about 1/2 hour to boot. :)
  • yeah that sounds nice..... I think W2K will finish booting by the time your plane arrives.
  • actually they should have written it mb instead of Mb - megabits instead of megabytes.

    288 megabits * 8 bits = 288 megabytes

    x2 = 576

    My guess for the wierd amounts is that they are building in parity.. 288 / 9 = 32

  • I thought it was so funny I laughed out loud. To each his own, but don't go trashing other people's sense of humor just cause you don't get it, k?


    If you can't figure out how to mail me, don't.
  • hence my many disclaimers.
    If you can't figure out how to mail me, don't.

  • But you missed the point - that kind of certification is not necessary in a large open-sourced project such as linux. The certification that occurs is the peer review of the driver team and coders. Such certification is only necessary for those companies that insist on a closed-source driver.

    And... well, if it doesn't work, it's just more work for them. And linux junkies will just go buy something that works. I've done that myself.

    So honestly I think that's a moot point...


    If you can't figure out how to mail me, don't.

  • Since you are commenting on the comments, has anyone noticed how damned FUNNY some of those "trolling" comments are in this article? Some of those responses to the trolls have almost had me falling off my chair laughing...

    Not to encourage them or anything... still, I couldn't help but notice...


    If you can't figure out how to mail me, don't.

  • shouldn't that be:

    # npu

    for Newspaper Pages Used?

    Maybe we could get an open source programmer to write the nputils kit... with:

    NPU - newspaper pages used
    DU - DU but in newspaper pages
    FREE - free but in newspaper pages

    Wouldn't we also have to rewrite the /proc filesystem code?

    Offtopic, while I write this, I'm watching an ancient "I've got a secret" and there is a whole line of fat guys in ballet suits dancing ballet to the "nurcracker suite"... talk about surreal...


    If you can't figure out how to mail me, don't.
  • I could wish it'd be the same way as well, but there are some differences:

    Different makes of cars require different parts. It is very rare, (although I'm not a mechanic) for two different cars from two different manufacturers to accept the same parts. Computers are not this way - in that most mainboards that are made in roughly the same time period take similar equipment.

    Also, a part in a car, because it is moving, is much more susceptible to wear than the computer parts are... this makes them in more constant demand and there's much less financial incentive for car companies to cut the prices on them. In fact, I've noticed that most parts get more expensive the older the car gets. Just because the car gets older doesn't mean the part's not going to be in demand.

    Back to your original point, I wish it were the same, too. Then I could buy a 1970 'vette for about $50.

    Oh well, enough of my blabbing for now. This is more appropriate than ever, but... your mileage may vary. I could be completely wrong.


    If you can't figure out how to mail me, don't.
  • heh heh, you said RIMM.

    Thats the coolest thing I've ever seen, heh heh

  • i know this is ot, so feel free to moderate

    but i damn near pissed myself laughing at this one.

    incidentally, there arent even any pics on that page

    sad pathetic spam. didnt even ask me for a credit card
  • to make us have to memorize yet ANOTHER random and situationally ambiguous set of numbers

    im tellin ya, its a conspiracy

    PCMCIA - people cant memorize computer industry acronymns

    *cry*
  • Serial is for modems and mice, not memory damnit!

    It's also for networks (Ethernet, including Gigabit Ethernet, ATM, SONET, etc...), high performance disks (Fibre Channel), digital video (Firewire, even VGA video ports), and probably many other things.

  • Cutler and the people he worked with at dec (about a 100 engineers) moved to MS, correct. Ofcourse Cutlers ideas about solving problem X and Y didn't change, nor did the ideas of his staff change about how to implement it. however, it's a long shot to mark NT as 'VMS derived material'. It's a new OS, with some solutions in it which are the same as implemented in VMS, and in other OS-es perhaps. VMS is not that unique :)

    Off topic ofcourse, I know, but I thougth I'd to put this into the discussionthread. :) the rumour you touch with your texts is as much true as marking apple as inventor of the gui with a mouse and windows.
  • The problem with the crappy drivers has one advantage (yes it has :): the reason why crappy drivers still can work in the system is also the reason why there are so many drivers for so many hardware/software. You see: MS has a tough testlab and they certify every driver they can (you should always use only certified drivers to be safe, if you don't do that in a production environment it's your fault :)) and can't certify every driver in a short period of time. Because a company is still able to supply 'a' driver to bridge that time, more hardware works on windows than on any other OS that only wants to have certified drivers or drivers which are granted permission by a committee. Win2K breaks this 'run it if you dare' syndrome and only allowes certified drivers to run. Perhaps we all get wise in the end ;)
  • Yes, computer scientists have frequently used newspaper pages as a unit of measure to provide laymen with a tangible feel for storage capacity. Generally one newspaper page is exactly equal to the number of characters on the latest New York Times. When referring to USA Today however, the unit is specified as "deep" pages to account for the additional color bits.

    For "historical reasons" the df command counts bilevel front and back pages and the resulting display can be very misleading. On my old fashioned system I need to remember to use df with the -k argument to convert to full pages (ie. front plus back). Many newer operating systems do this by default.

    # df -k
    total newspaper pages: 362790
    newspaper pages used: 7944
    newspaper pages free: 354856
  • um just incase you missed the first 40 answers 8*288Mb= 288MB 288MB*= 576MB hope this clears up all of the confusion
  • Yells the computer store clerk when this things hit the street and regular SDRAM prices take another dive.

    Ahh if only the new BMW would do _that_ to the Corolla I can *almost* afford.

  • From what I have seen, the Windows kernel is VERY stable and fast. However, it's basically an extensible design, like a micro-kernel architecture, that is expected to remain stable with a variety of extensions by other companies, mostly in the form of drivers, CODECs, etc. My opinion is that's where Windows stability problems come from. Linux has an advantage here because of the open source aspect and because "core" Linux drivers (i.e. those that get distributed with the mainstream kernel source) tend to be looked at by very knowledgable individuals before being circulated. In the Windows world, anyone can deliver a driver, DLL, what have you and, since the source is never released, many common mistakes may never get fixed, or get fixed only after many users have suffered with the problem.

    Just $0.02 from a MCSD (since 1996).

  • Anyone have a technical reason why Samsung is using 288MB instead of 256MB? Looking at their press release doesn't help. I am not familiar enough with RAMBUS to know more...

    0.17 micron seems small. Can they expect to produce enough of this? I thought I read that there are only a few fab plants that can produce 0.18 micron chips - I imagine it must be even smaller 0.17 micron chips.

    RAMBUS - vaporware (or will I be shocked?).

    -- Justin
  • Maybe they're talking about the advertising section as well. Takes more space to scan in and save all those pictures of brassiere models.
  • From the article:

    "Sixteen 288Mb Direct Rambus® DRAM components can be configured together to create a 576MB module."

    Can someone please explain to me how 16 x 288 = 576? Thanks!
  • Missed that darn little "b". Thanks for pointing that out to all who replied.
  • Parity... Basicly, you take the least significant bit of the sum of the other 8 bits (or, to put it differently, you add the 8 data bits together, if it's even the parity bit is 0, else it's 1), and dump it in the parity bit.


    >So they are marketting it as 288 when it can only store 256 of real data?

    Of course they are...
  • Yes, but still, this press release meant to let the public know: look, we've made really cool ram. And they're making it look better than it is by doing it in bits.
  • How many bytes of data flow over Niagara Falls in a second? Now THAT's memory bandwidth!

    Would an Empire State Building worth of magnetic core memory come close to one of these RIMMs?

    If you outfitted pirahna with this RAM, would they be able to skeletonize a cow in less than a millifortnight?

    For that matter, how many picoY2Ks of hype does this press release contain?
  • My guess is, it's parity RAM. Eight bits of storage space getting one parity bit makes 256MB + 32MB = 288MB.
  • I wasn't aware that a "newspaper page" was a standard unit of measuring data:

    # df
    total newspaper pages: 725580
    newspaper pages used: 15887
    newspaper pages free: 709693

    --- Dirtside | "Spirituality" is the irrational belief in the supernatural

  • So that means that technically, they can only store 288 * 8/9 = 256 MB of data, and that 32 MB are there just as "checksum" or something? (How exactly does this work?)

    So they are marketting it as 288 when it can only store 256 of real data?

  • OK, this was my understanding: (keep in mind this is what I heard via hearsay and it may bear little to not resemblance to the truth!)

    Sometime ago, the chief coder of VMS, a Dave Cutler (I think) got hired away from DEC to be the chief coder for WNT (notice that WNT is one character away from VMS). It was also my understanding that along with the coder came a bunch of the code for VMS (this is what I heard). They used it in the kernel for the first Windows NT release. But somehow they got caught - they even kept the *comments*.

    So DEC struck a deal with Microsoft - you can keep using our technology *if* you port NT to Alpha.

    Hence a very stable kernel, because it was essentially VMS.

    Then, when 4.0 was released, some yahoos over at MS had the bright idea to start tinkering with the kernel - moving graphics calls into kernel space. IMO, I don't care what the performance gains are, that's a really BAD IDEA. But it was my understanding that the speed gains this created were offset by the instability this change introduced into the kernel. It is a documented fact that MS has been moving many things that arguably shouldn't be there into the kernel, and destabilizing it in the process. A miswrote graphics call shouldn't crash the kernel. But I digress...

    Anyway, this was what I heard and my understanding. Someone knowledgable feel free to correct me. The point I'm trying to make though is that arguably MS *has* been destabilizing their kernel in order to attain nebulous performance gains. I have no doubt that the MS kernel is *reasonably* stable - read just fine for desktop or light load use. But I don't and wouldn't trust it with my mission critical stuff...

    Anyway, this has ended up offtopic...


    If you can't figure out how to mail me, don't.

  • I personally don't build kernels to be small, I build them to be functional. I really dislike modular code and on "production" machines I even build the kernels without module support at all. It just makes things easier.

    That being said, I bet you could build them damned small. But you know, I bet the core kernel code for windows NT is small too. It's the stuff they build around it that sucks.

    Windows NT, from what I understand, is basically a modified VMS kernel, at least in 3.1 or 3.51. Notice how the kernel destabilized the further away you got from that release? As I understand it, it's because the Windows coders started to tinker around in the kernel and destabilized it.

    Just my $.02


    If you can't figure out how to mail me, don't.

  • What I don't like about rambus is how they seem to be pushing it not on its merits but because it'll be good for their (Intel's) bottom line. I also don't like the thought of paying at least 4 times as much per megabyte (though that would probably equalize). I like the SDRAM just fine, thanks.

    That being said, it's very heartening that it seems to have almost no support in the industry at this point. Although putting 288mb on a chip isn't a small achievement...


    If you can't figure out how to mail me, don't.

  • You've got it right -- latency vs throughput (Score:3)

    by Anonymous Coward on Sunday January 09, 2000 @09:48PM (#1388105)

    While PC100 SDRAM and 800MHz single-channel RAMBUS have the same theoretical peak rates of throughput, RAMBUS appears to be better at sustaining higher real-life throughput levels. On the other hand, SDRAM's latency is lower -- sometimes *much* lower (60ns vs 120ns) -- and for a large class of applications main memory random-access latency is a more important performance limitation than memory throughput. Just to be fair, there are also many applications (mostly floatingpoint-intensive workstation applications) which benefit more from high throughput than from low latency.
    &nbsp
    On the gripping hand, the growing disparity between the operational rates of memory busses and microprocessors is making latency increasingly important. In the next two years the time spent filling cache from main memory will comprise a significant fraction of total runtime. This paper is several years old, but thusfar the industry is proceeding according to the schedule predicted by its authors:
    The Memory Wall [flyingcroc.net]

    One thing to keep in mind when reading this paper is that SDRAM's low sequential access latencies are equivalent, in the context of this paper, to a form of caching. Think of it as extra caching going on in the memory module itself, and use 60ns as the latency of a main memory access (ie, a cache miss).

    One thing that RAMBUS is supposed to do (but thusfar has not) is make processors and motherboards easier and less expensive to build. It is much, much easier to get an 8-bit-wide bus working correctly than a 128-bit-wide bus, and chip manufacturers are having to come up with new and gruesome ways of sinking more IO pins into their products, which is a costly pain in the ass. The PCB real estate consumed by a motherboard's 128-bit-wide bus isn't cheap, either. Whether RAMBUS will ever succeed in delivering on this particular promise remains to be seen.

    -- Guges --

  • No. It's a scam. (Score:3)

    by Wakko Warner ( 324 ) on Sunday January 09, 2000 @08:55PM (#1388106) Homepage Journal
    It's 8 bits wide instead of the 64 bit width you get for normal SDRAM. RAMBUS can only hit 800 MB/sec, the same throughput as PC100 SDRAM. PC133 SDRAM has it beat by, of course, 33%. Real-world testing puts it either as fast as, slightly slower than, or slightly faster than SDRAM, depending on which site you go to and what day you visit it. Regardless, you won't see much difference at all, but you'll sure pay a lot.

    Don't even start comparing it to DDRAM (double-rate DRAM), which will be appearing soon and which can be produced on the same production line as SDRAM is with minimal retooling. This means, of course, manufacturers will love it.

    RAMBUS is, plain and simple, a scam.

    - A.P.
    --


    "One World, one Web, one Program" - Microsoft promotional ad

  • RIMM? (Score:3)

    by A4Joy ( 54907 ) on Sunday January 09, 2000 @07:16PM (#1388107)
    That cracks me up. 'Well, I sent the motherboards in for an upgrade last week and we're still waiting on the RIMM job.'

    Tasteless, yeah, I know... I am so ashamed of myself.

    ---
    Tempfiles fugit.

  • This just in.. (Score:4)

    by myconid ( 5642 ) on Sunday January 09, 2000 @07:00PM (#1388108) Homepage
    This just in:
    Samsung has began a new program to help American's afford their new 288Mb Rambus DRAM. We have been informed that Samsung is taking "trade ins" of whole human appendages and key organs. They say that for an arm, or a leg, they will trade a single 288Mb RDRAM. For any major organ, 2 288Mb RDRAM, and for essentail organs, 4 288Mb RDRAM modules.

  • Re:Why the weird memory size? (Score:4)

    by RevRigel ( 90335 ) on Sunday January 09, 2000 @07:12PM (#1388109)
    My guess would be that it's parity ram. 256 = 32 * 8. 288 = 32 * 9. Extra bit for parity checking and all.

  • Ram technology review (Score:5)

    by Scott McGuire ( 4080 ) on Sunday January 09, 2000 @07:44PM (#1388110)
    Ars Technica [arstechnica.com] had this link [techrepublic.com] to an article about ram technologies. RAMBUS didn't look so good there.

Slashdot Top Deals

Single tasking: Just Say No.

Close