IBM Unveils Second Cell-Based Blade For High-Performance Computing

With a peak performance of 460 giga floating point operations per second, a single BladeCenter chassis of QS21s can achieve 6.4 teraflops, IBM claims.
IBM on Wednesday unveiled a Cell processor-based blade server that can run in the same BladeCenter chassis as blades currently powered by x86 or Power PC chips. IBM is positioning the revised Cell blade servers for medical imaging, oil and gas exploration, financial analytics, and other high-performance computing applications.

The BladeCenter QS21 is the second blade server released by IBM in less than a year using the Cell Broadband Engine, the same eight-core microprocessor used in Sony's PlayStation 3 video-game console. IBM developed the Cell chip with Sony and Toshiba.

The QS21 is different than the previous QS20 in that the new blade can operate alongside other blades running in a standard BladeCenter. "You can co-mingle it with other blades anyway you want," Paula Richards, product line management for Cell Systems at IBM, told InformationWeek. In addition, the QS21 delivers twice the memory, density, and input/output throughput as the previous version.

A maximum of 14 QS21s can fit in a BladeCenter versus the seven that max out the QS20. In addition, the QS21 offers an additional 2Gbytes of I/O buffer memory. The I/O throughput, which is a measurement of the ability of data to travel in and out of the system, supports 16 lanes of single data rate InfiniBand. With a peak performance of 460 giga floating point operations per second, a single BladeCenter chassis of QS21s can achieve 6.4 teraflops, while a standard 42U rack more than 25.8 teraflops.

IBM said it plans to ship the new blade servers Oct. 26, offering them to companies that run highly visual 3-D applications, or those running high-powered algorithms for real-time analytics, such as those used on Wall Street. IBM offers the Linux operating system on top of Cell to provide a more flexible development environment than competing products that use a semiconductor device called a field-programmable gate array, or FPGA. Such devices are programmed once for specific functions.

"We are really trying to take high-performance computing workloads from the ivory tower to the masses," Richards said.

In doing that, IBM also introduced on Wednesday the Software Development Kit for Multicore Acceleration Version 3.0. The SDK, which can be plugged into the open source Eclipse integrated development environment, includes libraries that provide system services, debugging tools, and performance analysis and tuning tools. It also includes sample code, demonstration programs, and compliance tests. The SDK, scheduled for release Oct. 19, also include early release features, such as automatic partitioning of an application across a hybrid computing platform consisting of an x86-based system and one or more QS21s.

In addition, IBM plans to announce next month tools for interoperability between Windows desktops and multiple Cell-based blades running Linux on the back end, Richards said.

Editor's Choice
Brandon Taylor, Digital Editorial Program Manager
Jessica Davis, Senior Editor
John Abel, Technical Director, Google Cloud
Cynthia Harvey, Freelance Journalist, InformationWeek
Christopher Gilchrist, Principal Analyst, Forrester
Cynthia Harvey, Freelance Journalist, InformationWeek