Want to read Slashdot from your mobile device? Point it at m.slashdot.org and keep reading!


Forgot your password?
DEAL: For $25 - Add A Second Phone Number To Your Smartphone for life! Use promo code SLASHDOT25. Also, Slashdot's Facebook page has a chat bot now. Message it for stories and more. Check out the new SourceForge HTML5 Internet speed test! ×

Single-Chip DIMM To Replace Big Sticks of RAM 100

MrSeb writes "Invensas, a subsidiary of chip microelectronics company Tessera, has discovered a way of stacking multiple DRAM chips on top of each other. This process, called multi-die face-down packaging, or xFD for short, massively increases memory density, reduces power consumption, and should pave the way for faster and more efficient memory chips. Multi-die face-down packaging is exactly what it sounds like, with memory dies stacked on top of each other like roofing tiles. Much like a normal desktop DIMMs and laptop SO-DIMMs, each of the stacked dies is wired to each other in series — but in this case, the connections are much shorter, as they only have to run a few micrometers to the chip below it. This is where all of the power and speed enhancements come from: shorter interconnects mean less power is needed (and thus less heat is dissipated) and signals propagate faster."
This discussion has been archived. No new comments can be posted.

Single-Chip DIMM To Replace Big Sticks of RAM

Comments Filter:
  • by Anonymous Coward

    Are these still considered DIMMs?

  • For one horrible moment of puke-inducing fear, I thought you wrote "Invensys".
    One letter makes a big difference, sometimes.
  • by TheReaperD ( 937405 ) on Wednesday September 07, 2011 @05:00PM (#37332786)

    The questions is, will the patent fees be reasonable enough that we will see this technology for less than $200 a DIMM?

    • by Baloroth ( 2370816 ) on Wednesday September 07, 2011 @05:11PM (#37332970)

      Patent fees? Why would Tessera charge itself patent fees? I think you have been staring at software patents too long.

      They may or may not license this to other companies, and once they start building them they will have to have low enough prices to be competitive with existing DRAM technology. The world of hardware is not quite like the software world where companies routinely submarine others in areas they often don't even make product. In hardware, you can patent an excellent technology, but you either have to build it yourself or license it for affordable rates to actually make money off it. Unlike software where you can look at someone else's product, patent it, then sue their asses off and get a settlement. AFAIK that has never worked in hardware (it probably has, but it is certainly much, much rarer.)

      • by TheReaperD ( 937405 ) on Wednesday September 07, 2011 @05:30PM (#37333206)

        Though I have been following software patents closely, it has no bearing on my question/comment. One of two things will happen, they will either license this to other vendors for a fee where they can manufacture it or they will not and only build them themselves. If they license it, they can charge a fee that is either reasonable or exorbitant. If they build it in house they can charge whatever they want. Though either option is their right, I, as a consumer, would like to see this product come to the consumer market at a reasonable price, thus my question/comment.

        And no, they do not have to be price competitive to make a profit. This has been proven many times over. Since they have, according to TFA, a superior product, they have the option of producing it in low volume and charge a high price for the high end server and gamer market. If you insist on a citation, just look at Apple. They produce in low volume, charge a high fee and make a large profit because their customers believe they make a superior product. And I don't have to agree with it for the last sentence to be true.

        This invention means jack to me, as a consumer, if they take the Apple route. Thus, my original comment.

        • by Nutria ( 679911 )

          If you insist on a citation, just look at Apple. They produce in low volume, charge a high fee and make a large profit because their customers believe they make a superior product.

          Like the world-conquering FireWire?

          Oh, wait. That's a dismal failure because Apple charged too-expensive license fees so the technically way inferior USB conquered the world...

          • by dgatwood ( 11270 ) on Wednesday September 07, 2011 @06:35PM (#37333898) Homepage Journal

            Wasn't the licensing fees. They were never all that expensive.

            USB is so ubiquitous in large part because the silicon for USB devices is much, much simpler, and thus much, much cheaper. USB devices can be dumb as a post, whereas FireWire devices have to actually understand a lot more about the bus topology, etc., IIRC.

            Also, there's no such thing as a slow FireWire bus. S100 is the bottom limit. Therefore, it isn't a great match for really trivial devices like mice and keyboards.

            Also, Intel supported USB very quickly, and drug their heels on FireWire until... well, I'm not sure if they've ever shipped a southbridge with integrated FireWire.... So for computer manufacturers, FireWire was an extra part that they had to pay for, not just an extra connector.

            And there were no doubt other factors. I'm not convinced that the licensing was a significant one, though. By 2001, it was something on the order of a quarter per device. I think that's less than a tenth what the actual silicon costs. Even back when it was a dollar per port, it was still a tiny cost compared with the silicon.

        • Why would they, though? Apple doesn't charge high prices (and low volume? Apple does decidedly none-low volume runs for most of their products) because they are targeting high-end markets with superior performing equipment. No, they generally target middle-range consumers looking for stuff that works and is easy to use. High end consumers and low end consumers go elsewhere. They can do this because of a whole host of reasons, but it has little to do with a superior product. Generally speaking, if they can m

      • by spauldo ( 118058 )

        Rambus [wikipedia.org] did the hardware patent troll thing. It was all over /. back in the day.

        Short story, Rambus was a member of an industry group designing the new RAM chips (SDRAM, pentium II and III era). The new designs used technlology they had patented, but this wasn't a big deal since all members of the group were supposed to license any applicable patents they held under "reasonable" terms.

        Rambus didn't like that, so they pulled out and started suing everyone who made SDRAM. Intel had started using Rambus memo

      • Tessera makes all its money from patents on packages, it doesn't make the DIMMs themselves. They create these patents, license them and the licensees, such as a Samsung, Crucial, Hynix, et al would made the DIMMs, and others. They have some pretty impressive package technologies
    • This is one of those times when prior art is an absolute motherfucker. Back in the XT days (or was it the actual PC days?) you could double your memory on some platforms (details hazy) through the same technique. Apparently there were parts available such that the address and select lines would work out and you could bank select the stacked set of chips in software. You just stacked the DIP chips on top of the other ones and soldered them on. I've only seen it once but when I saw it the old DOS guy who cack

      • by Chas ( 5144 )

        This is how VisionTek started out.

        They'd buy old memory from people disassemble it and reassemble it into larger DIMMs through stacking. Actually bought some of it from them back in the day.

      • by hawk ( 1151 )

        That predates the XT by a couple of generations.

        The Model I TRS-80, for example, had only 7 bits of video memory. To conver to upper case, you glued another 2102 on top of one, soldered 14 of the pins to the chip below, and ran the other two by wire.

        There was anothernhack, iirc, that did this with even two extra chips on top of each main ram chip, allowing expansion from 16k to 48k in the main unit rather than an external "expansion interface." (yes, odd as it sounds today, the extra ram was close to a foo

        • There was anothernhack, iirc, that did this with even two extra chips on top of each main ram chip, allowing expansion from 16k to 48k in the main unit rather than an external "expansion interface." (yes, odd as it sounds today, the extra ram was close to a foot of wire away).

          Well, again I only go back to the PC era, but it doesn't sound strange to me because I had an IBM PC-1 (yes, really, I am not confused) with 64k onboard and 384k on an 8-bit AST ISA card, where the RTC also lived.

      • by mirix ( 1649853 )

        Yeah, this goes back to the dawn of time. You can stack parallel interfaced SRAM and ROM because everything is common ... /RD /WR and all the addr and data lines. You just had to separate off the chip select line... double the ram or rom instantly, assuming you had enough address space. I presume DRAM would be similarly packaged, plus the refresh lines, I've not personally stacked it though.

      • by cbope ( 130292 )

        I can confirm that. I had an old PC-AT motherboard I bought at a swap meet back in the late 80's, in fact I believe it was probably from one of the original IBM PC-AT models. It had a 6MHz 286 cpu and the memory, which was in DIP packages back then (individual chips), was double-stacked 2 to a socket. I don't remember how much memory it had, but damn it was a lot of chips when they were double-stacked like that.

    • The questions is, will the patent fees be reasonable enough that we will see this technology for less than $200 a DIMM?

      The monopoly granted by a patent doesn't mean what you think it means.

  • I can't wait. Past that.... what is there to say?
    • by DamonHD ( 794830 )


      • by jd ( 1658 )

        If you just have two dies and align them vertically rather than flat on the PCB, you've got the same cooling surface as you would have with two independent chips. Beyond that, you'd need to interleave the dies with a heatsink and then you're in for all kinds of funky problems. Surely it would be better to increase the total size of the die whilst keeping the same resolution (since you're eliminating even more connections and any supporting bits and pieces and can also exploit immediately any improvements in

        • They're not talking about orienting the chips vertically (e.g. ZIP packaging [wikipedia.org]), but stacking two (or more) chips flat against each other horizontally. This may make a slightly thicker package, but since many of the drivers, buffers, & latches for the external connection can be shared, it doesn't need 2x the power/heat. Silicon is a good conductor, so dissipating the extra heat shouldn't be a major issue. The key will be maintaining good thermal conductivity between the two chips and between the upper chi

  • by jandrese ( 485 ) <kensama@vt.edu> on Wednesday September 07, 2011 @05:02PM (#37332824) Homepage Journal
    The problem with stacked chips like this in the past has been cooling the wafers in the middle of the stack. While DIMMs don't run as hot as processors or GPUs, this is still a concern for them. I wonder how they're going to handle this? Or are they only going to target low power low performance parts?
    • by Anonymous Coward

      Plus, you have to remember less power consumption overall as well.
      To what extent is another question, but it might just be enough to not need much heating.

      It probably wouldn't be hard to add heatsinks to it anyway, will increase complexity ever so slightly, but for extreme high-performance RAM, worth it so your RAM doesn't actually somehow explode.
      Besides that, the modules themselves are really close, on the micrometre scales at that. A single heatsink on top would probably still be enough for it.

      • by rossdee ( 243626 )

        Just keep the ambient temperature below absolute zero and you'll have no problems.

        • Pfft, is that all?

        • "His dart throwers had been sealed and 'washed' against snoopers, then maintained at minus 340 Kelvin in a radiation bath for five SY to make them proof against snoopers." -- Frank Herbert Heretics of Dune

          Why, that's even colder than the null-entropy bin in my Harkonen No-Globe!
        • Just keep the ambient temperature below absolute zero and you'll have no problems.
          Would this be i-energy?
    • Or are they only going to target low power low performance parts?

      Like mobile? I didn't RTFA, but that's the area that seems to me to be screaming for this. With all the crap we're running on phones lately, we're going to need more memory. You can't exactly stick a SO-DIMM in a phone.

      • by Artraze ( 600366 )

        They mention mobile, but that's not terribly interesting: Someone (TI, IIRC) has had stackable memory for a while. In particular RAM and Flash that can be soldered directly onto their CPU (though I'm not sure how many of either it supports). That saves routing/board design costs and can make the overall device smaller. There's not too much point in having a stack of RAM elsewhere as you're probably only going to have 2 chips at the most... Current densities are 1GB/chip, so unless you're looking for 2+G

        • by ajlitt ( 19055 )

          Samsung (and Micron I think) sell a multi-chip BGA with flash and DRAM stacked in the same way. Some of these models are meant to fit on top of an SoC like Samsung's Hummingbird or TI's OMAP in a scheme called PoP [wikipedia.org]

        • If you design the chips for this purpose, you can share many of the drivers, buffers, and latches for the external connection, thus lowering the power consumption. Also, this won't materially lengthen any of the connections. Half as many packages and ~ half as many driver circuits for a given capacity should produce a notable power savings, even though each package will draw slightly more power than a single chip package would. Or, you could get 2x the capacity with a less than 2x increase in power consumpt

    • by Mashiki ( 184564 )

      Well in the past with computers, and supercomputers the only way to get around this problem is by using immersion in an inert liquid which can properly draw heat. Plenty of current DIMMs require head spreaders once you cross into PC3-10666, one of the few ways around it is to lower the input voltage, but even then they can get toasty.

      I'm using some g.skill eco(1.3v) in my system simply to keep the temperature down but under load those things will still hit around 64C with proper cooling across the spreader

    • The problem with stacked chips like this in the past has been cooling the wafers in the middle of the stack. While DIMMs don't run as hot as processors or GPUs, this is still a concern for them.

      True, however shorter wires can be driven with less power, creating less heat.

    • But Sir, it is only wafer thin!
  • Seems like I've seen this article a half dozen times over my career, and nothing ever comes of it. Usually by the time they get the bugs worked out a higher density generation of RAM comes along and the stacked wafers can't compete on price.
    • by ajlitt ( 19055 )

      Except by now flash manufacturers have the stacked die process down pat, fitting many geebees in a single BGA. Presumably this is using the same manufacturing process, using bond wires on one edge of the stagger to connect to the substrate.

  • by vlm ( 69642 ) on Wednesday September 07, 2011 @05:09PM (#37332930)

    The marketing release implies most of the power is being dropped resistively in the leads instead of in the dies. Just doesn't work that way.

    Think about it for a second... The voltage on the die is only a tiny bit less than the voltage on the bus... You know the bus impedance too so that gives away current flow. Do a little ohms law on that tiny little drop and the tiny little current and compare it to what the die drops.

    Or look at it from a thermal engineering perspective... they put heatsinks on the dies, not on the leads...

    Now there will be some savings, probably lower capacitance and inductance and all that makes life easier for the bus drivers. But you're still gonna roast the dies in the middle of the sandwich. So you got three charcoal bbqs stacked on top of each other. No matter how fancy you make the cooking grate the burgers in the middle are gonna fry even if the guys on the end are raw ...

    • by Memroid ( 898199 )

      No matter how fancy you make the cooking grate the burgers in the middle are gonna fry even if the guys on the end are raw ...

      So if I understand you correctly, we need to devise a ram-stack rotisserie. Wendy's may be able to provide guidance, at least regarding a double stack implementation.

  • Old Atari heads know that you can stack RAM on top of the existing RAM packages and solder them in the 520 and 1040 ST machines.

    This is basically doing the same thing, but inside the package.


    • I was going to say, stacking Nand has been around for years so I'm not understanding how this is any different for DRAM.

      • Well for one NAND does not require constant refreshes to retain the data, and thus uses almost no power when not in use. DRAM on the other hand needs to be refreshed constantly, creating significantly more heat issues than nand.
    • But doing it inside the package, directly stacking chip on chip has significant advantages over stacking packages. Lower height, better heat dissipation, shorter interconnects, etc. And if the chips are designed such that they share the drivers, buffers, & latches, etc for the external connection, that can save quite a bit of power. There are many things you can do in package, that are impractical or impossible off package.

  • This will merely increase the density of individual memory modules. However, with processors using multiple memory channels (for performance reasons) you will still require a separate memory unit per memory channel. For Intel Core i5/i7 processors this would be two units. For Xeons it would be sets of three.
  • i had an alpha driven compaq xp1000, it had ram with 2 chips stacked on it...

    also, heat can be led out from the middle of the sandwitch by thin metal plates, glued to the chips with some epoxy...

    • That's stacked packages, this is stacked chips in the package. There are numerous advantages to doing it in package (see my post [slashdot.org] above.)

  • Neural Net CPU [wikia.com]

    I think the terminators want their technology back. Is it time for SkyNet yet?
  • Doesn't higher memory density result in a greater chance for cosmic radiation to flip bits?

    With greater power savings and more memory per module, adding ECC to the mix shouldn't be too painful.

    • Not in this instance. This actually helps avoid that. Smaller geometry increases the changes of a bit flip due to cosmic rays (capacitance discharge due to cosmic ray). This allows 2x (or more) the memory in a given package (may be slightly thicker), without going to a smaller geometry.

      On chip ECC isn't called for (at least not yet), and in fact chipset based ECC has several advantages, including having a single ECC controller for all memory, and being able to detect errors in the memory bus, not just error

  • Right now, I can only get 2x 2gb sticks inside most laptops.

    Given the inherent doubling of chip density this offers, when can I expect to be able to purchase 4gb SODIMM packages?

    • by hawk ( 1151 )


      I bought them from owc for my MacBook about a year ago.

      They now sell 8g modules, at least for the new iMacs (bit they're unGodly expensive, more than $1k each when imchecked a few weeks ago)


    • What? I was speccing a work laptop almost 3 years ago, and I asked if I could get a total of 8 GB. I was told 4 GB SODIMMS are bloody expensive, but they were available anyway.
  • Can they also start stacking cpu's ? 12, 24, 48 cores ? It would have to have cooling pipes running though them, or thin separator plates connected to a cooling system and of soon you would need 220v outlets in your bedroom to power your 48 core system with cooling. Seems per-core cpu seeds have not gained much in the last few years, they are a faster per mhz though from better optimization but seem to mainly faster due to more cores per cpu.
    • Doesn't everyone have 220V outlets in their bedroom? I certainly have....

      • What kind of current do you get out of that 220? In the states the 110 generally gets 15-20 amps per circuit.
        • There are countries outside "The States"... That was pretty much the point of my post.. My country is listed as having 230V [wikipedia.org]. I think you can pull up to 16A.

      • I have 240V...

Real Programmers don't write in PL/I. PL/I is for programmers who can't decide whether to write in COBOL or FORTRAN.