Vanguard Astra - Petascale ARM Platform for U.S. DOE/ASC Supercomputing
summary
Astra is a 2.3 petaflop ARM prototype at Sandia with about 2,600 nodes (dual‑socket Cavium ThunderX2, 28 cores @2.0 GHz, eight DDR4 channels per socket) connected via Mellanox EDR ConnectX‑5; the system is housed in a new data center using a thermosyphon and consumes ~1 MW. Software work centers on the Advanced Tri‑Lab Software Environment, OpenHPC builds, Singularity containers and Lustre client tests. Early runs hit 1.5 PF at ~80% efficiency and strong memory‑bandwidth performance (HPCG ∼36), proving the architecture viable for DOE ASC workloads.
AI-written summary · about
chapters (12)
- 0:00Introduction to Vanguard program
- 2:59Astra system overview
- 7:40Cooling and data center challenges
- 8:00Software stack and ecosystem
- 15:28Performance and early results
- 18:17Containers and virtualization on ARM
- 21:10Optimizations and future work
- 23:04OpenHPC collaboration and lessons learned
- 25:13Memory bandwidth and HPL vs HPCC
- 28:53GPU considerations and future plans
- 30:10Workload focus and emerging trends
- 31:12Q&A and closing remarks
