Cray announced Wednesday that its next-generation Cascade supercomputer will now include the option of Nvidia’s next-generation “Kepler” GPUs.
The Kepler GPU, part of the Nvidia Tesla family, will be offered alongside the Intel Xeon Phi coprocessor, allowing customers to customize the final design to their high-performance computing tastes. Cray expects the Cascade supercomputer to be “widely available” in 2013.
Nvidia did not say which Kepler product would be included, although it could be the Tesla K10, an accelerator that contains a pair of 2 Kepler GK104 cores. Nvidia claims the K10 is appropriate for applications like seismic processing, signal and image processing, and video analytics. Performance-wise, the combined GK104 cores offer 190 Gflops of peak double-precision floating-point performance, 4,577 Gflops of single floating point performance, and 320 Gbytes of memory bandwidth, with ECC off.
Unlike a traditional microprocessor, graphics chips are optimized for dedicated and specific operations repeated millions of times per second. That means a traditional GPU cannot actually serve as the foundation for a server. But Nvidia’s architecture (as well as similar cards from ATI) has been designed to bridge the gap between GPU and “general-purpose” graphics processing, serving as effective coprocessors for visualization and scientific computing.
“This is an exciting announcement for us, and for our customers, as it proves that we remain committed to our vision of integrating a range of advanced processing technologies into a single, scalable architecture,” said Peg Williams, Cray’s senior vice president of high performance computing systems. “Adaptive Supercomputing is about building a robust hardware and software environment that ultimately supports giving our customers choices.”
The Cascade will include improvements to the Cray Linux Environment, Cray’s HPC-optimized programming environment. In addition, Cray will now support the Intel Xeon CPU, not just the Xeon Phi—a first for Cray’s high-end systems. Cascade will also include the improved Aries interconnect technology.
Combining Aries and Intel’s Xeon is probably not surprising, of course, given that Cray sold off its interconnect business—including the “Aries” and “Gemini” interconnect intellectual property and engineers—to Intel this past April for $140 million.
Intel’s Xeon Phi, launched in June, delivers teraflop performance on highly parallelized applications. A list of Intel’s own benchmarks for the Xeon Phi can be found here.
Although AMD has not yet announced its latest APUs, this week it launched the FirePro W9000, W8000, W7000 and W5000 cards for the workstation. They’re all based on what AMD calls Graphics Core Next, AMD’s first architecture designed for both graphics as well as computational computing.
Cray has announced a number of design wins for the Cascade: a $40 million contract for a Cascade and Conexion storage system at the Department of Energy’s National Energy Research Scientific Computing Center (NERSC) in Berkeley, California; a similar $21 million contract with Commonwealth Scientific and Industrial Research Organization (CSIRO) to provide the Pawsey Centre in Perth, Australia; plus deals at High Performance Computing Centre Stuttgart in Suttgart, Germany and Kyoto University in Japan.