IBM announces PowerXCell 8i, QS22 blade server

Tuesday 13th May 2008, 11:23:00 PM, written by Carl Bender

IBM today made formal the announcement of an HPC-targeted variant of the Cell B.E. processor, given the name PowerXCell 8i. Featuring re-engineered SPEs with enhanced dual-precision performance and improved memory addressability, the processor is capable of ~102 DP GFlops and support for up to 16GB of DDR2 memory at ~25GB/s. Originally described at the Cool Chips conference in April of 2007, the PowerXCell 8i is the processor IBM will use as the workhorse in a heterogeneous setup powering the Los Alamos National Laboratory's Roadrunner supercomputer, expected to be the world's fastest when it debuts later this year. Having undergone cost/benefit trials since the original project announcement back in September 2006, by late last year Los Alamos scientists had succeeded in improving key algorithm performance (molecular dynamics, radiation transport, plasma behavior) by a six-fold average through the use of the Opteron/Cell compute nodes vs Opteron-based nodes alone.

Following its successful test period with the Department of Energy, the PowerXCell 8i will become available commercially next month through IBM's newly announced BladeCenter QS22, a blade-based solution rated at 217 GFlops of DP performance in a single-slot form factor. With over fifty customers presently migrating code to the QS22, IBM sees an opportunity in a market niche it estimates is worth roughly $8-10 billion. The PowerXCell 8i will be aimed at seven industries where the architecture is particularly well suited for HPC performance gains, including: Aerospace/defense, Health care/life sciences, Petroleum exploration, Financial, Digital media, Electronics, and Government. In addition to the architectural refinements of the new Cell, IBM has continued work on the software and tools, allowing the QS22 to function in an increasingly transparent fashion inside of heterogeneous BladeCenter environments. Through the development of new API/library functionality, customers are able to take advantage of QS22-accelerated code vectorization on low-hanging fruit (such as Monte Carlo) without the need to make significant alterations to the original x86 or POWER code.

Developed as an offshoot of the Cell partnership with Sony and Toshiba, IBM has indicated that in the future, high-volume/performance-intensive applications such as game systems are likely to continue as the point of origination for new high-performance architectures. It can be expected that to some extent the eventual STI Cell '2' will reflect IBM's design gambit in a post-Larrabee/GPGPU computing world, the massively parallel space in which Cell v1.0 itself was an early player.

Discuss on the forums

±

AMD Bulldozer microarchitecture analysis
Really Techy Worldly Hot Par 2010 coverage
NVIDIA's Parallel Nsight finally released
NVIDIA GeForce GTX 460 - GF104 breaks cover
PhysX87, ancient tragedy in 5 acts by RWT
Carmack Rage: either console could look superior
Beyond Programmable Shading CS448s first slides available
ATI releases OpenGL4.0 preview driver, for great justice
32nm sixsome over at RealWorldTech
Khronos release OpenGL 3.3 and 4.0