Swiss boffins go nuts for Cray supers
Cuckoo clocks, chocolate, and XMT-2 ops
Supercomputer maker Cray hasn't even finished building the next-generation of is massively multithreaded XMT supercomputer, and the Swiss National Supercomputing Centre (CSCS) in Manno, Switzerland already knows it wants one.
The XMT-2 machine, as it is known outside of Cray at the moment, is expected to be launched sometime this year. The XMT design is the result of the work of Tera Computer founder Burton Smith, who now is a technical fellow at Microsoft Research, and James Rottsolk, who used to run Cray after Tera bought the name-brand super maker in 2000. The MTA and then follow-on XMT designs are not heavy on the number-crunching or cache memory, but are instead packed with lots of threads, deep instruction pipelines, and fast switching between threads to keep the machine humming. The current "ThreadStorm" processor used in the XMT machines has 128 threads and the XMT system can scale up to a mind-bobbling 8,192 threads with 128 TB of shared memory across all those threads.
It is exactly the kind of machine you want to have if you want to compare patterns on absolutely immense datasets very quickly. That's why the U.S. government likes the XMT and is probably going to love the XMT-2 even more when it ships out of Cray's Customer Engineering division later this year.
Cray has not said what the feeds and speeds are for the XMT-2 box, but it seems likely, as El Reg has said before, that the ThreadStorm-2 processor will sport more threads and fit into a G34 processor socket used by the current Opteron 6100 processors from Advanced Micro Devices. The original MTA and MTA-2 processors created by Tera had their own sockets, but with the "Eldorado" boxes that eventually became the XMT machines,
Cray got smart and made the ThreadStorm processor plug into the 1,207-pin Rev F socket used by Opteron 8000 series chips. The MTA-2 to ThreadStorm jump saw clock speeds bump up from 220 MHz to 500 MHz and processor count jump from 256 CPUs to 8,192 CPUs using the XT4 chassis and its SeaStar2 interconnect. The "Gemini" XE6 interconnect used in the current XE6 machines can scale to millions of processor cores, and Cray could no doubt ramp up the clock speeds on the threads, add more threads, and boost processor count to some unimaginable number.
Whatever Cray is going to do, CSCS is very happy to be the first customer to sign up to buy one. Just as it was the first HPC shop to get its hands on an XE6 super back in July 2010. CSCS also has a 141.6 teraflops XT5 system called "Monte Rosa" that it installed in June 2009. The XE6 machine is toy with only two cabinets with 352 of AMD's twelve-core Opteron 6100s. The XT5 machine has 3,688 of AMD's six-core Opteron 8400 processors, for a total of 22,128 cores. CSCS also has a few smaller XT4 systems.
"Many researchers are faced with massive volumes of data through experiments, observations, and simulations on a vast array of scientific applications such as material sciences, medicine genomics, high-energy physics, climate research, and astrophysics," explained Dominik Ulmer, general manager at CSCS, in a statement.
"The next-generation Cray XMT will enable our scientists to perform data analysis applications that differ significantly from the current high performance computing workloads in that the data structures are often irregular - based on strings, trees, graphs and networks - without the high degree of spatial and temporal locality seen in physics-based simulations using regular matrices."
Like I said, these XMT thingamabobs suck at math and do great at sifting through huge volumes of data very quickly and doing pattern matching, predicting behavior, or running scenarios. ®