(“World fastest GPU accelerator” as per 2014 — http://images.nvidia.com/content/tesla/pdf/nvidia-tesla-k80-overview.pdf )
10 New Zeus HPC Broadwell-CPU based nodes have 10 x 2 Nvidia Tesla K80 GPU accelerators on board (2 per node). Theoretical performance of each K80 GPU in double precision is 2.91 TFlops (8.7 TFlop in single precision) , which gives theoretical GPU power of these nodes in order of 58 TFlop in double precision calculations.
18 Older Sandybridge CPU-based compute nodes of Zeus have 36 Nvidia Tesla K20 GPU accelerators (2 per node), each of K20 has max. theoretical performance of 1.2 TFlops (double precision), which gives overall max GPU compute power of 43 TFlop. Obviously this is just an indication of the amount of max. possible compute power at ideal scaling (realistically you can’t just add these numbers together).
In comparison HPL benchmark performed on CPUs of all new 56 Broadwell nodes (1792 CPU-cores) gave 34 Tflop and 1152 CPUs of older 8-Core Nehalem based nodes gave 9.5 TFlops. and 18 x 12-core Sandybridge CPUs produced about 3.6 TFlops. See details here: http://zeus.coventry.ac.uk/wordpress/?s=HPL
——
Alex
Recent Comments