US supercomputers are having to grow up, with a wider set of tasks the machines at Lawrence Livermore and Oak Ridge National Laboratories are being asked to do by the researchers, national security teams, and others given access to them demanding a change in architecture in order to keep them flexible. IBM and NVIDIA are upgrading two supercomputers – Sierra at Lawrence Livermore, and Summit at Oak Ridge – using IBM’s support for the open-source OpenPOWER standard, increasing the interconnect speed of the CPU and GPU processors responsible for doing all the heavy-duty crunching. In fact, so the companies claim, there’ll be as much as a tenfold cut in processing time for real-world applications.
At the heart of it is the new NVIDIA NVLink interconnect technology, which increases the speed between processors by five to 12 times what they currently have. NVLink is being integrated into the IBM POWER CPUs along with NVIDIA’s own next-gen Volta GPUs.
Finally, there’s a new interconnect tech developed by IBM and Mellanox, which is said to use “built-in intelligence” that improves how data is handled.
While supercomputing is often seen as an excuse to wave big petaflop figures around and get a high spot on the “world’s most powerful” shortlist, IBM is claiming a real-world improvement of 5-10x when it comes to actual commercial and high-performance applications.
At the same time, the two systems will be around five times more energy efficient.
Post-upgrade, Sierra and Summit will muster more than 100 petaflops at peak, and each have in excess of five petabytes of dynamic and flash memory. In fact, they’ll be able to move data around their respective systems at 17 petabytes per second.
IBM points out that’s the equivalent of moving over 100 billion photos on Facebook, in just a second, though it’s unlikely to be one of the typical tasks either supercomputer is expected to do.
Instead, rather than just the “hardcore scientific problem solving” which so far they’ve been relied upon for, implementing techniques such as modeling and simulation, the two systems will be able to turn their abilities to broader schemes.
For instance, tasks related to healthcare, genomics, economics, financial systems, social behavior, and visualization of large and complex data sets are likely to be on the agenda.
It’s not just a case of switching out processors, however. IBM expects the OpenPOWER upgrades to be completed in around 2017-2018.