Urbana (IL) - The National Center for Supercomputing Applications (NCSA) at the University of Illinois in Urbana will soon receive another supercomputer upgrade.  Complementing their existing highest-end machine called "Abe," the new machine will be "Lincoln" and will provide roughly two thirds the same capacity.  However, as technology continues to improve the components inside of Lincoln will be notably different than those of Abe, costing less to operate, consuming less floor space in the center and ultimately providing more bang for the buck.

In late June, 2007, TG Daily was honored and privileged to receive an all-access pass to the NCSA's facilities on campus for a guided tour hosted by four very spontaneous, knowledgeable and perky people.  In fact, for this journalist it's the most memorable interview I've ever had to date.  That team gave the world a bird's eye view of what exactly goes on inside a supercomputing facility, and we are all the better because of their willingness to share their love and their passion.

During that visit, the latest edition to their supercomputing family was just getting ready to be launched.  Abe, a sustainable 61 Teraflop (90 peak) cluster comprised of 1200 Dell PowerEdge 1955s blades, housing 2400 physical processors with 9600 cores with 1 GB of memory per core, was launched to operation on July 1, 2007 and debuted at #8 in the world in terms of performance.  And what a difference 15 months can make in the world of supercomputing as Abe has fallen to #23 in the world, and the specs of Lincoln are quite different than Abe's innards.  Even though Lincoln will only be about 65% as capable as Abe in terms of raw processing power, it will be comprised of far less equipment.

Lincoln will consist of only 192 Dell PowerEdge 1950 III dual-socket blades, each with quad-core Harpertown 2.33 GHz processors and 16 GB of memory.  In addition there will be 96 Nvidia Tesla S1070 accelerator units in the cluster.  Each Tesla unit itself is capable of providing up to 500 GFlops of double-precision (64-bit) and 16 GB of memory.  Nvidia donated the hardware and works with team members to make its CUDA software development kits adaptable for use in a supercomputing model.

The use of Nvidia's multiple GPU-accelerated clusters at NCSA, one of the largest of its kind in the world, is providing researchers with previously untapped potential in applications as a single machine can now harness vast computing resources that previously had to be distributed over several machines.  The GPU-accelerated clusters may ultimately prove to be the true ingredient in creating the lowest-powered, highest-performing, lowest cost general purpose supercomputers in the world, the kind that could ultimately fit inside an average person's living room with little more than a fan blowing on it to keep everything cool.  Of course, that's a few years off still, but it does appear to be coming.

Lincoln will go live next month in October and will bring the NCSA's total theoretical computing capacity up to 210 TFlops, with a more realistic sustainable throughput of nearly 170 TFlops realized.  In September, 2007, the NCSA was chosen as the site to receive the upcoming 1-2 petaflop machine called Blue Waters, scheduled to be installed over the next few years and become operational in 2011.

As of June, 2008, only one of NCSA's existing supercomputers was still in the top 100, Abe maintained its presence at #23.  Some other changes from July, 2007 were:  T3 had fallen from #47 to #134, Tungsten had fallen from #90 to #390, Mercury had fallen from #160 to completely off the list, as did Cobalt from #246 to completely off the list.  The slowest computer in the Top 500 list from June 2008 was crunching data at 9.00 TFlops.  Such a machine would've been just six slots away from the top 100 in June 2007.  And in June 2006 it would've been #47 in the world.  It's worth noting also that many private companies own, house and maintain their own supercomputing clusters that are not part of the top 500 list.

It is very likely that there are additional supercomputers in existence which the general public knows nothing about which are at least as powerful as the top 100 computers.  Nvidia is one such company who does all of their research and development via simulation on a huge supercomputing cluster of undisclosed size.

GPGPUs have changed everything

Sep 11, 2008 15:11     

With ATI and Nvidia, we have Teraflop performance in our PCs. Two 4870x2 in Crossfire produce around 4 teraflops of performance. A year ago, that would have been mid-pack in the Top 500 list. As more supercomputers get GPU accelerators like Lincoln, it won't be long before 100 teraflops or more is required to be on the list, and multi petaflop machines dominate the top 100. Truly amazing times.
EEngineer report abuse vote down vote up Votes: +0