First details on the National Nuclear Security Administration (NNSA) Trinity Supercomputer show that the 42 PF/s system costing $174M USD will run a combination of Intel Haswell and Knights Landing processors. In particular the Intel Xeon Phi devices will use Micron’s Hybrid Memory Cube technology, which will greatly help memory bandwidth and memory capacity limited applications. Intel is asserting that KNL will deliver 4x the performance of an NVIDIA K40 on some applications. The Trinity supercomputer will be housed at the Los Alamos’ Strategic Computing Center (SCC) in the Nicholas C. Metropolis Center for Modeling and Simulation.
Trinity is expected to be the first platform large and fast enough to begin to accommodate finely resolved 3D calculations for full-scale, end-to-end weapons calculations. Large 3D jobs are common in combustion research, CFD (Computational Fluid Dynamics) and a variety of other application areas of interest to the Trinity supercomputer shareholders such as LANL, Sandia, and the NNSA. The presentation by Doug Doerfler – the Trinity architecture lead (“Trinity: Next-Generation Supercomputer for the ASC Program“) provides more details including (1) the Trinity supercomputer has been sized for large 3D jobs and (2) it is expected to be delivered in 2015. (A 2015 delivery date is also expected for the KNL-based NERSC Cori supercomputer).
Note that the Micron Hybrid Memory Cube technology provides 15 times more bandwidth than DDR3 DRAM and draws 70 percent less energy, with five times more bandwidth than the emerging DDR4 memory. The Knights Landing chipset will also have DDR4 controllers, but Cray could not say whether the supercomputer would use that memory type, which is understandable as HMC technology is very new making both pricing and manufacturing yields uncertain.
Federal review radio has an interview with Thuc Hoang, Trinity Project Manager, National Nuclear Security Administration on the procurement.
Following are some relevant slides: