eWEEK content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More.
Cray is unveiling a powerful new cluster supercomputer packed with GPU accelerators and offering a peak performance of more than 11 teraflops per node to address increasing demands for more performance and scalability in such business segments as oil and gas, financial services and life sciences.
The new CS-Storm offers up to eight Tesla K40 GPU accelerators from Nvidia and two Intel Xeon E5-2600 v2 processors per 2U (3.5-inch) compute server, and each standard 48U (84-inch) rack can hold 22 of the servers and offer more than 250 teraflops. A CS-Storm system with four cabinets can deliver more than a petaflop of peak performance, according to the company.
The new system, which is based on Cray’s CS300 supercomputer, “fills a gap in Cray’s portfolio” to meet the growing demand for more highly accelerated systems that can process thousands of calculations at the same time, according to Barry Bolding, vice president of marketing and business development at Cray. More workloads in the high-performance computing space need higher accelerator performance and bandwidth, and the new system answers that need, Bolding told eWEEK.
OEMs increasingly are using GPU accelerators from Nvidia and Advanced Micro Devices or x86-based co-processors from Intel to help improve the performance of their high-performance computing (HPC) systems while keeping down the power consumption. On the latest Top500 list of the world’s fastest supercomputers published in June, 62 of the systems used accelerators, up from 53 in November 2013. Forty-four used Nvidia GPUs, two used Radeons from AMD, and 17 ran Intel’s Xeon Phi products.
“We definitely see a growing demand from our customers for more accelerated workloads,” he said, adding that not all HPC applications will need all that power. However, an increasing number of applications in such areas as high-frequency trading, digital imaging analysis, weather forecasting and genome sequencing can take advantage of the large number of GPUs on the new system.
Nvidia lists almost 280 current applications that can benefit from GPU-accelerated processing, according to Bolding.
Nvidia introduced the K40 GPU accelerators in November 2013, with officials saying they offer twice the memory and 40 percent the performance of their predecessors, and 10 times the performance of systems using only CPUs.
The CS-Storm will hold up to 176 Tesla K40 GPUs in a rack, and will offer 4.7 times more performance density and 4.4 times more power efficiency than CPU-only systems, Bolding said.
He described an environment that is coming together to push the use of GPU accelerators, from the evolution of the GPUs themselves to the maturation of the surrounding ecosystem, including the CUDA development technology for creating applications that can take advantage of GPU accelerators, application libraries, and programmers that are willing to develop code for them.
Cray also offers a range of capabilities for the CS-Storm system, from its Advanced Cluster Engine management software and programming environment to the power and cooling capabilities that enable the systems to run at a full 300 watts without overheating. Bolding said organizations that use these supercomputers need to be able to run all eight GPU accelerators continuously at top speed, which the company’s air-cooling system enables.
The systems also are built to enable them to embrace future generations of accelerators, though he declined to say whether those future accelerators will only be Nvidia GPUs, or GPUs from AMD or Xeon Phis from Intel.
Such capabilities are what separates Cray’s offering from those developed with eight GPUs in a compute by white-box makers like Tyan and others like SuperMicro, Bolding said.
Cray’s other supercomputers in the XC and CS lines have a ratio of GPU accelerators and CPUs of 1:1; the CS-Storm’s ratio is 8:2.
Organizations that want to take advantage of the CS-Storm’s capabilities will have to spend about $1 million for a fully loaded rack, he said. It’s being used by a limited number of organizations now, and will be generally available in the fourth quarter.