Quadro GP100 brings the most potent Pascal to workstations

Nvidia's GP100 graphics chip has shown up in two different forms for datacenters already: first as a Tesla P100 NVLink mezzanine card for HPC systems, and second as a Tesla P100 PCIe accelerator. The green team has now finished its GP100 hat trick with the Quadro GP100, a graphics card that brings the most potent Pascal chip to workstations for the first time.

It's not this small in real life, we promise

According to Anandtech, the GP100 GPU in this Quadro arrives on the desktop with every bit of its frightening floating-point performance intact. That means FP64 calculations will tick along at 5.3 TFLOPS, FP32 at 10.6 TFLOPS, and FP16 at 20.7 TFLOPS. The card's GP100 GPU shares its interposer with 16GB of HBM2 RAM. Unlike its Tesla counterparts, the Quadro offers four DisplayPort 1.4 outs, one DVI-D out, and optional provisions for stereo goggles.

The Quadro GP100 will also mark the first appearance of Nvidia's NVLink interconnect in a traditional PC-compatible form factor. Using appropriate bridges, workstation users will be able to team a pair of Quadro GP100s for more performance and a 32GB pool of RAM.

The GP100's (note: I hate GPUs and graphics cards having the same name) next-closest sibling in the Quadro lineup is the GP102-powered P6000, but the new hotness doesn't necessarily supersede the older here. The P6000 offers higher single-precision performance at 12 TFLOPs, and it also offers more RAM in the form of 24GB of GDDR5X.

Anandtech notes that the GP102 card can't come anywhere near the GP100's FP64 throughput, however, and it also won't offer the double-rate FP16 performance from packed instructions nor the ECC RAM support that'll likely be critical to demanding professionals. Consequently, the choice of monster Quadro for a given job will be more nuanced than just hammering "buy" on the most expensive option in Dell's or HP's workstation configurators. CAD professionals will be able to get their hands on the Quadro GP100 starting in March.

Comments closed
    • psuedonymous
    • 3 years ago

    Well this has confirmed a few things that were previously only rumoured:

    1) That GP100 actually DOES have the hardware on-die necessary for direct image rendering output.
    2) That NVLink has a PHY variant that can tolerate jumping two card-edge connectors (though with those ‘notched’ bridge designs, I’d guess it’s not passing over a PCB between the two but probably a ribbon of twinax cables or a printed homogenous substrate flex).
    3) That a card can have both NVLink and PCIe active at the same time (unclear yet if the PCIe link is on the other side of an NVLink-PCIe bridge chip though).

      • jts888
      • 3 years ago

      Why do you doubt it’d be a PCB bridge? Stripline traces can’t have much worse performance over those lengths than micro coax ribbons and are massively cheaper to make.

        • willmore
        • 3 years ago

        My guess is that they’re a flex and that they’re twisted as they go over the card. Hence the C shape.

      • the
      • 3 years ago

      1) This was indeed an open question to this point.

      2) Unsure how difficult this was actually. The traces between mezzanine cards looks like it can be longer vs. going over these two edge connectors. The benefits of routing in 3D space vs. 2D (motherboard). There is signal loss going over the edge connectors but similar (but smaller) loss is encountered with the mezzanine pins.

      3) I thought GP100 could always have nvLink and PCIe active simultaneously. This is used for the Xeon + GP100 mezzanine setups. Only the POWER8+ setups using nvLink don’t utilize PCIe with the GP100.

        • psuedonymous
        • 3 years ago

        [quote<]"There is signal loss going over the edge connectors but similar (but smaller) loss is encountered with the mezzanine pins."[/quote<] It's that signal loss that is the key, card-edge is pretty harsh for high-speed connections. More specifically, card-edge connectors are not really [i<]reliable[/i<] connectors, two adjacent fingers may not have the same properties in terms of impedance, signal reflection, etc. A big problem if the two halves of a pair do not match. Pin-based connectors like the mezzanine connectors the compact GP100 uses (or the LGA pin-beds CPUs use) make contact more reliably.

    • tipoo
    • 3 years ago

    128 ROPs connected to HBM2…
    Complete overkill, but that would be a crazy gaming card that could possibly do 5K most of the time ?

      • Jeff Kampman
      • 3 years ago

      I’d love to find out, but getting our hands on one of these could be difficult at best.

        • the
        • 3 years ago

        PLEASE PLEASE PLEASE PLEASE PLEASE
        *puts cherry on top*

    • ptsant
    • 3 years ago

    Crysis?

    • Krogoth
    • 3 years ago

    Methinks a “Titan X Pascal Black” refresh is in the works based on this guy.

      • ImSpartacus
      • 3 years ago

      Really, you expect a 2017 Titan based on GP100? Interesting thought.

      I feel like Nvidia would be more likely to use a 24GB fully enabled GP102 for that role (perhaps using the new 11-12 Gbps GDDR5X just out of sampling).

      Using a >600mm2 chip with four stacks of expensive HBM2 and a big interposer sounds pricey when you could get literally the same gaming performance from a <500mm2 chip using cheaper GDDR5X and no interposer.

      But GP100 is a year old now and will probably be replaced by a “GV100” in the next 12 months, so maybe Nvidia is willing to let loose a little.

        • Krogoth
        • 3 years ago

        Titans are really either just “Quadro” rejects that ate too much power or just Nvidia trying to sell off excessive “Quadro/Tesla” stock.

          • ImSpartacus
          • 3 years ago

          And yet, Nvidia always releases one titan per year like clockwork.

          I think Nvidia puts more thought into the Titan line than you think. It’s their constant halo card and we know how important it is to Nvidia that they maintain a halo card.

            • Krogoth
            • 3 years ago

            The Titan X Pascal is from 2016.

            It is 2017 and Vega is coming out soon. Nvidia is just waiting until it comes out and decides to do the refresh since Volta isn’t ready yet.

            • the
            • 3 years ago

            Volta [i<]might[/i<] arrive at the end of the year. Still that leaves ~6 months with Vega on the market before hand so a Pascal refresh is likely to help fill that void.

            • ImSpartacus
            • 3 years ago

            Yeah, Volta could easily arrive in late 2017, but it’ll do so in the same fashion that Pascal arrived.

            We’ll probably see a GV100 that is dedicated to supercomputers and other high margin pro stuff while the gaming Volta cards arrive in 2018.

            Nvidia is in the business of making money. They aren’t a charity. With Pascal, they proved that they could get away with giving priority to profitable segments first. What makes you think they won’t do it again?

            • the
            • 3 years ago

            No argument that nVidia is in the business of making money.

            However Volta is to arrive on a new process which entails risk for large chips. Yields are a variable that has to be controlled to keep costs in check. This may mean that the larger die products arrive in the second or third wave of releases. Knowledge learned from mass production of the smaller dies is slipped into the larger die designs. nVidia does not a repeat of what happened with the GF100.

            nVidia may announce the GV100 chip first but I would be surprised if it was the first chip of that generation to actually ship.

            • jts888
            • 3 years ago

            TSMC’s 12nm is just a refinement of their 16nm plus gobs of marketing spin. Improvements will likely be very marginal, but if any “transition” can happen quickly, it’ll be this one.

            [b<]Edit:[/b<] If you guys really don't believe me... [url<]http://www.digitimes.com/news/a20161128PD206.html[/url<] TSMC's "12"nm process is a rebrand of the 4th gen/variant of 16 nm they've had planned for a long time.

            • ImSpartacus
            • 3 years ago

            I wouldn’t be surprised in TSMC’s 12nm is a bit of a marketing show based on how long we had 28nm.

            But do you know any sources of info that suggest 12nm will be kinda “meh”?

            I honestly don’t have anything to say otherwise, but I haven’t read much on 12nm, so I’m curious.

            • jts888
            • 3 years ago

            I don’t think anything but the barest of marketing fluff is know to non-NDA partners yet, but my EE friends say their smarter materials science friends say it should be an improvement of “several” percent on one of the half pitch dimensions with slightly better leakage characteristics. So it might yield something on the order of a 10% power or density improvement if we’re lucky, but it’s really more of a node variant than even a half-node jump.

            Other online speculations say the relabel is partially a reaction to GloFo’s slightly earlier announcement of a “12nm” planar FD-SOI, but that’s harder to buy into, since the latter is more of a play on RF processing territory.

            • ImSpartacus
            • 3 years ago

            GP100 arrived on a new process and they made it work.

            I mean, it’s basically at the reticle limit from the get-go. They didn’t even start with done G#104 foreplay, just went in raw with the big boy.

            And it worked great.

            I see no reason why they won’t attempt the same thing again.

            • the
            • 3 years ago

            GP104 shipped first though and from all indications hit fabs first too. (GP100 with its interposer and HBM does take longer to manufacturer and test).

            • ImSpartacus
            • 3 years ago

            Do you have any source that GP104 shipped first?

            [url=http://www.anandtech.com/show/10222/nvidia-announces-tesla-p100-accelerator-pascal-power-for-hpc<]GP100 was announced in April[/url<] in the form of the mezzanine P100 while [url=http://www.anandtech.com/show/10304/nvidia-announces-the-geforce-gtx-1080-1070<]GP104 was announced in May[/url<] in the 1080 & 1070 that we all know and love. It's always tough to determine when a nonconsumer product releases, but Nvidia claimed that the DGX-1 would ship in Q2 2016 while we know the GP104 shipped at the end of May and the end of April. We do know that Nvidia had supercomputer contracts to meet by summer-ish 2016 and I can only assume that those went of without a hitch. And anecdotally from the rumor mill, I recall rumors for GP100 far before GP104 rumors.

            • the
            • 3 years ago

            Well the first production DGX-1 system shipped in August:
            [url<]https://blogs.nvidia.com/blog/2016/08/15/first-ai-supercomputer-openai-elon-musk-deep-learning/[/url<] Per previous communication from nVidia, the GP100 chip would first ship in DGX-1 systems, then stand alone mezzanine cards and then the discrete PCIe. Not mentioned were the super computer contracts which could take president. However, no GP100 cards made it into June's Top500 list so I would have think that if nVidia were shipping before then, a few would have made it into the lower ranks.

            • jts888
            • 3 years ago

            GP100 was near vaporware until quite recently. IIRC, there were still fewer than 100 customers as of early January, most of which were trial installations in research groups in universities and corporations.

            I would be surprised if the number of chips in the wild today exceeded the hundreds, but that should presumably finally start changing if quadros (i.e., not the $130k appliances) are being announced.

            • ImSpartacus
            • 3 years ago

            Wow, I didn’t know that dgx-1 took that long.

            When I said super computers, I meant things like this.

            [url<]https://www.theregister.co.uk/2015/11/16/nvidia_noaa/[/url<] [url<]https://www.hpcwire.com/2016/04/06/europes-fastest-supercomputer-get-pascal-gpu-upgrade/[/url<] I don't know hpc, so I can't tell you what the implementation timelines were in those situations. But it seemed like Nvidia would have to supply a couple thousand gp100 chips for some of these supercomputer contracts throughout spring/summer 2016.

      • NTMBK
      • 3 years ago

      I doubt it; the GP100 would make a worse gaming chip than the GP102. GP102 is far more efficient for FP32 workloads, which makes up the majority of games.

        • Krogoth
        • 3 years ago

        Not for general compute that doesn’t need ECC support. 😉

        • the
        • 3 years ago

        Only if the game itself is shader bound. If it is fill rate limited, this Quadro would be the better solution if 128 ROPs turns out to be true.

        I would predict that the P6000 and GP100 cards switch places depending up on game and resolution.

        • jts888
        • 3 years ago

        Agreed, GP100 uses a lot of space on non-gaming bits:[list<] [*<]fp64 ALUs that will most certainly not be unlocked in <$2k consumer cards [/*<][*<]lots of NVLink controllers and PHYs [/*<][*<]4x HBM2 controllers - i.e., twice the bandwidth of Titan X (Pascal) with a lot less than 2x the compute power [/*<][*<]~14 MiB of aggregate register files meant for more complicated or more concurrent GPGPU warps/wavefront in flight[/*<][/list<] With no guarantees of 128 ROPs or comparably beefed up geometry/rasterization units and the die already near reticle limit, the GP100 feel very unlikely to be a gaming monster to me.

          • ImSpartacus
          • 3 years ago

          Yeah, GP100 would simply be wasteful.

          GP102 exists for a reason.

            • Krogoth
            • 3 years ago

            They are rejected Quadro/Tesla-tier “GP100” cores so you might as well sell as something else rather than recycling them.

            That’s what Titans were especially the Kepler-generation units. That’s why Kepler Titans were such massive hit with the general compete market. You got near-Quadros/Tesla for a fraction of the price where you don’t need ECC or certified drivers.

            The funny part is that aforementioned Kepler-based Titans were a hot-selling items among the non-gaming crowd. Maxwell and current Pascal-based Titans are lackluster by comparison due to crippled general compute and being marginally better then their regular gaming cards.

            • the
            • 3 years ago

            The Maxwell and Kepler Titans weren’t exactly crippled, at least in the traditional binning or market segmentation fashion. Full FP64 support was enabled for the hardware, just that the design didn’t emphasis FP64 unlike the Kepler generation. Curse of being stuck on 28 nm for too long.

            The other oddity is GP102: a chip designed to set between the high end Geforce (GP104) and compute (GP100) as a Titan and serve the Quadro market. I wouldn’t have expected there to be enough volume in those segments to merit its own chip. Similarly it was an open question just how much the GP100 chip had for actual graphics tasks. Apparently enough for this new card being announced so that emphasizes the question of why the GP102 chip exists. nVidia could utilize GP100 for the Quadro and compute markets from the start. The benefit to nVidia is the higher volume of sales for a particular chip and savings by not having to develop another piece of silicon.

            • ImSpartacus
            • 3 years ago

            What about the Titan Black?

            That was Kepler and it ended up being effectively the best GK110 implementation that anyone could ever get.

            And it was the “refresh” of the previously cut-down Titan.

            What makes you think we won’t see the exact same situation with the 2016 & 2017 Titans? The 2016 Titan is a cut-down GP102 with mediocre clocks and the 2017 Titan is a fully enabled GP102 with higher clocks and more VRAM.

            That seems much more plausible than the scenario where Nvidia decides to switch to a very expensive part that won’t game any better than an equivalent GP102.

            I’ll say it again, GP102 exists for a reason. We haven’t had a G#102 before, so I think there’s a very good reason for its existence.

          • the
          • 3 years ago

          It really depends if the expected ROP figure is really 128. If so, then it would push more pixel than the GP 102. That would enable the GP100 to perform better at higher resolutions than the GP100 in fillrate limited scenarios. Even then, the

          nvLink may improve SLI scaling versus the previous generation connector. This may turn out to be advantageous to gaming. However nVidia only supports dual card SLI scaling for gaming now, so the additional nvLinks to support 8-way SLI is indeed wasteful in that context.

          The larger register file may also be advantageous for gaming, though how much likely depends on the game and how well nVidia’s shader complier is vs. standard Pascal. More overkill than wasteful.

    • juzz86
    • 3 years ago

    I like the look of the NVLink bridges, very tidy.

      • dodozoid
      • 3 years ago

      well, it is a render…
      someone would have to be fired immediately if even renders didnt look tidy

        • willmore
        • 3 years ago

        Don’t worry, gritty will be back in favor before too long.

        • juzz86
        • 3 years ago

        Tell that to every motherboard manufacturer ever. Can’t even line the ports up sometimes!

    • ImSpartacus
    • 3 years ago

    Yeah, kinda shitty that this is literally named “GP100”. It forces you to preface with “Quadro” to ensure clarity.

      • lycium
      • 3 years ago

      After they started adding the “Pascal” suffix to Titan X (the X already a later-added suffix) basically their marketing team took a massive bong rip and declared “trolololol it actually doesn’t matter what we call it, to the moon with suffixes and prefixes!”

        • jihadjoe
        • 3 years ago

        I think they were trying to do an Apple (or like car makers) and call each model the same thing. Actually not a bad idea, IMO; but they forgot to include a differentiator like the date.

        • ImSpartacus
        • 3 years ago

        The most recent “Titan X” is just “Titan X”.

        “Titan X Pascal” is just a name that the community came up with.

        Read the official page for the card. It mentions “Pascal” often, but never makes it part of the official name.

        [url<]https://www.nvidia.com/en-us/geforce/products/10series/titan-x-pascal/[/url<] Personally, I wouldn't be surprised if the 2017 Titan also gets named "Titan X" and we end up unofficially using years to talk about the various models, e.g. "2016 Titan X," "2015 Titan X," etc.

      • Chrispy_
      • 3 years ago

      I wish they’d do this for all their cards.

      1080 = GP104
      1070 = GP104 LE
      1060 6G = GP106
      1060 3G = GP106 LE
      1050 Ti = GP108
      1050 = GP108LE

      I mean that stack of 6 is a mess all by itself;

      Harvested GP104 dies are given their own new product segment.
      Harvested GP106 dies are differentiated by memory size.
      Harvested GP108 dies lose the Ti moniker.
      What’s next, will harvested GP102 dies be differentiated by [i<]his 'n hers[/i<]?

        • jts888
        • 3 years ago

        The ~$1200 TXP is already the clipped GP102 variant of the ~$5k P6000.
        Differently binned and clocked GP102s are the most likely candidates for a 1080 Ti and TXBP should Big Vega be strong enough to merit new models.

Pin It on Pinterest

Share This