1. This site uses cookies. By continuing to use this site, you are agreeing to our use of cookies. Learn More.

AMD Working on Vertical (3D) Stacking of DRAM onto processors

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Mar 18, 2019.

  1. Hilbert Hagedoorn

    Hilbert Hagedoorn Don Vito Corleone Staff Member

    Messages:
    34,851
    Likes Received:
    4,050
    GPU:
    AMD | NVIDIA
    chispy and fantaskarsef like this.
  2. nevcairiel

    nevcairiel Master Guru

    Messages:
    462
    Likes Received:
    126
    GPU:
    MSI 1080 Gaming X
    I wonder if at some point we'll actually lose the freedom to pick the amount/speed of DRAM we want, or if this will remain limited to select embedded use-cases.
     
  3. Ridiric

    Ridiric Member Guru

    Messages:
    167
    Likes Received:
    92
    GPU:
    GTX 1080Ti@2068mhz
    Eventually we will get to the point where the only way to increase ram speed will be to have it on package with the CPU, you can already see this a little bit with how overclocking focused boards often only have 2 ram slots placed as close as they can get them to the cpu to increase ram overclocking ability.

    So I wouldn't mind it at all if we loose the ability for upgrading ram or changing ram speed without replacing the whole CPU package, as long as the ram speeds and efficiencies on those packages are better than what would be possible with the current method of having separate ram modules.
     
  4. Backstabak

    Backstabak Master Guru

    Messages:
    309
    Likes Received:
    85
    GPU:
    MSI GTX 970
    I would very much mind loosing the freedom to choose what amount and what kind of RAM I will have in my system. There are huge differences in both price and performance between modules and I doubt that either intel or AMD will have so much combinations to satisfy everyone's needs.

    So even though I agree that to get better performance something like this must happen, I hope that they would introduce possibility to install RAM modules as a service or DIY slot on the CPU, but I guess that is far into the future, as they first need to make it a reality and show that it is actually significantly better than what we have now.
     

  5. Kaarme

    Kaarme Maha Guru

    Messages:
    1,433
    Likes Received:
    329
    GPU:
    Sapphire 390
    Couldn't we get a bunch of truly fast RAM inside the CPU package, but then the user could add more in the typical DIMM form? It would be slower, but if the system ran out of what's shipped with the CPU, the system could use it. It would still be faster than an SSD (bandwidth and latency wise). Who knows how much the CPU package would initially ship with anyway.
     
    schmidtbag likes this.
  6. K.S.

    K.S. Maha Guru

    Messages:
    1,262
    Likes Received:
    199
    GPU:
    1080 Ti SEA HAWK
    Yeah "right to repair" n all that.. definitely skirting further n further away and closer to none at all.

    No way the CPU OEMs are going to match anything anywhere near the current dedicated modular RAM market - it's just not profitable or realistic

    Any smart memory manufacturer is lobbying to protect their piece of the market. Not like they are against performance - loss of profit. Likely a halfway solution between them & the cpu makers will arise.
     
    Last edited: Mar 18, 2019
  7. Fox2232

    Fox2232 Ancient Guru

    Messages:
    8,612
    Likes Received:
    1,565
    GPU:
    -NDA +AW@240Hz
    Honestly, if AMD starts selling 8C/16T APUs w/ solid iGPU + 8/16GB HBM2 or other embedded memory... people will not complain.
    45W CPU part + 85W iGPU part + 10~20W memory + VRMs 25W of heat. Size of that system... Tiny, but nasty gaming box. (Just 220~250W PSU.)

    Due to actual heat producing components being nicely in one place, cooling solution can be quite optimized.

    And what that can do for mobile...
    This is no different than HBM for GPU and that's not much different than GDDR5. Memory is integrated one way or another.
    Those "Memory Assembling Companies" are doing exactly that. Memory chips themselves are made by very few.

    Is memory stick in danger with this? No. Server business requires components which can be easily replaced and upgraded.
    Is that problem for notebooks/tablets? No, tablets can't be upgraded in 99% of cases and those which have actual slot of any sort (M.2) are quite new.
    In notebook market, most of them are not upgraded through their life cycle in terms of system memory. It is nice option, but if it is integrated, it trades ability to exchange memory for smaller thickness and possibly power efficiency. And as long as that reduces cost of device, one can pay for extra warranty.
     
  8. HWgeek

    HWgeek Master Guru

    Messages:
    345
    Likes Received:
    243
    GPU:
    Gigabyte 6200 Turbo Fotce @500/600 8x1p
    I already though about it since first Ryzen came, if we are getting to the point where 8C/16T can be bought for reasonable price- what going to come later to make you upgrade?
    Now we are looking for 16C/32T Desktop CPU's- what could be a reason for upgrade in 2021+?
    I think in Future desktop CPU's will come with Ram on CPU, Starting at 4GB low end~ upto 64GB high end - same story like on Ram/Storage on smartphones.
    This way AMD/Intel gonna keep making good $$$, for example:
    8C/16T comes with 16GB Ram, for you 8C/16T is more then enough but you want more then 16GB- then you have to spend more and buy 12C/24T with 24GB or 16C/32T with 32GB and etc'.

    On HEDT/Servers - they will keep using Ram modules for long time.

    P.S- Memory manufactures will be happy with this method- think about it: for years you could upgrade your MB/CPU and keep using old DDR3 memory or DDR4 memory= no new sales for Memory manufactures , but if you will be force to replace slower CPU with built in Ram with newer CPU with built in Ram- then Memory manufactures just sold you another memory modules!.
     
    Last edited: Mar 18, 2019
    Fox2232 likes this.
  9. dexterav

    dexterav Active Member

    Messages:
    51
    Likes Received:
    9
    GPU:
    Evga GTX1080Ti FE
    well AMD have well promising HBM, i think is only promising and at the its not that big deal, expensive, many compromises, DDR5X/DDR6 looks better with lowest price...
    embeded ram is ok for oem, bude bad for rest, i agree with HWgeek
    i see in future total bad combination...
    week cpu with so many ram, strong cpu with half of this ram etc
    say simple, no good choices :/
    or in best case good choices but with abyssmal price
     
  10. HardwareCaps

    HardwareCaps Master Guru

    Messages:
    452
    Likes Received:
    154
    GPU:
    x
    I love how 4 years ago it was like:"We don't need more than 4 cores, cpu's are perfectly fine for now"
    after there's some competition, everyone is like: "We have to push to 10 cores & use 3D stacking techniques"

    [​IMG]
     

  11. theoneofgod

    theoneofgod Ancient Guru

    Messages:
    4,020
    Likes Received:
    26
    GPU:
    RX 580 8GB
    [​IMG]
     
  12. Kaarme

    Kaarme Maha Guru

    Messages:
    1,433
    Likes Received:
    329
    GPU:
    Sapphire 390
    [​IMG]
    *Not an actual quote
     
    schmidtbag likes this.
  13. Evildead666

    Evildead666 Master Guru

    Messages:
    904
    Likes Received:
    112
    GPU:
    Vega64EKWB 240PP12V
    I can see it being just a large L4 cache.
    4GB would be enough to enable low-end PC's without any extra RAM at all.
    It would obviously have to be much much faster, and lower latency than DRAM, otherwise there is literally no point.

    4 or 8GB of HBM on package could be used as L4, and shared with the integrated GPU.
    The RAM would then be the next Memory storage tier, and would be able to be a lot slower.
    The Large L4 would mitigate memory problems.

    I could quite possibly happen, especially if one of the chiplets could be replaced by HBM.
    I see that happening probably in the Threadrippers first, for HPC work, and then the Mobile chips, for integrated graphics.
    Eventually, it would hit the enthusiast line on AM4 or AM5.
     
  14. HWgeek

    HWgeek Master Guru

    Messages:
    345
    Likes Received:
    243
    GPU:
    Gigabyte 6200 Turbo Fotce @500/600 8x1p
    Did any1 tested the influence of GDDR5 on ZEN in Subor Z+ console?
     
  15. JamesSneed

    JamesSneed Master Guru

    Messages:
    452
    Likes Received:
    126
    GPU:
    GTX 1070
    Seems like a rather logical evolution to making an APU with AMD's chiplets. The thing that hurts the APU's the most today is memory bandwidth so this solves that issue getting it on die.

    Looking at how much DRAM samsung is getting into a phone, 12GB, it may be possible we see system ram for APU's on die. LIkely this would start on console APU's, and trickle out.
     
    Last edited: Mar 18, 2019

  16. schmidtbag

    schmidtbag Ancient Guru

    Messages:
    3,778
    Likes Received:
    850
    GPU:
    HIS R9 290
    It was only a matter of time until this started to happen. Many ARM platforms have been doing this for years.

    I was thinking the same thing. As far as I'm concerned, we can think of this like an L4 cache. I'm glad they're pushing for HBM and more memory channels too, since that's exactly what AMD needs right now.
    I think if AMD supplies 0.5GB per CPU core and maybe 0.25GB per GPU core, that ought to be plenty sufficient to start with, as long as you get to supplement with discrete DDR4 memory. The way I see it, DDR4 is relatively slow but it's fast enough where it should be able to load in new data without being a bottleneck, while the integrated RAM is more for "current data".

    I guess one major concern I have is how much this may hinder overclocking. But, it looks like we're heading to a point of x86 PCs becoming SoCs. Although it's a necessary evolution, it's a little sad to me, in that this is the beginning of the end for hardware enthusiasts.


    I really hate that quote, not just because Gates never said it, but also because it's always taken out of context. 640K, at the time, was enough for anybody.
     
    HonoredShadow, HandR and Kaarme like this.
  17. OnnA

    OnnA Ancient Guru

    Messages:
    7,673
    Likes Received:
    1,213
    GPU:
    Vega 64 XTX LiQuiD
    Finally :D
    I know of this for some time.
    But we need 3D+2.5D Stacked GPUs chiplets w Infinity Fabric + HBM
     
  18. MBTP

    MBTP Active Member

    Messages:
    83
    Likes Received:
    3
    GPU:
    HD4600
    Sincerely, i don't see why anyone would be concerned by upgradability, i think this is the right, and probably the only way foward.
    1. This can co-exists with DDR ram stick modules.
    2. Smartphones are like this already, seems most people don't care.
    3. The major market for this being small portable notebooks, ultrabooks, and laptops, are generally not upgradable anyway, and when they are most people don't upgrade ever.
    4. The overall cost of the chiplets will be in no time very cheap.
    5. This style will probably be sold in packages so in the end will be cheaper to the end consumer, with highly optimized cooling systems.
    6. This could work as a L4 cache or not as someone said already, and let's remeber that Non volatile "rams" are coming to the consumer market, this could be extremely compact with it.
    7. This could be the next step to modular systems, as antagonic this statement can sound, with AM4 as and example, more optimized cooling systems, less components on the motherboard itself, reduced motherboard footprint, problems like having to upgrade a motherboard just to support the next memory technology will not need to exist.
    Ex: APU+HBM2.0 "AM5" socket, will probably be able to be upgraded to APU+HBM3.0/4.0, since the limitation will not be on the MB traces, but you be limited maybe on the I/O speed of last gen. This will be highly dependent on MB suppliers and CPU sockets...
     
    Last edited: Mar 18, 2019
  19. Venix

    Venix Master Guru

    Messages:
    707
    Likes Received:
    182
    GPU:
    Palit 1060 6gb
    With apus this is not a surprise apus are very much bottlenecked by ddr4 and ddr5 will not be enough. Tiered storage is how computers work since forever l1-l2-l3-ddr now it will be l1-l2-l3-stuck-ddr
     
  20. Richard Nutman

    Richard Nutman Member

    Messages:
    32
    Likes Received:
    11
    GPU:
    AMD RX 580 8GB
    Broadwell DT already did this with 128MB of EDRAM for the integrated gpu. When using discrete graphics it acted like a L4 cache. It gave some serious performance boosts in games.

    It's a shame this tech didn't take off more.
     
    schmidtbag likes this.

Share This Page