|
How many supercomputers does it take to turn on a light bulb?As you must know by now, I love supercomputers, parallel processors, SIMD, MIMD—hell, I even like MISD. So maybe since all humans on earth know of my fascination with all things big and fast is why I have spent the last couple of weeks inundated with information about them. This is not a case of be careful what you wish for because, just like graphics, in supercomputers too much is not enough. But, my fascination and that of others doesn’t translate into revenue for anyone building such devices; not really. Yeah, I know, I’m supposed to be a big influencer, and people pay attention to what I say—ah-huh, I tell them at Starbucks and they say, “That’s nice, that’ll be $3.50.” Whereas I might influence a VC to invest in a startup, or an end user to buy a particular AIB or display, I have zero, less than zero, influence on the folks who buy supercomputers. Why is that? Because governments buy supercomputers, and they don’t necessarily, being governments, make those decisions on a rational or technical basis. Japan’s Earth Simulator, and the U.S.’s Blue Gene, are as much geo-political decisions as they are about petaflops, and maybe more so. These are expensive machines. The Earth Simulator cost $310 million, and the Blue Gene/L cost $290 million. Even governments don’t buy machines like that every year. However, if you look at the top 500 supercomputers in use (http://www.top500.org/list/2006/06/100) you should notice which of the machines are clusters, which means they are super servers tied together. It also means they are not all going to cost more than $100 million. In fact, the newest machine that is coming up in Japan, the MDGrape-3 (developed by Riken), costs only $9 million and is expected to take the title of the world’s fastest. And The Tokyo Tech supercomputer built by Sun with 10,480 AMD processors and 360 CSX600 Clearspeed coprocessors (each CSX600 contains 96 32/640bit FPP) is anticipated to be one of the ten largest supercomputers in the world at 85 TeraFLOPS When discussing these machines and the market, words are critically important. Not all parallel processors are “supercomputers.” A GPU is a parallel processor, and one of them is not a supercomputer, but it is conceivable that a nest of them could form a supercomputer. So to avoid having to define and decide what makes a computer “super” marketeers have adopted other words like “high-performance computing” or “high-performance technical computing,” carefully avoiding the issue of defining what “high” is. When that’s done then the counters—folks like us—can count whatever they choose to include. So if Ambric makes a 360 processor chip that can hit a TeraOp is that a supercomputer? Probably not, but is it an HPC? Maybe, depends on the stuff around it (OS, I/O, memory, etc.). And I think we can assume Ambric’s chip will sell for considerably less than a million dollars, just as a GPU or Cell processor does. And yet, the counters would like to count them and a whole lot of other stuff when arriving at the “market size” or “market potential” for HPC or HPTC’ing. OK then, what is the market size or TAM for GP-GPUs? Answer: any number you want it to be. You can make it big by counting everything that is needed to make a GPU act like a supercomputer, which can include of course the boxes and power supplies, the OS, the middleware like Peakstream, the peripherals, displays, the power needed to run it, the cooling systems, the licenses, maintenance, installation costs, and the coffee the truck driver had who delivered the stuff. All of which brings me to this point. It’s silly in my opinion to use IDC’s or anyone else’s multi-billion-dollar cluster server market sizing when evaluating the market potential for chips, middleware, and other componentry in the super/HPC/HPTC market. And when I’m told such numbers it doesn’t impress me, it makes me question the presenter’s grasp of reality. Does he/she think I’m stupid, not paying attention, or easily impressed? |
| |||||||||
