By Timothy Prickett Morgan
Executives at SGI, by many measures still the market leader in high performance computing, are a bit hot under the collar about IBM’s Blue Gene supercomputer project. You’d expect this, given the fact that SGI is coming under increasing pressure from IBM in the supercomputing market – which SGI’s Cray unit once controlled the way that IBM still does control the mainframe market. But SGI’s vice president and chief scientist, John Mashey, has some valid criticisms and observations about the Blue Gene project.
High on Mashey’s list of bones of contention is the fact that the press and analyst community seems to be a little unclear on whether Blue Gene is a product IBM will eventually sell, like RS/6000 SP supercomputers, or a true research project, like the GF11 supercomputer that predates the SPs that ran for close to ten years in the TJ Watson Research Center in Yorktown Heights, New York.
The GF11 was a very specific kind of supercomputer designed to solve very specific quantum physics problems; it was designed by Monty Denneau, who is also the chief architect of Blue Gene and one of the main designers of the SPs. It was never a commercial product, and SGI wants to remind everyone that Blue Gene is more like the GF11 than a traditional SP. Monty Denneau, it turns out, told ComputerWire he has had more than a few phone calls from pharmaceutical companies and other research institutions who are telling him they don’t want SPs, but rather a piece of a Blue Gene machine.
He has obviously had to remind them that Blue Gene is not, as yet, for sale, and that it may never be depending on how the research goes. That said, whatever IBM learns from the Blue Gene product will eventually make its way into the SP line three, four or five years from now. Denneau says it will probably not take five years to build Blue Gene, although IBM Research has told everybody the project has a five-year development span so it has some slack. Three and a half years looks more likely, he says.
SGI’s Mashey also says that IBM is not the only one targeting the 1 petaflops level, but that every serious player in HPC is shooting for that performance level, too. He says that IBM’s endeavor is not unique, but that IBM’s proposal is nonetheless lofty. That said, Mashey is somewhat puzzled by the Blue Gene design. What IBM is proposing is quite bazaar: 1 petaflop on 1 million processors by 2005 with a $100mn budget. This would call for getting past silicon as we know it today, he says. Also, the typical progression of technology is to do more with less. Given the state of technology today, one would expect a proposal for far less than 1M processors, he says. And the prospect of managing one million processors confirms that this could only be acceptable for a very specific single-purpose application.
Ironically, Mashey says the Blue Gene project strongly validates the ccNUMA architecture that SGI is currently delivering with its Origin and custom high-end machines based on its design like ASCI Blue Mountain. These systems, designed to address customer growth, performance and manageability requirements, are delivering phenomenal levels of performance today and are well on the way to petaflops of power with well under one million processors. He cautions prospective HPC customers that it would be difficult to predict what Blue Gene will and will not do. From past experience, large computers with massive numbers of CPUs have often been announced, less often delivered, and rarely become commercial reality. To make his point, Mashey reminded us of the old joke: What can you build when you put six MIT graduates in the lab? Anything. But you only get one of them.
Mashey is also a bit perplexed by the relatively small amount of memory in the Blue Gene machine. He says that current generations of supercomputers tend to require between 512Mb to 1Gb per processor, depending on the type of supercomputer workload the box is running. Chemistry problems like the one Blue Gene will be solving tend to have big number crunching requirements and relatively modest memory needs, while fluid mechanics, finite element analysis and radiation simulations tend to be memory hogs. The point is, says Mashey, that commercial customers need to support a wide range of CPU-to-memory ratios. IBM’s Denneau has hinted that it would be able to tweak the Blue Gene chip design to accommodate more memory and fewer processors to support different workloads (CI 3812), but even still, IBM is talking about using only a few megabytes of memory per CPU at a maximum even if it does this.
Blue Gene represents an ultra-minimalist design approach and a radical departure from current supercomputer designs. The question remains whether or not IBM will use this approach on a smaller scale to go after SGI, Sun Microsystems Inc and Hewlett- Packard Co in the commercial HPC market, or stick to its SP nodes and high-speed interconnect. Given what IBM thinks it can do with Blue Gene, perhaps the forthcoming Power4 and Power5 chips are overkill and IBM should really consider a streamlined AIX or Linux supercomputer that looks more like Blue Gene and less like Deep Blue.
