Nvidia lets loose a flurry of Volta-based HPC products

Are you salivating over the Nvidia Tesla V100? If you are, then you best get a bib ready. Right after unveiling the V100, Nvidia showed the Volta-powered DGX-1 rack unit, the DGX Station, the HGX-1 GPU cloud computing unit, and a Volta-powered server add-in card.

If one Tesla V100 was already impressive on its own, how about eight of them in a rack unit? That's what the DGX-1 is. Nvidia claims that this box o' fun can replace 400 standard servers for machine learning tasts and should be good to deliver 960 tensor TFLOPS. The DGX-1 uses an NVLink Hybrid Cube interconnect. Price is set at $149K, and Nvidia expects it'll deliver the units in Q3 of 2017. Customers who order the Pascal-powered DGX rack now will get a free upgrade to the Volta version when it's out.

Not everyone can just waltz into their own personal datacenter to do some number-crunching, though. Nvidia has apparently been bagdered over the years with requests to make a Tesla-powered workstation, and the company is now delivering it. The DGX station is a water-cooled tower "PC" packing four Tesla V100 cards, three DisplayPort outputs, and capable of punching through 480 tensor TFLOPS. You'll need a lot of juice to make this machine go at full tilt: all of 1500W. Nvidia will sell you a DGX Station for $69K.

Nvidia also made a Volta-powered "GPU cloud computing" unit called HGX-1, though it didn't offer many details on it. We surmise that this is a version of the previous HGX-1, except it's based on Tesla V100 cards. The box uses a VLink Hybrid Cube interconnect and apparently has eight Tesla V100s inside it.

Last but by no means least, Nvidia made an apparently-unnamed add-in card with a V100 chip aboard tuned for inference tasks. The full-height, half-length appears to connect using a PCIe x16 interface and draws 150W of board power. Nvidia says that 33 "nodes" of V100 (we assume thirty-three of these cards) can deliver inference performance equivalent to 500 servers with 1000 CPUs in them, equating it to a 15x reduction in cost for the same performance.

Comments closed
    • End User
    • 3 years ago

    Saving my pennies for a summer release of the GTX xx80.

    • DeadOfKnight
    • 3 years ago

    [quote<]The DGX station is a water-cooled tower "PC" packing four Tesla V100 cards, three DisplayPort outputs, and capable of punching through 480 tensor TFLOPS.[/quote<] ...but can it run Crysis?

    • Airmantharp
    • 3 years ago

    I’m actually more interested in GV102 and GV104, the consumer/gaming-oriented versions, assuming Nvidia follows their part-to-product naming conventions.

    I think this announcement is more for shareholders, rumormongers, and fanboys ;).

      • Krogoth
      • 3 years ago

      It is entirely for [b<]shareholders[/b<] and professionals. The gaming crowd doesn't about general compute stuff.

    • Krogoth
    • 3 years ago

    Does anybody else getting SGI-vibes from the DGX?

      • tay
      • 3 years ago

      Came here to post this. Nvidia has become the new SGI. Considering how many SGI folk were at 3dfx and nvidia I guess this kind of an expected evolution. Nvidia has given up on the smartphone GPU market (low end) and is really pushing up into the high end.

      I wonder if a new low end player – like nvidia did all those years ago (Riva TNT!!) will emerge to take them on.

        • bjm
        • 3 years ago

        Ironically, the only new players that may even be capable of making an attempt would be those that found success in the smartphone GPU sector.

    • Bumper
    • 3 years ago

    Good on nvidia. Im glad a new market has developed to support more graphics innovation. AI is super cool and it’s great to taste the fruits in consumer cards a year later.

    Btw 15+TFLOPS sounds low for that big of a chip. Am I wrong in thinking that?

      • Liron
      • 3 years ago

      It seems the extra space is mostly taken by the tensor units which deliver 120 TFLOPS on tensor stuff.

        • Bumper
        • 3 years ago

        ok. this thing sounds like a monster.

        • the
        • 3 years ago

        120 TFLOPs sounds impressive until you realize that it is FP16.

          • BryanC
          • 3 years ago

          It’s not, though. The tensor cores are mixed precision.

      • ImSpartacus
      • 3 years ago

      It’s got tons of non-FP32 compute resources that don’t really contribute to that FP32 FLOPS figure.

      [url<]http://images.anandtech.com/doci/11367/volta_sm.png[/url<] Check out those Tensor units. Those exist purely to do a very specific matrix operation. The INT units aren't going to be helpful either. Overall, Nvidia is trying to make a balanced GPU for this rapidly growing market. FP32 performance is just one metric for success.

        • NTMBK
        • 3 years ago

        NVidia’s block diagrams don’t really bear any relation to actual unit sizes on chip.

    • psuedonymous
    • 3 years ago

    OK, a V100 PCIe card right out the of gate has me [b<]VERY[/b<] surprised, doubly so because it is a 'neutered' chip compared to the full V100 (half power, interconnect restricted to 1/30 of NVlink2, and I'd bet probably some HBM stacks disabled). Even if we assume that with the mahoosive die size of v100 they're going to be left with stacks of chips that don't pass QC, it took P100 the better part of a year to go from the premium full-fat big-margin P100 in the GP100 to a cut-down version on a PCIe card. If Nvidia are going to be 'eating their own lunch' right away with 'budget' versions (budget for HPC anyway) of V100 competing with full V100, either they expect the market to be so hungry that they are confidant they are going to sell any and every V100 they can fab at whatever price they set, or Nvidia have gotten even better at fine division binning than with Maxwell and Pascal. In they can fab massive dies that can survive multiple defects and just bin off tiny areas rather than chopping off entire GPCs for only a minimal performance hit, that's a massive increase in [i<]effective[/i<] chip yield.

    • tsk
    • 3 years ago

    Looks like you were right Chuck.

      • Neutronbeam
      • 3 years ago

      Now you’re just encouraging him. On the other hand, I think TR should start paying him.

        • chuckula
        • 3 years ago

        ME THREE!

Pin It on Pinterest

Share This