Cray® XC™ Series: Adaptive Supercomputing
Extreme Scalability and Sustained Performance
Cray has an established reputation for regularly running the biggest jobs on the largest numbers of nodes in the HPC industry. The Cray® XC™ series puts even more focus on solving extreme computational challenges. Cray XC series systems scale hardware, networking and software across a broad performance spectrum to deliver true sustained, real-world production performance.
Production-Ready Arm-Based Supercomputing
With the addition of Arm-based processors, Cray XC50 users now have options for choice and flexibility. Compute blades can be mixed and matched with Intel® Xeon® Scalable processors, Intel® Xeon Phi™ processors and NVIDIA® Tesla® GPU accelerators. The expanded ecosystem makes HPC more accessible for a wide variety of applications and users.
A production Arm-based Cray XC50 supercomputer featuring a complete HPC-optimized software stack, including the Cray Linux® Environment and Cray Programming Environment, gives researchers, scientists and engineers complete access to libraries and tools optimized for running complex HPC workloads.
Aries™ Interconnect and Dragonfly Topology
To provide this breakthrough performance and scalability, Cray XC series supercomputers integrate the HPC-optimized Aries interconnect. This innovative network technology, implemented with a high-bandwidth, low-diameter topology called Dragonfly, provides substantial improvements on all of the network performance metrics for HPC: bandwidth, latency, message rate and more. Delivering global bandwidth scalability at reasonable cost across a distributed memory system, this network gives programmers global access to all of the memory of parallel applications and supports the most demanding global communication patterns. The open architecture of the Cray XC series offers intranode flexibility, empowering users to run applications with either scalar or accelerator processing elements depending on their requirements for parallelism.
Cray XC series systems exploit a Dragonfly network topology — constructed from a configurable mix of backplane, copper and optical links — providing scalable global bandwidth, avoiding expensive external switches and enabling easy in-place processor upgrades. Cray XC air-cooled systems utilize backplane and copper cabling only to reduce costs for simulation, test, development, analytics and AI environments. Depending on whether a cabinet is air- or liquid-cooled, each cabinet can be populated with up to three chassis (one for air-cooled systems), culminating in up to 384 sockets per cabinet.
The Aries ASIC provides the network interconnect for compute nodes on Cray XC series base blades and implements a standard PCI Express Gen3 host interface, supporting a wide range of HPC processing compute engines. The universal nature of Cray XC series open architecture allows the system to be configured with the best available devices today, then augmented or upgraded in the future with the user’s choice of processors, coprocessors and accelerators using processor daughter cards.
Intel® Xeon® and Intel® Xeon Phi™ Processors
For Intel processor support, Cray XC series compute blades come in a variety of configurations.
- Intel Xeon Phi and Intel Xeon Scalable processors are packaged in a dual-socket node configuration, with four nodes supported in each system blade. The Intel Xeon Scalable processors provide more than 10,000 cores and enable more than 700 teraflops per Cray XC liquid-cooled cabinet, and 230 teraflops per Cray XC air-cooled cabinet.
- Intel Xeon Phi processors are packed in a single-socket node configuration, with four nodes supported in each system blade, providing up to 586 TF/cabinet of peak performance.
NVIDIA® Tesla® GPU Accelerators
Cray XC series supercomputers support CPU-hosted NVIDIA Tesla GPU accelerators. Two options are available: the NVIDIA Tesla K40 for the XC40 system and the NVIDIA Tesla P100 PCIe for the Cray XC50 system. NVIDIA’s P100 GPU accelerator delivers over 3,500 embedded cores and flexible mixed-precision computing options. The P100 offers flexible double-precision, single-precision or half-precision compute operation and also integrates high-bandwidth memory into the package, enabling up to 3x memory bandwidth improvements over prior-generation external-memory GPU solutions.
The Cray XC50 system with the Tesla P100 delivers superior application performance, memory bandwidth and performance per watt. Cray also supports multiple programming models for the P100 GPU accelerator, including the Cray compiler, OpenACC directives-based coding and CUDA.
Cray XC series supercomputers support Cavium ThunderX2™ processors. With 16 DDR4 DIMMs per node, users can access up to 512 GB per node at up to 240 GB/s, addressing a common performance bottleneck in high-performance computing workloads.
The ThunderX2 processor is supported by Cray’s complete HPC-optimized software stack including the Cray Linux Environment and premium programming environment. The enhanced Cray compiler and programming environment improve the performance of the Cavium ThunderX2 processors.
Custom and ISV Jobs on the Same System — Extreme Scale and Cluster Compatibility
Based on generations of experience with both environments, Cray has leveraged a single machine architecture to run both highly scalable custom workloads as well as industry-standard ISV jobs via the powerful Cray Linux Environment (CLE). CLE enables a Cluster Compatibility Mode (CCM) to run Linux/x86 versions of ISV software without any requirement for porting, recompiling or relinking. Alternatively, Cray’s Extreme Scalability Mode (ESM) can be set to run in a performance-optimized scenario for custom codes. These flexible and optimized operation modes are dynamic and available to the user on an individual job basis.
ROI, Upgradeability and Investment Protection
Besides being customizable for each user’s requirements, the Cray XC series supercomputer architecture is engineered for easy, flexible upgrades and expansion, a benefit that prolongs its productive lifetime and the user’s investment. As new technology advancements become available, users can take advantage of these next-generation progressions deep into the life cycle before ever considering replacing an HPC system.