EECS Servers

Spring 2019

CPUsGPUs
program using stl::thread, et al., for example program using CUDA or OpenCL, for example
cycle1
  • 12 processors, each with 6 NUMA cores
  • ~124 GB shared RAM
  1. Quadro 6000
    14 multiprocessors (32 cores each) ⇒ 448 cores
    ~6.5 GB global memory
  2. Tesla C2075
    14 multiprocessors (32 cores each) ⇒ 448 cores
    ~5.6 GB global memory
cycle2
  • 24 processors, each with 6 NUMA cores
  • ~124 GB shared RAM
  1. Quadro 6000
    14 multiprocessors (32 cores each) ⇒ 448 cores
    ~6.5 GB global memory
  2. Tesla C2075
    14 multiprocessors (32 cores each) ⇒ 448 cores
    ~5.6 GB global memory
Details
cycle3
  • 12 processors, each with 6 NUMA cores
  • ~124 GB shared RAM
  1. Quadro 6000
    14 multiprocessors (32 cores each) ⇒ 448 cores
    ~6.5 GB global memory
  2. Tesla C2075
    14 multiprocessors (32 cores each) ⇒ 448 cores
    ~5.6 GB global memory
Details

"Multicore" computations – i.e., computations that are distributed across a networked collection of machines (e.g., across all three cycle servers) – can be programmed using OpenMPI.