Nvidia intros Kepler-based Quadro K5000

Today’s all about pro graphics, it seems. Not long after AMD’s hail of FirePro releases this morning, Nvidia announced a high-end Quadro graphics processor based on its new Kepler architecture.

The Quadro K5000 features 1536 ALUs and a 256-bit memory interface, just like the GeForce GTX 680, but it has twice as much memory: 4GB of GDDR5. The card has peak arithmetic rates of 2150 GLFOPs for single-precision computations and 90 GFLOPs for double-precision work. Nvidia outfits this bad boy with a dual-slot cooler and a mix of display outputs, including DVI-I, DVI-D, and dual DisplayPort. (The card can drive up to four displays at once.) Peak power consumption is rated at 122W.

Since this is a Kepler-based product, it delivers some of that architecture’s improvements to the professional graphics world. In particular, Nvidia touts bindless textures, which "give users the ability to reference over 1 million textures directly in memory while reducing CPU overhead," and TXAA antialiasing, which is meant to improve image quality and reduce crawling and flickering from one frame to the next. PCI Express 3.0 support is on the menu, as well.

Ths Quadro K5000 has a suggested retail price of $2,249, and Nvidia expects to make it available in October. Workstations that couple the Quadro K5000 with Nvidia’s Tesla K20 processor, in what Nvidia calls its second-generation Maximus platform, will be available in December from vendors like HP, Dell, Lenovo, Fujitsu, BOXX, and Supermicro. Nvidia’s Maximus technology allows a Quadro GPU to shoulder graphics work while a companion Tesla processor handles general-purpose computing duties. The first-generation Maximus platform debuted last fall with Fermi-based offerings.

Comments closed
    • Fluffmeister
    • 7 years ago

    nV’s current Quadro range fair excellent against AMD’s new cards:

    [url<]http://hothardware.com/Reviews/AMDs-New-FirePro-W8000-W9000-Challenge-Nvidias-Quadro/?page=1[/url<] You can say it's the drivers, but in this market poor drivers aren't acceptable 😛

    • ish718
    • 7 years ago

    $2,249 for the Nvidia Quadro K5000 and $3,999 for the AMD FirePro W9000.

    The FirePro W9000 does have 1 TFLOPS of DP and is just as good for 3D graphics work as the Quadro.

    • jamsbong
    • 7 years ago

    With NV throwing DP out of the window and ATI not having much market share. It seems that the only winner is Intel. Their CPUs have been consistent with DP performance throughout the different generations. Existing computing intensive softwares like CFD is still predominantly a CPU based code. Rewriting the code in OpenCL seems like a pointless exercise now.

    Work that requires SP will only need a cheap gaming card.

    That said, I’ve heard some rumors that GT100 was still in the works and that is suppose to be a serious monster. I guess that remains just a rumor to please NV fans.

      • HisDivineOrder
      • 7 years ago

      Papa Kepler will take care of that situation. Meanwhile, we make do with Ma Kepler and Baby Kepler for the time being until Pa is ready.

      It doesn’t hurt that Pa Kepler can take his time getting his shoes on ’cause AMD hasn’t done much except stare at the wall and keep prices high on parts that barely moved the performance level up a notch from last generation.

        • moose17145
        • 7 years ago

        Not gonna downvote you… but I will disagree with you because (from the way I read it), you make it sound like AMD isn’t doing anything to improve on the situation, when this simply (IMHO) is far from the case.

        For several generations now NV has been the king of DP, but they chose to throw that out the window with kepler and are now behind AMD on that front. So I would argue that NV done even less, and has in fact gone backwards in DP computing. This is obvious in the fact that their previous generation is better at DP than current gen kepler is. And yet people still be saying that NV has already won even though their current generation has lost because NV might be coming out with something in the future (meaning not yet available to anyone) that might change that. If it’s a future product it better be able to beat everything on the market… because it’s a FUTURE product.

        I guess with the way people are making it to sound like, AMD hasn’t come up with anything compelling this iteration, when their products have been more well rounded, performed as good/better, have had reasonable/competitive power consumption and thermal envelopes, and have come out about 6 months ahead of anything NV has come up with (obvious current exception being that the FirePro and Quadros seem to be getting released at the exact same time). And even with AMD’s current cards in this particular instance being more compelling than what NV has, people are still saying “oh well just you wait till BIG kepler comes out… It will all be over then…” It’s the same stuff I was hearing when the 7900 series first came out. “Oh well just wait till Kepler comes out… AMD will be hurting bad then…” Then NV brought out kepler, and it was barely better than what AMD had. Then everyone just started saying “oh uhh… well.. when Big Kepler comes out. THEN it will be all over… ” Idk I guess I just don’t see what so hard to admitting that for the time being… NV has lost this round…

        NV wins half the rounds and AMD wins the other half… For roughly half of the time I have been building computers (I am only 26, been building since about 14/15) I have been recommending NV cards… but likewise the other half the time I have been recommending ATI / AMD cards. I simply recommend whatever appears to be the best bang for the buck at the time given what that person says they are looking for at that time (same thing with Intel vs AMD CPU’s). If Big kepler finally fixes half of the glaring issues with current kepler in six months.. than that is totally awesome… six months from now.. But for the time being, I am sorry, for what these cards are meant for I just see AMD as the more compelling choice. Even for Just gaming I see the 7xxx series as being a more compelling option for most of the lineup. If you are looking for something purely for gaming in the 400 dollar price range then the 670 is a very compelling choice, and I have no issues recommending one at that price for the right person. But for anything else I will likely recommend a Radeon currently. They have a compelling and complete lineup outside the 400 dollar price range, and even for the FirePro’s are apparently offering multiple cards at various price points.

        Maybe this entire cycle the whole “oh well just you wait till NV releases XYZ six months from now… ” is getting rather old and jaded to me… but that’s just my two cents on the matter…

      • Silus
      • 7 years ago

      NVIDIA is not throwing anything away! Can you at least check what you’re commenting on ? This is a Quadro card, not Tesla. Tesla cards are the compute monsters. Quadro is a brand for professional graphics cards!

    • Chrispy_
    • 7 years ago

    Kepler = sacrificing compute power to make gaming graphics faster.
    Quadro = professional compute and double-precision cards.

    Something is not right here, I’ll take the AMD solution that is five times faster, thanks.

      • Jason181
      • 7 years ago

      Quadro is actually not a specific architecture, but a branding. It’s amusing that in the last generation, AMD and nVidia did the exact opposite.

      I’ll be curious to see how much of an advantage the AMD cards have in actual practice, but that won’t happen for awhile given that the Kepler isn’t due out until October.

      • Ryu Connor
      • 7 years ago

      Tesla is the double-precision card, not Quadro.

      Tesla:
      [url<]http://www.nvidia.com/object/personal-supercomputing.html[/url<] Quadro: [url<]http://www.nvidia.com/object/quadro-K5000.html[/url<] The K5000 has some sort of hybrid thing going on (Maximus), but with stats weaker than the last Tesla C2075 this isn't the replacement for last gen Tesla cards. Or someone made a mistake in the stats. [url<]http://www.nvidia.com/object/workstation-solutions-tesla.html[/url<] According to that link the Tesla Maximus replacement based on Kepler won't arrive until Q4 2012.

        • Jason181
        • 7 years ago

        500 GFlops dp on current generation; that’s more like it! Hehe.

        • Silus
        • 7 years ago

        The AMD fanboys don’t really care about facts! They just see some spec that’s lower than AMD’s and AMD is immediately better 🙂

          • jamsbong
          • 7 years ago

          I don’t think there is a ATI vs NV fanfare here.
          It is a fact that:
          the Tahiti architecture has much better DP performance.
          Kepler has << DP than Fermi.

          Having better DP is the only reason to have a GPGPU work anyway. Otherwise, just get a game card to crunch SP stuff.

    • Jason181
    • 7 years ago

    ~$1 per gigaflop, nice 🙂 Peak power consumption is very nice too.

    Edit: The impressive part is that they’re doing this with only 700 Mhz core.

    • chuckula
    • 7 years ago

    [quote<]90 GFLOPs for double-precision work[/quote<] Meaning that it is comparable (not favorably) to the ~100 -> 120 Gflops of double precision throughput you can get from a desktop-grade quad-core Sandy or Ivy Bridge CPU. Obviously Big Keppler will be in a different league, but these guys are not all that spectacular at DP number crunching.

      • CampinCarl
      • 7 years ago

      I think most people understand that most applications dont NEED DPFP data. Most of them are just fine in terms of accuracy, using SPFP. It’s a matter of understanding your actual problem, and not just saying “Let’s be SUPER accurate, even though we don’t need to be!”.

        • jensend
        • 7 years ago

        Needing double precision data is very different from needing double precision computations.

        If all you’re doing is taking the data and displaying it, sure, most applications don’t need DP. But anybody who’s had any training in numerical analysis and/or has had to work with floating point solvers, optimization algorithms, etc knows that SP is very frequently not enough for real scientific work. Error propagation may be too large even for very numerically stable algorithms solving well-conditioned problems, to say nothing of marginally stable algorithms or systems with relatively high condition number.

        There are good reasons why DP has been the standard in science for decades. [i<]Sometimes[/i<] you can get good enough convergence with SP if you're very careful about it, or you may be able to only do a few of the intermediate operations in DP and do the bulk of your work in SP if you are clever enough about it. But for a vast range of problems, DP allows you to use simple algorithms and have enough breathing room with your precision to be confident about the results. It's quite telling that claims that SP is enough usually come from companies which provide no DP support or lackluster DP performance and their apologists, not from impartial scientists, mathematicians, or numerical analysts.

          • Meadows
          • 7 years ago

          This card is for graphics design/modeling, not science. It doesn’t need double precision.

            • derFunkenstein
            • 7 years ago

            True to a point. Double-precision is reserved for Tesla cards, though it’d be nice if they’d unlock it for their expensive professional cards so they could pull double-duty. You’re already paying a huge premium

            • Silus
            • 7 years ago

            And destroy one of your markets by allowing products from another market to excel there ?

            Don’t ever open a business of your own…you will fail horribly!

            • BobbinThreadbare
            • 7 years ago

            1) the market wouldn’t be destroyed, 2) either way Nvidia sells a product. Do they care a lot whether it’s a tesla or a Quadro?

            It’s also possible they could create a new market, or build customer satisfaction which would lead to more sales later.

            • Silus
            • 7 years ago

            Precisely!

          • UberGerbil
          • 7 years ago

          And sometimes even DP is not enough, which is why the old x87 used 80bit internally (though clearly Intel no longer thinks so, since they’ve stuck to DP for everything since)

            • willmore
            • 7 years ago

            At least the fused operations should remove a rounding step in some operations.

        • Jason181
        • 7 years ago

        More than you’d ever want to know about floating point:
        [url<]http://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.html[/url<] I read a document awhile ago that discussed getting different answers from what appears to be equivalent code. Wish I could find that one.

      • codedivine
      • 7 years ago

      AMD introduced their FirePro with 1 Teraflop DP. Clearly, if you want DP, thats the way to go.

        • brucethemoose
        • 7 years ago

        4 TFlops SP too, according to them.

        • beck2448
        • 7 years ago

        Obviously you don’t work in the pro graphics world. Nvidia has had 80% + of that market for years because of their SOFTWARE.

          • Jason181
          • 7 years ago

          Pro graphics uses double-precision?

            • Silus
            • 7 years ago

            No it doesn’t.

          • UberGerbil
          • 7 years ago

          Obviously, [url=http://codedivine.wordpress.com/2012/07/11/c-amp-first-impressions/<]he does[/url<].

      • flip-mode
      • 7 years ago

      It seems legit to wonder if Big Kepler is real. Any credible rumors on that front? And if so, is it going to launch as a desktop product first and take another 6-8 months before it launches as a Quadro? If so, the next gen FirePro won’t be far behind.

        • Farting Bob
        • 7 years ago

        I think the biggest argument in favour of a Big Kepler (or ‘Phat K’ to give it’s proper 90’s rap name) is that the current high end chips we’ve seen from NV are only about the same size physically as the previous generations mid range. While it’s clear that 28nm yields are the biggest hurdle to making a truly massive chip right now, by the end of the year it will be much more doable to make a larger chip. Since The current Kepler is a ‘complete’ chip design (not a chip with parts disabled due to binning), we’ll likely see a slightly new design to take full advantage of the extra leg room.

        However, given the gaming performance of the 680 on a fairly small die compared to what we have seen before would suggest the need to target a bigger chip at gamers is low. They already have a $1000 card, there is no point in crowding the $600+ consumer GPU market. But in the HPC crowd, they’ll love a more beefy card as there is always demand for more performance.

          • UberGerbil
          • 7 years ago

          “Phat K” is good, but if we’re going 90s shouldn’t it be “Biggie Kep” (and “Lil’ K”?)

        • Silus
        • 7 years ago

        You’re still wondering if it’s real ?? All the articles about it after NVIDIA’s presentation aren’t enough ?
        Also, it’s pretty clear that it is focused on compute and thus a Tesla product. A graphics product will only exist if AMD does something to challenge them in the high-end.

          • flip-mode
          • 7 years ago

          Yeah, that’s why I asked. Sounds like you’re saying it’s 100% real.

Pin It on Pinterest

Share This