Nvidia Quadro GV100 stands ready to power a ray-traced future

As part of his GTC 2018 keynote, Nvidia CEO Jensen Huang revealed the company's latest professional graphics card: the Quadro GV100. This professional-visualization product offers 32 GB of HBM2 memory, up from 16 GB on past Volta V100 implementations, alongside the same 5,120 shader processors (80 of 84 Volta shader multiprocessors on the V100 GPU) that have appeared on past Volta products. Nvidia believes this card will be the optimal foundation for exploring its RTX real-time ray-tracing platform.

Nvidia claims the GV100 will deliver 14.8 TFLOPS of single-precision compute power and 7.4 TFLOPS of double-precision number-crunching power, along with 118.5 TFLOPS of tensor processing. Although the company didn't say as much, those figures suggest a 1448 MHz boost clock, down from Tesla V100's 1530 MHz.

Two Quadro GV100 cards can be joined across the coherent NVLink 2 interconnect to create a beastly number-cruncher with 10,240 shader processors, 64 GB of HBM2 memory, and 236 TFLOPS of tensor-core power. The Quadro GV100 will allow developers to access Nvidia's RTX technology through Microsoft's DirectX Raytracing API, the Nvidia Optix API, or through Vulkan in an upcoming version of that API.

The Quadro GV100 is available direct from Nvidia today at a price of $8,999.

Comments closed
    • ronch
    • 2 years ago

    Next – AMD scrambling to come up with something similar as soon as they found out about this announcement.

      • chuckula
      • 2 years ago

      What are you talking about? This is clearly Nvidia’s panicked response to Navi.

      AMD has already won.

      • NTMBK
      • 2 years ago

      You mean a workstation card with HBM2? They shipped one of those months ago.

        • chuckula
        • 2 years ago

        It’s great they shipped one card.
        Who got it? I may want to “borrow” it.

    • the
    • 2 years ago

    One of the things being over looked is that this card has its full compliment of ROPs enabled. This should be a record holder in terms of fill rate which comes in very handy at high resolutions.

    If scaling works as advertised with nvLink this might be a 5K or 8K capable for gaming, though at the stupid crazy price of ~$20,000.

      • Srsly_Bro
      • 2 years ago

      If the Titan V is a somewhat capable 4K card on recent titles, how can a few more ROPs enable it to push 4 times the pixels? Am I missing something?

      GTA V and other titles are around 50-60 fps @ 4k, for example. More ROPs allow it to push 4 times the pixels at a similar frame rate?

      • psuedonymous
      • 2 years ago

      I don’t think any gaming workload has been ROP-limited for many years now, unless you have a very basic title and are trying to push 120FPS at 8k or 16k through DXR.
      e.g. A 1080Ti has 88 ROPs and a fillrate of ~130 Gigapixels per second. A UHD image is ~8.3 Megapixels, for a raw fillrate of over 15 thousand UHD images per second.

      It’s a similar situation with vRAM and render resolution: the amount of vRAM used by render buffers is tiny compared to that used by texture data. A single 32-bpp (8/channel + alpha) UHD image is a hair over 33MB. Even if you use Deferred Shading with 4 buffers, and keep the last frame’s buffers for temporal effects, a collection of UHD framebuffers is just 256MB.

        • the
        • 2 years ago

        15 thousand UHD images sounds like a lot but you have to break that up across multiple buffers as you note which are constantly being read and/or written to. Other factors like HDR using 16 bit or even 32 bit values per color channel are eating away at that fill rate. Fill rate certainly doesn’t have the same weight on performance as it did two decades ago but it is a factor. For example, the Kelper -> Maxwell transition improved ROP efficiency which lead to some serious performance gains at higher resolutions.

        To your point of containing various buffers in a small amount of memory has been an idea that has always been floated about. Bitboys back in the day touted their on-die eDRAM as a solution to that very problem while using external memory for textures. More recently Intel’s Crystalwell parts have been doing something similar with the GPU having a say in how those buffers gets cached. AMD and nVidia has taken an alternative approach by incorporating large L2 caches within each ROP/memory controller cluster.

          • Laykun
          • 2 years ago

          This, plus overdraw tends to chip away at your fillrate. While we make large efforts to avoid overdraw it’s still an inevitable fact of any modern dynamic game.

            • the
            • 2 years ago

            Overdraw is correct but it isn’t the issue it used to be. Both AMD and nVidia have implemented some tile based rendering techniques and z-culling to reduce the amount of overdraw being performed.

    • Neutronbeam
    • 2 years ago

    Well, yeah, but can it play Minecraft?

      • chuckula
      • 2 years ago

      Well Ngreedia gives you the Tensor cores. It’s your job to use them to produce the AI that plays Minecraft!

        • Neutronbeam
        • 2 years ago

        And is that AI named…GLaDOS?!

          • K-L-Waster
          • 2 years ago

          It’ll be a huge success!

      • NovusBogus
      • 2 years ago

      Not gonna lie, on seeing the headline my mind was considering an entirely different type of mining…

      • freebird
      • 2 years ago

      Don’t you mean #Don’tMakeMeCry-sis?!?

        • Srsly_Bro
        • 2 years ago

        It’s #dontmakemecrysis because some people on here likely get beat up by their little sister.

      • ronch
      • 2 years ago

      Quake 2 is still better.

    • Voldenuit
    • 2 years ago

    so, no GTX 2080/Turing/Ampere?

      • DancinJack
      • 2 years ago

      No.

      • Leader952
      • 2 years ago

      Why?

      Nvidia does not announce consumer GPUs at the GTC conferences.

        • Voldenuit
        • 2 years ago

        Yeah, but if they had a new architecture, which would go into their Quadro, Tesla and Titan cards, I’d expect to hear something. Which is why I included Ampere/Turing.

      • JustAnEngineer
      • 2 years ago

      When they can sell their entire production run in $9K cards, why would they bother with a mere $2000 gaming card?

        • DoomGuy64
        • 2 years ago

        If sheer greed is what finally stops Nvidia from selling $2000 gaming cards, I’m all for it. Gaming cards have no business being over $600 even for the most ridiculously high end.

        The best option would be to enable all gaming features on the Quadro, and make the fanboys pay $9000 for the privilege. They certainly deserve it after letting Nvidia sell gaming cards first for $1000, then $2000. What’s a few more grand to those people? That way they get their halo epeen product, while the rest of us will finally recognize those people for what they are. Crazy delusional fanboys with more money than brains, and shouldn’t be allowed anywhere near the rest of us if we want to stay sane or have affordable products.

          • Kretschmer
          • 2 years ago

          Jealousy is an ugly thing.

          At least until I use my 1080Ti to add supersampling. 🙂

            • DoomGuy64
            • 2 years ago

            I don’t know what you’re all excited about. Those things are obsolete now, and you have to pay $9000 to get the new cards. It’s the only way Gameworks 2.0 with mandatory ray tracing will get playable framerates.

            • Kretschmer
            • 2 years ago

            You buy what you need to fuel a given experience at a given time. Nothing is forever. The Voodoo3 AGP was awesome back in the day, but technology marches forward.

            That’s not to say that there aren’t bad values and good values, but “value” is totally subjective. I might consider a 1080Ti a good value in 2017 as the minimum part for the gaming experience I want, while someone else would consider it a ridiculous luxury. Alternately, a startup entrepreneur might consider a Titan to be a good value, because the allure of having THE BEST is worth more than the $500+ difference between a 1080Ti and a Titan (what’s $500 when your bank account has that many zeroes?).

            • DoomGuy64
            • 2 years ago

            You are correct in that value is somewhat subjective. The problem is that the cards I consider to be valuable are not being produced by Nvidia at reasonable prices, and they are deliberately ruining performance for anything that isn’t their most high end products.

            Get rid of Gameworks, and the constant bashing of everyone not purchasing a Ti, and I wouldn’t care that Nvidia sells those cards. Until then, those cards are pure cancer to the industry, and I fully support Nvidia limiting supply of their new chips to $9000 Quadros, or raising Ti prices to match the Quadro.

            The Ti isn’t, and never was a mainstream part. It’s extremely toxic to be marketed as one. Nvidia is trying to force Chevy truck users into buying Lamborghinis. There isn’t anything wrong with owning a Lamborghini by itself. The problem stems from forcing it on people who don’t need or want it, and know the capabilities don’t justify the prices outside of being a niche luxury item.

            • K-L-Waster
            • 2 years ago

            [quote<]Nvidia is trying to force Chevy truck users into buying Lamborghinis.[/quote<] Uhh, no. They aren't forcing anyone to buy anything. They're offering products at a particular price point. People are perfectly free to decline to buy them if they so choose.

            • DoomGuy64
            • 2 years ago

            So you’re just outright denying the existence of Gameworks, and how using those settings requires a Ti?

            No, Nvidia clearly is forcing people to buy the Ti, because they tie every single advanced function that they advertise into high end models.

            You want SLI? Not on the 1060. You want to overclock? Not on the 1050, because no PCIe power connector. You want high resolution? The 1070 is the minimum card capable of decent 1440p, same with VR. You want VRR? Nvidia CLEARLY is FORCING you to buy Gsync.

            There is no flexibility allowed on Nvidia mid-range products, as they tie all their best features into the most expensive products. Claiming that Nvidia is not doing this is a blatant lie, and quite obvious to anyone who actually looks at the capabilities of their lower end cards. You pretty much HAVE to buy the upper tier products to get any decent capabilities out of Nvidia’s cards. The only way you are “free to decline” and have the options you want, is to buy AMD, because it isn’t possible with Nvidia.

          • K-L-Waster
          • 2 years ago

          GV100 isn’t a gaming card, it’s a professional card. It’s predecessor was the Tesla line (GP100), not the 1080 TI or the Titan (GP102).

          I’m not sure what GP100 Teslas cost, but it was in the multiple thousands as well.

      • Krogoth
      • 2 years ago

      Nvidia is milking out Pascal with the crypto-currency craze until it crashes hard and the market gets flooded with “lightly” used Pascal inventory. That’s when they will launch customer-tier versions of their next generation architecture.

    • jihadjoe
    • 2 years ago

    Disappointed NVLink only does two cards. If it supported four then someone can build a workstation with four of these and call it the QuadBro.

      • the
      • 2 years ago

      I looks like nVidia changed nvLink a bit as their store offers special Volta nvLink bridges. Or it could just be the color to match.

      Still being able to share resources via nvLink to stitch together a single large GPU would be interesting to test.

    • just brew it!
    • 2 years ago

    25 years ago, a 50 GFLOP supercomputer filled an entire room. Now you can get the equivalent of nearly 300 of them on a single card…

      • chuckula
      • 2 years ago

      And Krogoth wept.

      • UberGerbil
      • 2 years ago

      I happened to be at a reception at the Living Computer Museum on Friday night (I mean, it was for [url=http://www.livingcomputers.org/News/Announcements/Totally-80s-Rewind.aspx<]this[/url<], how could I not?), and wandering around I found their (sadly non-operational) Cray 1. I can remember going on a tour of Boeing Computer Services in 1986 or so, and being awed when we got to their [url=http://archive.computerhistory.org/resources/text/Cray/Cray.X-MP.1985.102646183.pdf<]Cray X-MP[/url<]. The guy showing us around said he felt like it should be in its own room with dry ice smoke around it whenever the doors slid open, but it was just quietly humming in a corner. You could even sit on it. At the time it was the fastest computer in the world: 105 MHz, 256MB of memory, $15M (plus an extra couple of million for the 1MB SSD with its low 25 μs latency). 400MFLOPS (200MFLOPS per CPU). All of which is now dwarfed by the mid-range ARM chip in the phone in most people's pockets.

        • Mr Bill
        • 2 years ago

        You can’t always get what you want…

          • NTMBK
          • 2 years ago

          But if you try sometimes

          You’ll find

          You get chuckula

            • blastdoor
            • 2 years ago

            Holy forking shirts! You guys, THIS is the bad place!

            • chuckula
            • 2 years ago

            Well it’s true: Chuckula is what you need.

            • DoomGuy64
            • 2 years ago

            He’s got what plants crave. This neutral tasting water stuff sounds dangerous.

        • geniekid
        • 2 years ago

        From the linked PDF: “The optional SSD consists of four columns arranged in a 90 [degree] arc occupying 24 square feet (2.3 square meters) and is connected to the mainframe through one or two short aerial bridgeways, depending on model.”

        The highest option for the SSD was 1GB.

    • chuckula
    • 2 years ago

    [quote<]The Quadro GV100 is available direct from Nvidia today at a price of $8,999.[/quote<] I know what you are thinking, but don't worry: Unless you live in Alaska, Delaware, [b<][i<]Montana[/b<][/i<], New Hampshire, or Oregon, then there's sales tax. So the final price [b<]is[/b<] Over Nine Thousand!

      • Krogoth
      • 2 years ago

      [url<]https://www.youtube.com/watch?v=SiMHTK15Pik[/url<]

Pin It on Pinterest

Share This