IBM Delivers Promised Linux-Based Cell Blade Server

Published: September 19, 2006

by Timothy Prickett Morgan

As expected, server maker IBM has launched a blade server for its BladeCenter chassis that employs the Cell variant of the PowerPC processor. The blade, called the QS20, puts two of the Cell chips on a single board and into a blade enclose that eats two slots in the BladeCenter chassis. Because of the eight special processing units in the Cell chip, the QS20 blade has an immense amount of computing capacity that can be brought to bear for media processing or other kinds of number crunching.

Big Blue first demonstrated the Cell-based blade server back in February at a server event in New York. At that time, I saw a single Cell processor render the 3D graphics of Mount Rainier in Washington in real time from satellite images. You could fly through it, and it didn't flicker. Which is why Sony is using the Cell chip in the future PlayStation 3 game console. The Cell comes with a single 64-bit PowerPC core (basically a PowerPC 970 core, including the AltiVec vector unit) plus eight special co-processors, which IBM calls synergistic processor units. The chip also has 64 KB of L1 cache and 512 KB of L2 cache memory and an integrated Rambus XDR main memory controller.

IBM says that the QS20 is using two Cell chips running at 3.2 GHz, and that the blade server can deliver 410 gigaflops of peak performance. The QS20 has 512 MB of XDRAM memory per processor, which is not exactly a lot, and a 40 GB IDE disk drive mounted to the blade. It also has two Gigabit Ethernet ports to link the QS20 to the chassis backplane (which is used for blade management) and to Gigabit Ethernet switches (which link the blade server to the outside world). The blade, which is not particularly skinny, also has room for two Mellanox 4x InfiniBand host channel adapters, which is what most supercomputing customers will need to build supercomputers from these blades. IBM could eventually offer 10 Gigabit Ethernet connectivity on the QS20, if enough customers ask for it. The blade has no slots beyond the ports for InfiniBand adapters.

When you do the math, a single BladeCenter H chassis with seven of the QS20 blades should pack about 2.8 teraflops of raw performance, and a full rack of these blade servers would deliver about 17 teraflops. IBM warns, however, that you cannot mix the QS20 blades with other Xeon, Opteron, or PowerPC blades within the same chassis; the company did not explain why, but it could have something to do with thermal issues. This QS20 blade is the device that makes up the bulk of the 1.6 teraflops of computing power in the Roadrunner supercomputer that IBM is building for Los Alamos Laboratory in New Mexico, a deal IBM cut with the U.S. Department of Energy two weeks ago.

While the Linux 2.6.16 kernel can support the Cell chip and its special math units, only Red Hat's Fedora Core5 development release provides something akin to a production-level Linux for the chip.

The QS20 blade will be available on September 29. The InfiniBand adapters for the QS20 will be available on October 27. A QS20 blade has a list price of $18,995. The InfiniBand adapter kit costs $1,950. That sticker price makes the QS20 one of the most expensive blade servers on the market--much more expensive even than Hewlett-Packard's BL60p Itanium-based blade server. But if you can tweak your Fortran code to make use of those extra processor cores, you can get something probably approaching 60 percent of the peak capacity of the machine and then such a blade server is a bargain.

Do the math. Los Alamos is paying $35 million for an 80 teraflops Opteron cluster from IBM, which is the first phase of the Roadrunner box. That's $438 per gigaflops. To build out Roadrunner, Los Alamos will add 8,000 QS20s, which will yield about 1.52 teraflops of computing power. For just the blades alone--not including chasses and other hardware--8,000 of these units would cost $152 million at list price. And you can bet Uncle Sam isn't paying anything close to list price for Roadrunner. But even at list price, that works out to $100 per gigaflops. My guess is that the Cell portion of the Roadrunner machine, including software and services, comes to somewhere around $80 million, which puts the price of a gigaflops in the range of $50 using the Cell blades. This is a lot less expensive than the Opteron side of Roadrunner. Which is why the advent of the Cell chip is an important thing.

Which brings me to my next point. Sony will begin shipping the PlayStation 3 on November 17. It will cost $500, and it will have one of these Cell chips, some XDRAM memory, a 20 gigabyte disk, and three Gigabit Ethernet ports. How long do you think it will be before someone hacks Linux onto this PlayStation 3 box and then builds a supercomputer cluster? At $500 a pop instead of $18,995, I know someone is going to try it.

RELATED STORIES

IBM to Build 1.6 Petaflops Super for Los Alamos Lab

PowerPC Cell Chip Gets Fedora Linux Support

IBM Announces BladeCenter Kickers