First of all, why should we care? The big capability machines are designed not only to do science but also to address a set of problems termed Grand Challenge Problems. A Grand Challenge Problem is simplistically defined as… one that cannot be solved in a reasonable amount of time with today’s computers. – Wikipedia. So there’s a race to build machines that can solve these problems that, by their very definition, have significant economic or social impact. Some examples of these problems include or have included: climate modeling, human genome mapping, semiconductor modeling, vision and cognition to name a few. Some of the Grand Challenge problems will, at some point in the future, require machines with performance on the order of 1 ZETTAFLOP (10^21 FLOPS, that’s equal to one BILLION ASCI Red machines!) Again, taking the compound aggregate growth rate of the large machines and extend that trend into the future, we wouldn’t expect to see such a machine until ~ 2029. Just in time for my 70th birthday.
High Performance Computing (HPC) has a broad definition. Most of the time when HPC is mentioned, many think about the large capability machines. To find out about the highest performing machines in the world today, all one needs to do is go to the ‘Top 500 Supercomputer sites’ website at: http://top500.org/lists. The machines are rated by their LINPACK benchmark number. As one would expect, number one on the list is the machine that achieved the maximum performance. One note, the first TFLOP machine (ASCI Red), introduced in 1997, was based on the Intel PentiumPro processor. TFLOP was the top rated machine for an as yet unmatched seven consecutive releases of the list .It remained on the Top 500 list until 2005, eight years after being introduced as number one. Since November 2006, the top rated machine has had a performance of ~280 Trillion Floating Point Operations per Second (10^12 Floating Point Operations per Second – TFLOPS) and consists of 131,002 processors. We expect that the first Peta FLOP machine (10^15 Floating Point Operations per second) will arrive in the 2008/2009 timeframe, assuming that the rate of compound aggregate growth in computing remains consistent with the trend from 1994 to today. There are some common attributes to these systems. First they require lots of research $$ to allow companies to invest in new technologies…technologies that may be too risky to immediately integrate into high volume products. Government programs help with a significant portion of the R&D, which can make building such a system feasible for a company or group of companies. The business isn’t exactly the most lucrative. There are a few players in this segment. Some do it because building these machines is their business model. Still others do it for the same reason that car companies have racing teams…develop new technologies, test and perfect them in the high performance, high stress environments and, eventually, some of these technologies will make it into the mainstream product(s). Some technologies do make it to mainstream product. One example is game physics algorithms/logic that’s used in high end gaming and game console systems. Also noteworthy are the pipelined vector processing units introduced more than three decades ago by Seymour Cray: today they power the graphics units used in personal computers. Another example is the Unix operating system which began its life in high-end servers and lives on today not just in those servers but even as the core of Apple’s OS-X operating system. Greater computing demand requires higher processing capability (1-10 TFLOP per processor), improved power efficiency (today’s high end systems consume on the order of 10’s of MegaWatts.), much higher memory capacity and bandwidth requirements that make today’s memory subsystems pale in comparison, greater IO bandwidth, very fast and efficient light-weight kernel OS’s, and applications that can scale up to millions of threads. Pretty challenging but the benefits can be huge. Not all systems on the TOP 500 list are big, proprietary, capability systems. Some of them use standard, common-off-the-shelf (COTS) components. Have a look at them when you get the chance. They will be the subject of the next blog.↧
