System C kugui_Hardware

HOME > Supercomputers at the ISSP > System C kugui > System C kugui_Hardware

Hardware Configuration

HPE Apollo 2000 Gen10 Plus/HPE Apollo 6500 Gen10 Plus

The system consists of a standard configuration CPU node equipped with AMD Milan featuring 128 cores per node,
and an ACC node equipped with 4 NVIDIA A100 GPUs per node.
It has a total theoretical computing performance of 0.973 PFLOPS and a total main memory capacity of 35.25 TB.

 

Hardware
  – CPU node x 128 nodes

    – CPU: AMD EPYC 7763, 2.45GHz, (64core)×2
    – Main memory capacity: DDR4-3200, 16 GB×16 = 256 GB/node
    – Theoretical performance per node
      – Theoretical calculation performance: 5.0176 TFLOPS (2.45 GHz x 64 cores x 16 instructions x 2 sockets)
      – Memory transfer performance: 409.6 GByte/sec 3,200 MHz x 16 channels x 8 Bytes)
      – Cores per node: 128 cores/node

      – Memory capacity per core: 2 GB/core
    – Total theoretical performance: 642.25 TFLOPS
    – Total capacity of main memory: 32 TB

 

  – ACC node x 8 nodes

    – CPU: AMD EPYC 7763, 2.45GHz, (64core)×1

    – GPU: NVIDIA A100 40GB for HGX ×4
                https://www.nvidia.com/ja-jp/data-center/hgx/

    – Main memory capacity: DDR4-3200, 32GB×32 = 256GB/node
    – Theoretical performance per node
      – Theoretical computing performance
        – CPU units: 2.5088 TFLOPS (2.45GHz x 64cores x 16 instructions x 1socket)
        – GPU unit: 38.8 TFLOPS (9.7 TFLOPS x 4)
      – Number of cores per node
        – CPU unit: 64cores/node
      – Memory capacity per node
        – CPU unit: 256GB/node

        – GPU unit: 160GB/node
    – Total theoretical performance: 330.4704TFLOPS
    – Total main memory capacity
        – CPU unit: 2TB
        – GPU unit: 1.25TB

           
  – Node-to-node communication
    – Topology: FAT-tree
    – Node interconnect: HDR200 InfiniBand
           
  – File system
    – Lustre file system
    – /home 200TB, /work 950TB