By Calder Phillips-Grafflin
If you haven’t heard already, we’re getting a supercomputer. Yes, thanks to IBM Senior Vice President John Kelly’s generous contributions, Union is about to become a proud owner of an IBM Intelligent Cluster. While this isn’t exactly a surprise, this past weekend’s announcement is the first official announcement about a process that has taken almost two years. In addition, the announcement contains the first official description of the capabilities of the new cluster, and those capabilities are pretty impressive.
Supercomputers, also known as “high-performance computing” or HPC, are almost unheard of at liberal arts colleges, mostly because of their cost and complexity. That doesn’t mean they aren’t useful; a wide range of data analysis and scientific problems can be addressed more easily using high-performance computing. Everything from modeling in economics to the analysis of classic Greek texts can benefit from the parallel processing power and large memory capacity of a cluster, let alone the multitude of applications in computational chemistry, neuroscience, biology and engineering.
The particular cluster we’re getting from IBM is an Intelligent Cluster. According to the announcement, we will be getting a system with 88 nodes and 1056 processor cores. Comparing this to other cluster configurations, it appears that we are getting a custom configuration, since the example setups provided by IBM only contain 42 or 84 nodes. In addition to the 176 Intel Xeon processors, the cluster can also support NVIDIA Tesla computational graphics cards. Depending on configuration, adding these cards can increase performance by a factor of 8.
Flops are the standard measure for comparing computers; they stand for “floating point operations per second”, and modern high-performance computers have performance ranging over a quadrillion flops. Most, however, including the computers which make up the TOP500 list of supercomputers; have performance in the trillions of flops, which is referred to as a teraflop.
25 teraflops isn’t quite high enough to make it onto the TOP500 list, but that isn’t the only configuration possible. If, instead, our system has 2 NVIDIA tesla cards per node, the performance could be over 200 single-precision and over 100 double-precision Teraflops.
Performance at that level would place our cluster at about 80 on the TOP500, which would be extremely impressive. A more conservative estimate with teslas in only half the nodes would still produce about 115 single-precision and 58 double-precision teraflops, which would rank it at about 140.
How does this compare to computers at other colleges and universities? American universities have heavily adopted high-performance computing, and more than 24 of the TOP500 computers either belong directly to a university or are used almost exclusively by one. The graphic below shows a number of these institutions. You’ll notice that all of the schools are much larger than Union, and this makes our new cluster even more impressive. Not only are we one of the smallest institutions to have a supercomputer, we are the only one where the computer is easily accessible to both liberal arts and sciences disciplines.
The new cluster will be installed this summer, and personnel from ITS will be trained to run and maintain it. Once it’s ready, it will be a powerful tool for research and a unique capability for Union.
Beyond that, if we’re lucky with the configuration, we’ll have an internationally-ranked supercomputer right here on campus. And yes, we might just have a better computer than RPI.