Perlmutter node architecture
Perlmutter is operated by NERSC · Berkeley Lab in United States. This profile covers the documented compute inventory.
Cluster interconnect: HPE Slingshot 11 · 200 Gb/s per NIC. Topology: Dragonfly.
GPU · A100 40 GB
- Node count
- 1536
- CPU
- AMD EPYC 7763 · 64 cores
- CPUs per node
- 1
- Accelerator
- NVIDIA A100
- Physical accelerators per node
- 4
- Host / unified memory
- 256 GB DDR4
- Accelerator memory
- 40 GB HBM2 per GPU
- Local accelerator links
- NVLink 3 · all GPU pairs
- CPU-to-accelerator link
- PCIe 4.0
- Storage
- Shared Lustre scratch; local compute disk not documented
GPU · A100 80 GB
- Node count
- 256
- CPU
- AMD EPYC 7763 · 64 cores
- CPUs per node
- 1
- Accelerator
- NVIDIA A100
- Physical accelerators per node
- 4
- Host / unified memory
- 256 GB DDR4
- Accelerator memory
- 80 GB HBM2e per GPU
- Local accelerator links
- NVLink 3 · all GPU pairs
- CPU-to-accelerator link
- PCIe 4.0
- Storage
- Not verified
CPU only
- Node count
- 3072
- CPU
- AMD EPYC 7763 · 64 cores
- CPUs per node
- 2
- Accelerator
- No accelerator
- Physical accelerators per node
- 0
- Host / unified memory
- 512 GB DDR4
- Accelerator memory
- Not verified
- Local accelerator links
- Not applicable
- CPU-to-accelerator link
- PCIe 4.0
- Storage
- Not verified
Scope and uncertainty
Compute nodes only. GPU groups contain 64 nodes; CPU groups contain 256. Network has 16 switches per group. Overview samples groups, not physical cable routes.
Not verified means a detail has not been confirmed in this catalog. Diagrams are schematic; node indices are illustrative, not physical hostnames. Data checked 13 September 2026.
Sources
Browse all cluster architecture pages →
CLUSTER