Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
Cray XK6 System (cray.com)
12 points by evo_9 on April 17, 2012 | hide | past | favorite | 7 comments


Check out this page: http://www.cray.com/Products/XK6/Technology.aspx The network hardware is connected directly to the HyperTransport bus!


So Cray used to make unique, iconic supercomputers and now their secret sauce has been reduced to some interconnect for gluing together mass-market bits like the next ODM?

Who drove them out of the supercomputer market? I know IBM still sells a bunch of mainframes running operating systems you and I have never heard of or used.


You can no longer make a supercomputer the old Cray way, by having Seymour Cray design the fastest, most optimized CPU he possibly can and sticking it in a box. Now you take thousands of processors (with an increasing trend toward low-power low-heat processors rather than just the most powerful ones) and glue them together--and the interconnect becomes one of the most important parts!

Huge companies like IBM can afford to make their own custom chips for their supercomputers (see Blue Gene L, P, and Q), but Cray isn't in that category. They do create a lot of software for using their systems, and the hardware is very well integrated, which is probably why they still do pretty good selling to national labs.

We're realizing that now while you can buy a bunch of computers and a 10 GbE switch, it can be worth it to get a nice fully-integrated supercomputer shipped to your door. It's ready to go, the software is installed, the cables are the right length (and neatly arranged!), and when a node burns out you just mail it back to Cray for a replacement.


That "interconnect" is the secret sauce of any modern supercomputer. Cray's proprietary link connects CPUs directly, almost like they were all on the same motherboard. It's easy enough to buy tons of cores. The hard part is reducing latency between them, and maintaining reliability if a link fails. Cray's networking package forms a "hypetorus" to reduce hops between all nodes, and will automatically route around damage. It's incredibly intelligent for such a fast system.

And I guess you didn't see the software environment either? It's really impressive. Check out the "Brochure" PDF for a lot more information.


These days racks are so standardized it would be difficult, but it would be cool if they made more of a nod to their past designs.


That shit cray.


Darn. You beat me to it.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: