AMD slaps 32GB of RAM on its FirePro W9100 graphics card

While things are still quiet on the consumer graphics card front, a small battle is being waged in the professional graphics arena. Nvidia recently refreshed the M6000, its flagship workstation graphics card, with 24GB of RAM. This put it ahead of AMD’s flagship, the FirePro W9100. Not to be outdone, AMD has announced the release of a 32GB version of the W9100.

Graphics memory is a big deal for workstation graphics cards. For datasets of the right size, this change could be a boon. That said, the rest of the specs for FirePro W9100 remain unchanged. Its fully-enabled Hawaii GPU is the same chip that the 16GB W9100 shipped with in 2014. The card retains the same 44 compute units, 930MHz core clock, wealth of display outputs, and the same, well, everything as the 2014 model. The extra 16GB of graphics memory will cost you, though. AMD quotes an MSRP of $3,999 for the 32GB W9100, a thousand-dollar premium over the current retail price of the 16GB version.

Comments closed
    • ronch
    • 3 years ago

    16GB for a grand. Move aside, App!e.

    • fellix
    • 3 years ago

    One area that would benefit from more on-board memory is accelerated off-line rendering (video compositing, effects and 3D models at once) with large and complex assets, all cached close to the GPU, eliminating the long swap chains with the system memory.

    • Krogoth
    • 3 years ago

    Will it run Crysis at 4K?

    • anotherengineer
    • 3 years ago

    Dammit, win8.1 x86 doesn’t see it with the 4GB of system ram 😉

    Who makes a modern OS 32-bit……………….oh, wait………………….

      • Laykun
      • 3 years ago

      You’d be surprised the amount of support claims I see where users have something like 16GB of ram, a 4GB graphics card and a 32bit install of Windows.

      • DoomGuy64
      • 3 years ago

      Aren’t there extensions that can address that to some degree? Microsoft used to have 32-bit server OSes that could use larger than 4gb, so I don’t see why win8 couldn’t.

        • fade2blac
        • 3 years ago

        I think you mean [url=https://msdn.microsoft.com/en-us/library/windows/desktop/aa366796%28v=vs.85%29.aspx<]this[/url<]. Unless the page is out of date, Win 7 was the last 32-bit OS to support PAE.

          • the
          • 3 years ago

          I think that page is indeed out of date per [url=http://windows.microsoft.com/en-us/windows-8/what-is-pae-nx-sse2<]this page[/url<].

    • Longsdivision
    • 4 years ago

    I think at this point, they should just stick some ram slots on the video cards.

    /s

    srsly….someone thought I was serious? its friday….

    My bad, forgot to put the “/s” before it spawns more discussions on the physical limitations of video cards.

      • jts888
      • 3 years ago

      The looming wall for GPUs is bandwidth power efficiency, and I/O to memory modules takes substantially more power than I/O to on-PCB chips, which in turn takes more power than talking through a silicon interposer, which is again still less efficient than full 3D stacked memory-on-logic.

      I strongly suspect that the “next generation memory” on AMD’s Navi (2018?) will actually be direct stacking, using the lessons learned from HMB1/2.

      At that point, the DRAM architecture will probably have to be modified to be less space efficient (more sub-arrays with shorter columns and much narrower rows) in order to keep the wattage down, but that becomes more feasible when modules interfaces are wide enough to push a cache line or more per cycle.

      [b<]Edit: [/b<] sarcasm is hard to infer sometimes, but especially when you're not pretty familiar with the person speaking, and I though the details would be worth mentioning regardless of how serious you were being. You can see below that some folks actually agreed with your notion...

        • DoomGuy64
        • 3 years ago

        I don’t think businesses that need it really care about the 2% efficiency they’ll lose by gaining the ability to add ram. The real reason why we don’t have ram slots isn’t efficiency, it’s price gouging and vendor lock-in.

        It’s like how intel changes sockets for every single damn cpu they make. You have a haswell system and want to upgrade? Sorry, you have to buy a new motherboard and ram.

        These tactics are why I’ll be going AMD next time I upgrade, and I guarantee that businesses would do the same if they had an option for ram upgrades in video cards. It’s not about efficiency, it’s about profiteering.

          • DPete27
          • 3 years ago

          Yup, and you’ll have to buy a new AM4 motherboard when you do upgrade. Have fun. Hypocrite.

            • xeridea
            • 3 years ago

            AM3 has been around since 2009, buying a new MB when Zen comes out is hardly something to complain about. Intel has so many sockets I don’t even care to know what they are, you may as well assume you need to buy a MB whenever you buy an Intel CPU.

            • ImSpartacus
            • 3 years ago

            Only because Amd has done effectively nothing in the cpu space for the past five years, lol…

            • DoomGuy64
            • 3 years ago

            LOL. If I’m already buying a new board, I’d rather buy one that can get 2 upgrades instead of one. There’s no logic in your comment, but then again I don’t think you were trying to make a logical point in the first place.

            • DPete27
            • 3 years ago

            You’re right, I was trolling.
            However, AMD isn’t nearly as good about maintaining motherboard compatibility across multiple generation CPU upgrades as they used to be. I think this is mostly due to the fact that most functionality/controllers are on the CPU these days whereas back when everything was on motherboard controllers, the IO from the CPU didn’t [need to] change much.

            • derFunkenstein
            • 3 years ago

            I think this is due to the fact they don’t release upgraded CPUs.

          • xeridea
          • 3 years ago

          I think really the reason is physical space. With cards needing large heatsinks on the front, it makes it difficult to allow swapping RAM. You could put it on the back, but then you would have a really fat card. There are also a lot of different bus widths and such from card to card, and high speed GDDR5 needs cooling… It would be difficult to try to do some sort of standard. So it is far easier for 99% of people to just keep it simple and put it on the board. Amount of RAM on cards is paired pretty well with how much the GPU can feasibly use anyway (for consumer cards). For workstation cards, it would be nice to be able to add RAM, since some things would need lots of RAM, but not a ton of horsepower, or tons of both, though it would cause other issues.

          • jts888
          • 3 years ago

          GDDR5 already spends ~60% of it’s power on I/O over short, direct, and spread out PCB traces.

          DDR(x) DIMMs already need 200-300 pins per 64b+8b data channel and use more power signaling the modules than directly soldered modules.

          Trying to get 256b or wider memory buses with interchangeable memory modules at the same clocks could easily suck down an additional 50%+ more power, make the PCBs more expensive, and the final cards more complicated to cool, all for a feature that the vast majority of customers would never want or use.

          Modern GPU architectures do a better job of transparently supporting transfer of memory assets across the PCIe and system interconnect buses, and most working sets can be intelligently managed to be less than on-board GPU memory availability anyway.

      • Voldenuit
      • 3 years ago

      Remember the old Soundblaster cards with RAM slots so you increase your samples filesizes?

      • the
      • 3 years ago

      This isn’t as sarcastic as a request as you’d think. Take the next iteration of the Xeon Phi has HMC for high bandwidth and six channels of DDR4 to back it up with lots of capacity. An AMD or nVidia GPU with similar configuration would be quite impressive as there wouldn’t be an inherent limit on scene/texture complexity any more. Well technically there would be but going from 8 GB to potentially >64 GB is game changing. Looking at cards like the Fury Nano, it is feasible to put a couple of DIMM or SO-DIMM slots on a standard PCIe card.

    • gmskking
    • 4 years ago

    I need at least a TB of ram on my graphics cards.

      • SomeOtherGeek
      • 3 years ago

      And they need to be bootable!

Pin It on Pinterest

Share This