Since 1986 - Covering the Fastest Computers in the World and the People Who Run Them

Language Flags
February 2, 2012

Cray Adds Baby Blue Waters to HPC Lineup

Michael Feldman

Cray has launched a new cut-down configuration of its XE6/XK6 supercomputer targeting the mid-range market for high performance computing. The XE6 is Cray’s flagship supercomputer based on the company’s Gemini system network and AMD’s Opteron 6200 (“Interlagos”) CPUs; the XK6 is an XE6 variant that pairs the Opterons with NVIDIA Tesla GPUs. It is this architecture that will power two of the most powerful multi-petaflop supercomputers in the world: Titan at ORNL and Blue Waters at NCSA.

Essentially Cray has come up with small node-count configurations of these machines that start at about $200 thousand. According to Barry Bolding, Cray’s VP of Storage & Data Management and Corporate Marketing, that price will get you a 48-socket machine that delivers about 6.5 teraflops. The price also includes some storage, the basic XE6/XK6 software stack and development tools, and a standard warranty. Adding GPU blades for a mini XK6 configuration would run up the cost accordingly. “Basically, you can get a mini Blue Waters supercomputer for $200 thousand,” Bolding told HPCwire.

Prior to this, the smallest XE6/XK6 you could buy from Cray would be their XE6m/XK6m variants, but the starting price there was in the neighborhood of $500 thousand. The new mid-range machines are essentially variants of those variants. Reducing the socket count further enabled the engineers to come up with a less dense configuration and employ less costly power distribution units (PDUs), thereby lowering the overall price.

Basically, the XE6/XK6 line now scales from $200 thousand to $200 million. Upgrading from one end to the other would entail more than just adding boxes, but it is unlikely if anyone would want to make that 1000-fold jump. The significant aspect for users is that the software running on a baby Blue Waters should run without change on the 20 petaflops version.

An enabling technology that makes this all possible is Cray’s cluster compatibility mode (CCM) that the company introduced into its Linux bundle in 2010. CCM allows users to run standard Linux codes from ISVs and others on Cray’s proprietary Gemini network. The idea was to open up Cray’s flagship supercomputing line to commercial applications, allowing the company to penetrate industries like manufacturing and aerospace, which depend upon a lot of ISV software.

Cray has achieved some success with the CCM capability, selling systems to GE, Exxon, and a few others. Of course, a lot of companies can’t afford $500K-plus machines, so the new mid-range configuration could help the supercomputer maker collect more sales in the commercial HPC market from firms with more modest budgets.

The new baby XE6/XK6 configurations replace the CX lineup that Cray has been offering for a couple of years. Specifically, the mini supers replace the CX1000 HPC cluster, which, along with the CX1 deskside machine, will no longer be offered. The CX1000 was the company’s entry- and mid-level “supercomputer” offering, built like a traditional HPC cluster, that is, with Intel Xeon CPUs (and optional Tesla GPUs) glued together with InfiniBand. These systems scaled from up to 8 blades (16 sockets) on the CX1 to up to 16 blades (32 sockets) on the CX1000.

The CX line was using “Nehalem” Xeon CPUs, so by coming up with the baby XE6/XK6 machines, the company was able to forego the trouble of upgrading those systems to the Sandy Bridge Xeon chipsets and possibly FDR InfiniBand. It also allows the company to dump the rest of the ecosystem it was dragging along with this lineup, including support for Microsoft Windows HPC, various cluster managers, and so on.

So for the time being, Cray will, once again, be without an Intel-based offering, but presumably that will be remedied when the “Cascade” supercomputer is launched. At that point, one would expect the company to offer the same spread of capabilities — from entry point HPC to elite supercomputing — across two x86 architectures.