Aurora (ALCF)

Aurora is the Cray EX supercomputer operated by the Argonne Leadership Computing Facility (ALCF). It achieved a max HPL score of 1.012 EFLOPS using 87% of the system in June 2024¹ after debuting six months earlier with 585.3 PFLOPS.

Although this system uses significant Cray EX technology including the cabinet and Slingshot network, Intel was the prime contractor and provided its own compute blades.

System overview

Aurora has:²

• 10,624 compute nodes
• 21,248 Intel Sapphire Rapids HBM CPUs (Intel Xeon CPU Max)
• 63,744 Ponte Vecchio GPUs (Intel Data Center GPU Max)
• 166 cabinets

Node architecture

Although Aurora uses the Cray EX platform, the compute blades were manufactured by an Intel subcontractor, not Cray. Thus, they don’t have a Cray model name, and you cannot buy PVC blades from Cray.

Each Intel node has:³

• 2x Intel Xeon CPU Max 9470C (Sapphire Rapids HBM)
  • 52 cores per socket, 2 threads per core
  • 64 GB HBM per socket
  • 512 GB DDR5
• 6x Intel Data Center GPU Max 1550 (Ponte Vecchio)
  • 128 Xe cores per socket
  • 128 GB HBM2e per socket
• 8x Slingshot-11 NICs

These nodes are so large that there is only one node per blade, or 64 nodes per cabinet.

Network architecture

Aurora uses Slingshot-11 in a dragonfly topology with compute and storage integrated on the same fabric. There are:⁴

• 166x compute groups
• 8x storage groups containing DAOS servers
• 1x service group containing login nodes and other ancillary servers

The inter-group connectivity is:

• 2 links per compute-compute group
• 2 links per compute-service group
• 2 links per compute-storage group
• 8 links per storage-service group
• 24 links per storage-storage group

Each group has 32 switches, and each switch has:

• 16 ports down to endpoints (L0)
• 31 ports to other switches in the rack (L1)
• Up to 16 ports to other groups (L2)

The blocking factor within the compute subsystem is 1.55:

Blocking factor $=\frac{512\text{ endpoints}}{165\text{ other groups}\times 2\text{ global links}}=1.552$

Storage subsystem

Aurora uses DAOS instead of a traditional parallel file system. This storage system is composed of 1,024 Intel Coyote Pass servers, each containing:⁴

• 2x Intel 5320 CPUs (Ice Lake)
• 512 GiB DDR4 DRAM (16x 32GB DIMMs)
• 8192 GiB Intel Optane 200 (16x 512 GB DIMMs)
• 244.8 TB Samsung PM1733 SSDs (16x 15.3 TB)
• 2x Slingshot-11 NICs

It has a total of 230 PB of capacity using 16+2 erasure coding and over 31 TB/s of bandwidth.⁵

Footnotes

1. ISC’24 recap (glennklockwood.com) ↩

2. Aurora_FactSheet_2024.pdf (anl.gov) ↩

3. Hardware Overview - ALCF User Guides (anl.gov) ↩

4. DUG22 - DAOS Community - Confluence (atlassian.net) ↩ ↩²

5. Aurora | Argonne Leadership Computing Facility (anl.gov) ↩