Cray XC50 supercomputer can probably run Crysis

Cray isn't a name you hear about every day. For the unitiated, the company has been making supercomputers ever since "computers" were a thing. Today at the SC16 conference, the company pulled back the curtain on the XC50, which it calls its fastest supercomputer ever. Cray says the XC50 ought to offer 1 PFLOP (yes, that's petaflops) per cabinet—and the entities buying these don't tend to stop at just one.

Each XC50 cabinet comprises three of what Cray calls "chassis." A chassis is in turn composed of 16 "compute blades," and each blade carries four nodes. Each node supports Xeon and Xeon Phi CPUs and Nvidia Tesla GPU accelerators. Just to give you an idea of what kind of horsepower we're talking about here, a single Tesla P100 Accelerator Blade offers four nodes, each with a Xeon host CPU and a Tesla P100 PCIe card. If all the nodes are stuffed with the maximum 128GB of RAM, that works out to a total of 24TB of RAM and 1.04 PFLOPS of computing power per cabinet. Yow.

 Those blazing-fast nodes wouldn't be much good if they couldn't talk to one another, and Cray includes one of its Aries routers in each blade. Each router is part of a Dragonfly network topology that handles intra-node, intra-chassis, and inter-chassis communications. That system cuts network latencies to just nanoseconds even when every node in a XC50 is talking to every other one at once.

The end result is that all this hardware can be put to work processing data at up to 500 PFLOPS. Before you ask, no, you can't afford it. But maybe your employer can.

Comments closed
    • Tulkas
    • 3 years ago

    Intel Xeon processors does not support NVlink, only the upcoming IBM Power9 and later versions of Power8 processors.

    • tipoo
    • 3 years ago

    Crysis turned 9 two days ago, so we’ll allow this one 😛

    Still looks rather incredible…If you take Crysis maxed out on a modern PC and said it was a game coming in the next year, people who didn’t know it by sight would totally believe it.

    [url<]http://www.neogaf.com/forum/showthread.php?t=1313628[/url<]

      • drfish
      • 3 years ago

      Crysis is the only game where the HD remaster would be a downgrade.

      You’re making me want to start a “let’s play Crysis again” movement for the 10 year anniversary.

    • Chz
    • 3 years ago

    I have an interest in the history of supercomputing, so pardon me for being a pedant here.

    Cray, as a company, is a relative newcomer. Aside from the name, barely anything of Cray Research survives to this day. Even then, old Seymour was making supercomputers for Control Data before he left and set up his own shop. Granted, it’s a matter of semantics as to whether the CDC-6600 or the Cray-1 was the first “super” but they were designed (and built by hand!) by the same guy, and the company(s) he founded no longer exist in any meaningful form.

    That’s not to say that Cray isn’t doing some cool things and shouldn’t be commended on it, but they’ve really nothing at all to do with Seymour Cray beyond having a license to use the name.

      • chuckula
      • 3 years ago

      IIRC, Cray the man was never a fan of clustered compute systems, which is ironic since basically all of the major supercomputers in existence are clustered systems.

        • Concupiscence
        • 3 years ago

        “If you were plowing a field, which would you rather use: Two strong oxen or 1024 chickens?” That was his quote, and it’s not that he didn’t have a point, it’s just a lousy analogy given the advances of technology over time.

          • faramir
          • 3 years ago

          He did have a point but didn’t have a clue (or perfect hindsight 😀 ) about the future of (super)computing. Then again who can really blame him, Intel was still dreaming about 10 GHz superfast CPUs ten years later …

      • just brew it!
      • 3 years ago

      For a while in the ’90s there were two Crays – Cray Research, and Cray Computer (which Seymour Cray founded after leaving Cray Research). Cray Computer filed for bankruptcy after a few years, but its ghost apparently still lives on as a company called SRC Computers.

      I remember seeing the Cray Computer booth at a trade show back in the ’90s. There was nobody hanging around the booth aside from the employees manning it (everyone was basically moving on to massively parallel designs at that point so interest in Cray Computer’s systems was waning), and the booth’s physical appearance even resembled a graveyard, with things that looked kind of like granite tombstones. The Cray people overheard a co-worker and me joking about the graveyard vibe; they were not amused.

      In that same general timeframe I did some work porting code to the Cray Research T3D system, as part of my job at Fermilab.

    • ptsant
    • 3 years ago

    A few of these with some deep learning software are probably more intelligent than the average human out there. Submit to your machine overlords.

    • Ushio01
    • 3 years ago

    Hmm the supercomputer top500 November should be out now.

    Yep one of these cabinets would get you to 118th place, damn.

    • drkskwlkr
    • 3 years ago

    Man, the cost of Windows Server 2016 licenses for this thing will ruin me.

      • Pwnstar
      • 3 years ago

      That’s what bittorrent is for.

    • CampinCarl
    • 3 years ago

    Actually, some of us would stop at 1 cabinet.

    That’s seriously impressive, Cray.

    • ozzuneoj
    • 3 years ago

    Some absurdly rich person should start a LAN gaming café, utilizing one of these cabinets, with each independent “node” networked with nanosecond pings.

    I wonder if you can have them downgrade the massively powerful Tesla compute units to something more pedestrian with video output, like a 1080 Ti. Each cabinet would house several independent systems, hooked up to a display for each one.

    No idea if this is actually possible, but it’s nice to think about…

      • Beahmont
      • 3 years ago

      It is an interesting thought concept, but I think you’d run into problems with latency in the cabinet to monitor connections unless you’re going to have your multi-million dollar machine in the middle of your gaming area.

    • SnowboardingTobi
    • 3 years ago

    But can it play global thermonuclear war?

    • RAGEPRO
    • 3 years ago

    [quote<]Before you ask, no, you can't afford it. But maybe your employer can.[/quote<]Hey, you don't know man. Maybe my bankroll is [b<]YUGE[/b<].

      • barich
      • 3 years ago

      Wouldn’t you have upgraded to a Rage 128 by now?

        • lycium
        • 3 years ago

        He’s considering splurging on a Riva TNT

          • Klimax
          • 3 years ago

          Can’t he wait a bit for those incoming Geforce? I heard they have that nice HW for T&L! That’ll be huge!

    • Waco
    • 3 years ago

    The bulk of those FLOPs are hidden behind PCIe though, so you better hope your dataset is small enough to not shuffle data every cycle.

      • TheRazorsEdge
      • 3 years ago

      They’re connecting with Aries routers, I’d be shocked if they aren’t using NVlink. It’s not like there are petty financial concerns at this scale.

        • Waco
        • 3 years ago

        NVlink is a POWER thing, not Intel…

          • TheRazorsEdge
          • 3 years ago

          Yeah, I skimmed and caught the Xeon Phi and Tesla P100 options but missed the regular Xeon CPU mention.

          They still get 16 GB/sec with PCIe, which is comparable to the main memory bandwidth of some modern computers. Not bad, just not as badass as the internode connectivity.

            • Waco
            • 3 years ago

            Internode on this isn’t 16 GB/s though. Feeding GPUs for simulation is hard, because PCIe is so slow compared to memory bandwidth (on HPC nodes).

            • TheRazorsEdge
            • 3 years ago

            True, internode is only ~10 GB/s for direct transfers—the 16 GB/s PCIe link is bottlenecked by the 2-port NIC links on the Aries. So the nodes are limited to 2/3 of their 16x PCIe connection.

            If they added 4 ports per router (one per NIC), the NICs could run flat out. Maybe 15 GB/sec is more than most people needed, or maybe they couldn’t make it work without sacrificing latency.

            Granted, it’s not much compared to the HBM2 bandwidth on the P100 itself. But still, 80 Gbps with 2-microsecond end-to-end latency is awesome.

            • Waco
            • 3 years ago

            Keeping those beasts fed requires a lot more than 15 GB/s. Think an order of magnitude better at the minimum for real simulation work…per P100.

          • DarkStar1
          • 3 years ago

          Not quite. NVlink is an Nvidia thing – DGX-1 systems use dual Xeons accompanied by eight P100s connected via NV-Link. I believe it was codeveloped with IBM, but it is not owned or exclusive to them.

            • Waco
            • 3 years ago

            Ah, good to know! I haven’t stayed in the loop much since they renamed it.

            • DarkStar1
            • 3 years ago

            Definitely good to know – hopefully it means those P100s stay well-fed.

    • tahir2
    • 3 years ago

    OK two things

    1. We know it can play Crysis but can it upsample a 1080p image to 1440p, upscale it to 4k and then downsample to 1080p?
    2. I employ myself and my cheap-ass employer told me I can’t buy one on company expenses… bummer!

Pin It on Pinterest

Share This