ABCI 3.0 is a 766-node, H200-based HPE supercomputer sited at AIST.

It debuted on the Top500 list at #15 at ISC25 with using 416 nodes (54% of the nodes), resulting in 43.6 TF/H200. The run took 33 minutes.

System overview

From ABCI 3.0: Evolution of the leading AI infrastructure in Japan:

Deployed in 2024.

  • 766 Cray XD670 nodes
    • 2x Intel Sapphire Rapids
    • 8x H200 GPUs
    • 8x NDR200 InfiniBand
    • 2048 GiB DDR5-5600
    • 2x 7.68 NVMe (U.3)
    • 1x HDR InfiniBand (for storage)
  • Three-level non-blocking fat tree
  • 75 PB of Lustre and S3 storage based on QLC SSDs