RealTime IT News

IBM's Cell Chip Moves Out From Gaming's Shadow

IBM's Cell Broadband Engine (Cell/B.E.) is not just for playing games any more.

The company today announced the third release of its Cell-based blade server, which jettisons some of the legacy features from chip's early role in the Sony PlayStation 3 game console, further evolving it into a complex math coprocessor.

The first blade server sporting the Cell chip -- had been the BladeCenter QS20, which share the same chip as the PlayStation.

The next generation, the QS21, modified the south bridge.

With the QS22, available later this month, IBM (NYSE: IBM) improved double-precision performance, greatly increased the available memory and ditched the Rambus XDR bus used in the PlayStation in favor of DDR memory -- a standard in the PC space.

The QS22 works in tandem with a standard server and can function as a big compute box to do high-performance computing (HPC) tasks. A dual-socket QS22 blade brings 380 gigaflops of processing power alone. For comparison, your average Intel dual core processor offers around 12 gigaflops of processing.

"We are embracing a hybrid computing model," Jim Comfort, vice president of workload optimized systems at IBM, told InternetNews.com. "Core computation stays on your platform, but when they require substantial digital media or transcoding or content creation, you move to Cell but keep the primary workload on the conventional processor."

"Working in a blade system allows us to implement a hybrid system a lot more readily," he added.

The hybrid computing model is also called heterogeneous computing, and is a concept Intel (NASDAQ: INTC), AMD (NYSE: AMD) and nVidia (NASDAQ: NVDA) have each been talking about exploring.

But IBM is out in front of all of them, said Dan Olds, a principal analyst with Gabriel Consulting Group.

"IBM is leading where others are promising to go," he said. "IBM is there because Cell has been around. The closest competitor to what I see they are doing now is nVidia, but there are some real differences. The nVidia thing has to evolve."

For starters, nVidia doesn't support error-correction memory, something needed in the important financial calculations in which the QS22 is used.

With the QS22, all calculations are native double-precision for a five-fold improvement in double-precision math. This could be a boon for financial services firms, which will be able to fire off huge mathematical computations to a QS22 server to perform the math, which would then send back the result to something like an x86 computer running Windows.

The Cell processor will be exposed as just another API, so any application can call on it to make calculations.

If users need to go closer to the metal and program directly to the processor, they can, but it's not required, Comfort said.

With the QS22, the memory limit also has been greatly increased over previous versions of the Cell -- from 2GB to 32GB per blade. Jettisoning Rambus XDR memory in favor of standard DDR helped address a bottleneck, Comfort said, adding that XDR had a lot of latency in it and didn't support large amounts of RAM the way DDR does.

The BladeCenter QS22 will fit in IBM's new BladeCenter HT chassis, and sports some new management features as well as design improvements. They include features that help system managers track and manage power requirements.

Olds said the BladeCenter QS22 makes HPC technology affordable to the masses.

"I'm looking at this for enterprise computing, and it just does a hell of a lot better on numerically-intensive garden variety stuff than anything you'll get out of Intel," he said.

"It's not like people will be putting these things together in massive configurations, but it will allow smaller customers to put together something to meet their needs," Olds added. "They will see it as a way to run the HPC-like workloads they've got and they can buy it in small, granular pieces."

The BladeCenter QS22 is available at the end of the month at a starting price of $9,995.