Nvidia’s Volta GPU to feature on-chip DRAM

Nvidia’s GPU Technology Conference kicked off today with a keynote address from CEO Jen-Hsun Huang. A range of topics were covered, with some of the most interesting revelations providing insight on the firm’s GPU roadmap.

The follow-up to the current Kepler microarchitecture is called Maxwell, and it’s due out in 2014. Maxwell will be Nvidia’s first GPU to employ unified virtual memory, which means the graphics chip will be able to see the contents of system memory, and vice versa. Huang says this capability will make programming the GPU easier, and I suspect it will be particularly welcomed by the Tesla crowd.

Source: Nvidia

Maxwell will be succeeded by Volta, which includes more memory innovations. Instead of relying solely on off-chip memory, the Volta GPU will share its silicon substrate with DRAM. This on-chip memory will be stacked vertically, and it will purportedly offer substantially higher bandwidth than what’s available via an external interface. Huang says the integrated memory will boast a whopping 1TB/s of bandwidth—more than three times what’s offered by the GeForce Titan.

Huang didn’t provide a timeline for Volta’s release, but I wouldn’t expect it before 2016. Nvidia has a history of debuting new GPU architectures every two years, with Fermi arriving in 2010, Kepler coming last year, and Maxwell scheduled for next year.

Comments closed
    • WaltC
    • 7 years ago

    Three years in this business is an ocean of time. I don’t believe *anybody* who tells me what he’ll be selling in three years…;) LoL…;) The reality is almost always either better than projected or worse. Often, so-called future technologies are dependent on manufacturing processes that themselves do not now exist.

      • NeelyCam
      • 7 years ago

      but, but… it has fin fez

      • samurai1999
      • 7 years ago

      Yes, but TSMC has a road map
      – no FinFETs at 20nm (Maxwell) but FinFETs at 14nm
      – and 20nm should be making an appearance in 1 year
      – and 14nm 2 years after that

      Although I’ve seen some talk of the 20nm/14nm transition actually happening faster as the feature sizes are all getting blurred (literally as well as metaphorically!)

    • 0g1
    • 7 years ago

    OMG they’re finally going to do embedded DRAM for PC. I’ve been waiting for this since … project ‘Dolphin’ aka N GameCube haha.

      • samurai1999
      • 7 years ago

      The embedded DRAM in the Gamecube & Wii was done differently
      – it was on chip, not stacked. It was 1T ‘SRAM’ from MoSys

      I think nV is talking about using the Stacked DRAM packages from Micron …

        • 0g1
        • 7 years ago

        Oh yeah, NGC was on die, this is on separate modules on the same substrate. 1T SRAM is just a clever way of making DRAM perform like SRAM.
        Stacked just allows more density as far as I know. Being closer to the GPU die allows faster speeds and having the memory stacked for more modules allows really wide IO pathways.
        I just read up on micro+intels efforts here. [url<]http://www.anandtech.com/show/4819/intel-and-micron-develop-hybrid-memory-cube-stacked-dram-is-coming[/url<] Maybe this Hybrid Memory Cube is the same thing as nVidia is using because its also 1tbit. They reduce the IO connections from GPU to each of the modules by having a logic chip to buffer the connected modules and output them at a higher frequency, probably similar to how DDR works.

    • Chrispy_
    • 7 years ago

    purportedly

    • chuckula
    • 7 years ago

    On chip DRAM! That’s Brilliant!

    Now watch those losers at Intel try to copy Nvidia’s innovative designs. I’m so sick of them ripping off everyone else’s PowerPoint innovations with their stupid silicon imitations.

      • shank15217
      • 7 years ago

      i always love on chip dram illustrations, it looks like a napoleon or a chocolate layer cake. too bad in real life no one would really be able to tell.

    • tipoo
    • 7 years ago

    Sure does look like a jetpack to me:

    [url<]http://i.imgur.com/Hl16dYY.jpg[/url<]

    • albundy
    • 7 years ago

    interesting new volta gpu…so pretty much OEMS are being excluded for memory configurations cus the gpu will house the memory. seems like nv is putting a tight grip on scalability, and it looks like the reference design will be less transparent. whats left for the OEMs to change? the video I/O’s, cooler, and circuit board color?

    on the other hand, it looks like it allows for a shorter board which allows the card to be used in smaller systems.

      • cynan
      • 7 years ago

      I don’t think the on board DRAM is going to be the [i<]only[/i<] memory. [quote<]Instead of relying [b<]solely[/b<] on off-chip memory...[/quote<] To me, this suggest it will serve as some form of large cache?

      • Laykun
      • 7 years ago

      Doubtful. This is more likely going to be like the Xenos GPU in the Xbox 360 where it has onboard ED-RAM AND external DDR ram (256 for the GPU). It’s likely to be in small amounts, most likely a sub-100MB number. Generally this is used for speeding up bandwidth hungry operations like antialiasing or anisotropic filtering.

        • Airmantharp
        • 7 years ago

        I don’t know how anyone got anything else from the article. It’s just a lower latency/higher bandwidth cache, not much different than the massive caches Intel puts on Xeons and Itaniums, except that it will be DRAM-based instead of SRAM-based for increased density.

        • Liron
        • 7 years ago

        Assuming the chip in the picture is roughly drawn to scale, we can check the possible amount. The diagram shows 6 chips in stacks of 4. For 100MB, each chip would have to hold only 4MB. How much area would a 4MB chip use at 16nm?

    • NeelyCam
    • 7 years ago

    Thinking back, there was an NVidia paper in ISSCC 2013 about a 20Gb/s super-low-power short-range link – the authors were mainly from the North Carolina unit NVidia poached from Rambus a couple of years ago, led by John Poulton.

    Overall, that was one of the top 3 papers I saw in the conference. Mighty impressive work, and I have a strong feeling it’s adopted for this Volta GPU-memorystack link.

    • anotherengineer
    • 7 years ago

    I don’t know why, but old Huang reminds me of Shang Tsung from mortal kombat………….

    yes? no?
    [url<]http://sydlexia.com/imagesandstuff/mk2/mk20739.png[/url<]

      • sweatshopking
      • 7 years ago

      cause he’s an asian man?

        • chuckula
        • 7 years ago

        Oh so Jen is a dude… oh.. that’s awkward.

        • MadManOriginal
        • 7 years ago

        cause of the supergay pose and rainbow?

    • Arclight
    • 7 years ago

    HMC b*tches, f*ck yeah.

    Edit
    Wait, what? 2016? Sigh

      • NeelyCam
      • 7 years ago

      Told ya.

      [url<]http://www.techreport.com/news/21807/tuesday-shortbread?post=587661[/url<]

        • samurai1999
        • 7 years ago

        Yes – It’s probably a Micron/Samsung chip stack in the same package

        • Arclight
        • 7 years ago

        [url=http://i.imgur.com/jfNxrJu.jpg<]Yes you did Sir. Also you are a scholar and a gentlemen[/url<]

    • R2P2
    • 7 years ago

    In the picture, the end of the video card makes it look like the guy has a jetpack on his back.

      • NeelyCam
      • 7 years ago

      Only a Boba Fett mask is missing

      • smilingcrow
      • 7 years ago

      So that’s the source of all the hot air he spews!

      • smilingcrow
      • 7 years ago

      As it looks like a massive PCI blanking plate stuck to his back maybe he is symbolically saying that the era of big PC iron is behind us and the future is mobile with Tegra.
      So a post modern, post PC symbolist joke! What next; Lady Gaga wearing an outfit weaved out of 22nm Atom SoCs dueting with Will.I.am dissing Alica Keys. They could call it something like; You is stuck on 28nm bulk silicon bitch.

        • Celess
        • 7 years ago

        More like: we are going to ride this back plate thing for as long as we can; I’ve even made a backplate shaped jetpack because I’m so excited.

      • tipoo
      • 7 years ago

      I registered after years of reading this site just for this:

      [url<]http://i.imgur.com/Hl16dYY.jpg[/url<]

    • NeelyCam
    • 7 years ago

    I don’t think this is the first time I’ve heard about this…. lemmethink… Oh, that’s right: Mars Volta

    [s<]WideIO probably, but the GPU underneath is going to be a [i<]royal pain[/i<] to cool - the stack of RAM on top is going to increase the thermal resistance quite a bit[/s<] EDIT: strikethrough for stupidity

      • Helmore
      • 7 years ago

      The GPU isn’t going to be underneath the RAM.

        • Firestarter
        • 7 years ago

        that’s what I was thinking, they’d layer the RAM on top of the GPU, then mount it flipped over

        • NeelyCam
        • 7 years ago

        Oh, yeah you’re right. My bad. Too busy trying to get an FP for the FIRST time in months!

        Looks like NeelyCam was right about something else, though… remember this:

        [url<]https://techreport.com/news/21807/tuesday-shortbread?post=587661[/url<] Because what Volta has is several HMCs, and the five-year prediction seems to line up pretty well

      • Flying Fox
      • 7 years ago

      From the diagram it looks like the memory stacks are surrounding the GPU area, not on top of the GPU area?

        • MadManOriginal
        • 7 years ago

        Yeah, it does. It makes sense as a step to transition from off-chip memory. Make one die that includes the RAM, but it’s connected by an on-die bus to the GPU memory controller the way off-chip RAM is right now. The next steps will be tighter integration of the on-die RAM and GPU, eventually to the point that there aren’t really memory channels but more of a true ring bus like Intel does instead of a crossbar.

          • Wirko
          • 7 years ago

          The precious GPU die would have to be three or four times as big in order to accomodate the DRAM stacks. For this reason alone I believe that the DRAM chips are off die, however, they do share the same package.

            • MadManOriginal
            • 7 years ago

            From the article:

            [quote<]Instead of relying solely on off-chip memory, the Volta GPU will [u<]share its silicon substrate with DRAM[/u<].[/quote<] As cynan said elsewhere, the word 'solely' may be key. Nonetheless, it is clear that some DRAM - and that means not just small amounts of cache - will be on-die. Also keep in mind this is slated for 2016, that means 'real' 14nm or maybe even 10nm-class transistors, so there will be a lot of die space available even with added graphics cores.

Pin It on Pinterest

Share This