Cray XK7
| Company type | Product / Supercomputer system |
|---|---|
| Industry | Supercomputing |
| Founded | October 29, 2012 |
| Products | Supercomputing systems (hybrid CPU-GPU architecture) |
| Parent | Cray |
XK7 represents a supercomputing system engineered by Cray and introduced on October 29, 2012. This marks Cray's second hybrid architecture that integrates both central processing units (CPUs) and graphical processing units (GPUs), necessitating programming strategies distinct from those used in CPU-exclusive environments. Organizations utilizing XK7 platforms often organize training sessions to familiarize researchers with these specialized programming paradigms. Among its notable deployments, XK7 powers Titan, which was ranked as the world's second fastest supercomputer in the November 2013 TOP500 listing. Other prominent users include the Swiss National Supercomputing Centre, operating a 272-node configuration, and Blue Waters, which runs a hybrid system combining Cray XE6 and XK7 nodes capable of approximately 1 petaFLOPS performance.
Overview
The XK7 architecture can be expanded to accommodate up to 500 cabinets, each cabinet containing 24 blades. Every blade incorporates four nodes, with each node equipped with one CPU and one GPU. The CPUs are 16-core Opteron 6200 Interlagos series processors, coupled with Nvidia Tesla K20 Kepler GPUs. Memory configurations include either 16 or 32 GB of ECC RAM for CPUs, while GPUs feature ECC memory capacities of either 5 or 6 GB depending on the specific model. Nodes are interconnected through Gemini Interconnect chips, where each chip links two nodes and provides a bandwidth of 160 GB/s. Power consumption per fully populated cabinet ranges from 45 to 54.1 kW, requiring either air or liquid cooling solutions.
XK7 systems operate on the Cray Linux Environment, which is derived from SUSE Linux Enterprise Server. Although the platform supports various programming languages, its hybrid nature demands tailored coding techniques; consequently, leading research institutions offer educational programs to aid developers in adapting to these requirements.
Usage
The XK7 platform was officially launched on October 29, 2012, coinciding with the completion of the Titan supercomputer at Oak Ridge National Laboratory (ORNL). Titan comprises 18,688 XK7 nodes, each outfitted with an Opteron 6274 CPU featuring 32 GB of memory and a K20X GPU with 6 GB of memory. The system achieves a theoretical maximum performance of 27.1 petaFLOPS and recorded a LINPACK benchmark result of 17.59 petaFLOPS, earning the top position in the November 2012 TOP500 ranking. Titan’s power draw is approximately 8.2 MW, placing it third on the Green500 list for energy-efficient supercomputers.
Illinois’ National Center for Supercomputing Applications (NCSA) operates Blue Waters, which integrates 3,072 XK7 nodes alongside 22,752 Cray XE6 nodes. The XE6 nodes are equipped with dual Opteron 6276 CPUs each possessing 32 GB of memory, while XK7 nodes contain Opteron 6276 CPUs paired with K20X GPUs. Blue Waters has achieved benchmark performances exceeding one petaFLOPS but has opted not to submit LINPACK results due to project leadership’s reservations regarding that benchmark’s applicability.
The Swiss National Supercomputing Centre (CSCS) upgraded its Todi system to the XK7 platform in October 2012. This system consists of 272 nodes featuring Opteron 6272 CPUs with 32 GB memory and K20X GPUs with 6 GB memory. Todi delivers a peak computational capacity of 393 teraFLOPS and attained a LINPACK score of 274 teraFLOPS, ranking it 91st in the November 2012 TOP500 list. With a power consumption of approximately 122 kW, Todi holds fourth place on the Green500 list following Titan.