Cray has introduced a new line of entry- and mid-level HPC systems, making good on its promise to fill the gap between its CX1 deskside systems and XT mini-supers. Called the CX1000, the new machine makes use of the latest Intel Xeon processors and, from a performance standpoint, picks up where the CX1 leaves off. Cray intends to leverage its existing CX1 ecosystem of more than 40 channel partners to sell and support the new product line.
While conceptually similar to the blade design of the CX1 deskside machine, the CX1000 is delivered in a more traditional chassis form factor and is designed to be installed in the typical datacenter environment, rather than the office. The CX1 theme of making high performance computing more a turnkey kind of experience is the same though. Like its CX1 predecessor, the CX1000 comes preinstalled with an integrated OS/cluster management stack, in this case, either Microsoft Windows HPC Server 2008 or a combo of Linux Red Hat plus the Cray Cluster Manager. Configurations of one to four chassis can be built, with prices that ranges from under $100,000 for a minimally-configured enclosure, up to around $700,000 for a full cabinet with high-end options.
Cray’s rationale for the CX1000 line is two-fold: offer an upgrade path from its CX1 line and provide a choice of architectures that are currently most favored by HPC customers. CX1, you’ll remember, was positioned as a product that could transition technical workstation users into the world of HPC mini-clusters. According to Ian Miller, Cray’s senior VP of the Productivity Solutions Group and Marketing, it just made sense to expand the CX family beyond the 8-blade limit of the CX1. According to him, the CX1000 will meet the needs of customers looking for additional computational power, but who don’t require the supercomputing capability of an XT machine.
To cover its architectural bases, Cray offers three models: a scale-up cluster (CX1000-C), a GPU-accelerated cluster (CX1000-G), and a scale-out SMP-type machine (CX1000-S). We’ll take these in order.
The CX1000-C fills the role of the standard entry-level cluster for distributed memory-style HPC. The C model comes in a 7U chassis that can house up to 16 dual-socket blades, where the sockets are populated by Intel Xeon 5600 series (“Westmere EP”) parts. Processor choices include both the 6-core and 4-core versions, but all running at under 3 GHz. (The top-of-the-line 130 watt Westmere parts are not an option here since these would tend to run too hot for the dense blade design.) QDR InfiniBand is used for the system interconnect, with an optional Ethernet switch available for connecting to an external network.
The CX1000-G is the GPU-accelerated variation of the C model. The double wide blades pair two Westmere EP chips with two NVIDIA Tesla modules, with a dedicated I/O hub for each CPU-GPU pair. The description on the Cray Web site specifies the current M1060 Tesla module, but since NVIDIA is expected to ship the new Fermi products in Q2, most users will probably hold out for souped-up GPUs. Since the blades are double wide, only eight blades will fit in the 7U chassis. That’s nothing to scoff at though. With Fermi parts, a fully-populated enclosure should deliver in the neighborhood of 10 double precision teraflops.
The CX1000-S is the odd one out. This model implements a mid-range SMP machine and is designed for non-distributed, big memory applications, such as electronic design automation (EDA). Most of the specs for the CX1000-S are still forthcoming, but Cray has divulged each node can house up to 128 cores and will make use of Intel’s QuickPath Interconnect (QPI) technology. That almost certainly means the CX1000-S will be using Intel’s 8-core Nehalem EX chips, which are expected to be released on March 30. Since Nehalem EX only supports eight sockets natively, Cray apparently has designed or repurposed a custom node controller to make a 16-socket (128 cores) machine possible.
As an aside, the introduction of the Westmere EP processor and the upcoming Nehalem EX and Fermi parts should encourage other vendors to expand their HPC portfolios as well. Thanks to the latter two chips in particular, SMP and GPU-accelerated computing now look much more attractive from a price-performance perspective than they ever have in the past. System makers should be able to build relatively-powerful SMP machines and GPU clusters for well under a million dollars, as Cray has done with the CX1000.
It’s not just about adding new platforms, though. Miller says it’s possible to mix and match the different CX1000 models into a single cabinet, depending on application needs. Both the Cray Cluster Manager (based on Platform Computing LSF technology) and Windows HPC Server are capable of provisioning these types of heterogenous environments. Dual booting is also supported, so Windows and Linux apps can be accommodated in the same system. In conjunction with that capability, Cray is working with ISVs to help port popular technical computing software onto these platforms so that customers can bring their codes with them. The ultimate goal, says Miller, is to “create a package for entry- to mid-range customers that helps them get productive quickly.”