IBM kills off second-gen Cell blade server
Power chips will do the math
IBM has killed off its QS21 two-socket Cell blade server, the second generation of Cell blades sold by IBM, which were announced in August 2007.
The move comes several weeks after IBM murdered a future Cell-based blade server, the QSZ2, that would have given Nvidia's Tesla and Fermi as well as Advanced Micro Devices' FireStream graphics co-processors a run for at least some of the money.
Today, IBM told customers using its QS21 blades, mostly in as supercomputer clusters, that they have until June 25 to order more parts. After that, no mas. The QS21 blade came with two 3.2 GHz Cell processors, each with one Power core and eight synergistic processing elements (SPEs) for doing complex calculations and 2 GB of XDR main memory; the blade sold for $9,995 and a chassis of 14 of the blades was able to deliver 6.4 teraflops of single-precision floating point math performance.
The QS22 was announced in May 2008, offering dual 3.2 GHz PowerX8i processors rated at 230 gigaflops single precision and 109 gigaflops double precision of floating point power each. This QS22 blade, which is still available, had about five times the double precision performance and InfiniBand mezzanine cards. This QS22 blade also sported 32 GB of DDR2 main memory, and it has quad data rate InfiniBand as well as Gigabit Ethernet connectivity. With 8 GB of memory, the QS22 costs the same $9,995; $13,539 will get you a blade with 32 GB.
The clock is ticking for the QS22 blade, of course. And not just because everyone knows that IBM is not going to invest in creating future blade servers based on Cell chips, as it admitted last fall. (The company had actually spiked the QSZ2 blade, which was to deliver around 500 gigaflops of double precision math performance per Cell chip. (The future Cell chip that IBM killed off appeared to have two Power cores and 32 SPEs each, delivering almost five times the double-precision math speed of the PowerXCell 8i chip).
The reason why the QS22's days are numbered is simple. IBM, say sources familiar with the company's plans, is to add specialty processing capabilities like those embodied in the SPEs in the Cell chip to the future Power chips beyond the current Power7 generation. Perhaps starting with Power7+ and definitely in full bloom with the Power8 generation.
This will take several years to come to fruition. And by then, IBM may decide that Nvidia is its new best friend and head off in a different direction entirely. That's how it goes in the HPC business. You can afford to be fickle when you are on the bleeding edge, and so can the customers who are buying the machines - largely with tax money levied by governments for HPC labs that exist for political as well as economic and technical reasons.
IBM has not announced a withdrawal date for the QS22 blade server based on the PowerXCell 8i chips, but it will probably sell them as long as it has customers that want them. Los Alamos National Lab has already upgraded the "Roadrunner" hybrid Opteron-Cell to the QS22 blades, so the big customer is already done and hitting the ceiling.
The Roadrunner blade cluster pairs a Cell chip with each Opteron socket in the machine, allowing the Opteron to offload calculations to the Cells and helping it hit that 1 petaflops performance level. Had the QSZ2 come to market, Roadrunner could, in theory, have been upgraded to 5 petaflops or more.
But clearly the Department of Energy has other ideas about how it is going to get to tens of petaflops of performance, or else IBM would have already delivered beefier Cell chips. The big government-sponsored super labs in the States will no doubt be watching how the Power7-based "Blue Waters" 1 petaflops super pans out at the University of Illinois when it is delivered later this year and will no doubt be pestering IBM about its plans for Power7+ and Power8 machines. ®