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.