CLUSTERATLAS
A field guide to supercomputing

Perlmutter

NERSC · Berkeley Lab

CPU Accelerator Interconnect
i
Reading this map

Perlmutter node architecture

Perlmutter is operated by NERSC · Berkeley Lab in United States. This profile covers the documented compute inventory. Specifications are listed below; anything the operator has not published is itemised in the documented-fields list.

Published benchmark (TOP500 · June 2026)

HPL rank
#34
HPL Rmax
79.23 PFlop/s
Theoretical peak Rpeak
113 PFlop/s
Ranked core count
888,832
Measured power
2945 kW
Energy efficiency
26.90 GFlop/s per watt

Cluster · GPU · A100 40 GB

Operator / site
NERSC · Berkeley Lab
Country
United States
Profile scope
Documented compute inventory
Data checked
13 September 2026
Manufacturer
HPE
Ranked core count
888832
Theoretical peak (PFlop/s)
113
Measured power (kW)
2945
Energy efficiency (GFlop/s per watt)
26.903225806451612
Operating system
HPE Cray OS
Network topology
3-hop dragonfly

GPU · A100 40 GB

Compute nodes of this type
1536
CPU model
AMD EPYC 7763 · 64 cores
CPU sockets per node
1
Cores per socket
64
Physical CPU cores
64
Accelerator model
NVIDIA A100
Physical accelerators per node
4
Host / unified memory
256 GB DDR4
Accelerator memory
40 GB HBM2 per GPU
Accelerator ↔ accelerator link
NVLink 3 · all GPU pairs
On-node fabric
NVLink 3 — four links between each GPU pair, 25 GB/s per direction per link
CPU ↔ accelerator link
PCIe 4.0
Network interfaces per node
4
Interface model
HPE Cray Cassini (Slingshot 11)
Interface speed
200 Gb/s (25 GB/s) per NIC
Interface-to-accelerator topology
Four NICs on each GPU node, attached to the CPU over PCIe 4.0
Local storage
Shared Lustre scratch; local compute disk not documented
Form factor
GPU cabinet: 8 chassis, each 8 compute blades and 4 switch blades; 2 GPU nodes per blade

GPU · A100 80 GB

Compute nodes of this type
256
CPU model
AMD EPYC 7763 · 64 cores
CPU sockets per node
1
Cores per socket
64
Physical CPU cores
64
Accelerator model
NVIDIA A100
Physical accelerators per node
4
Host / unified memory
256 GB DDR4
Accelerator memory
80 GB HBM2e per GPU
Accelerator ↔ accelerator link
NVLink 3 · all GPU pairs
On-node fabric
NVLink 3 — four links between each GPU pair, 25 GB/s per direction per link
CPU ↔ accelerator link
PCIe 4.0
Network interfaces per node
4
Interface model
HPE Cray Cassini (Slingshot 11)
Interface speed
200 Gb/s (25 GB/s) per NIC
Form factor
GPU cabinet: 8 chassis, each 8 compute blades and 4 switch blades; 2 GPU nodes per blade

CPU only

Compute nodes of this type
3072
CPU model
AMD EPYC 7763 · 64 cores
CPU sockets per node
2
Cores per socket
64
Physical CPU cores
128
Host / unified memory
512 GB DDR4
Network interfaces per node
1
Interface model
HPE Cray Cassini (Slingshot 11)
Interface speed
200 Gb/s (25 GB/s) per NIC
Form factor
CPU cabinet: 8 chassis, each 8 compute blades and 2 switch blades; 4 CPU nodes per blade

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.

Documented fields: 51 / 66

Every field this catalog checks. Fields listed above have a value; the remainder are shown here. A field marked absent is not publicly available in the sources we could verify, which does not mean the hardware is missing. Fields that cannot apply, such as benchmark results for an unranked system, are excluded from the total.

Cluster — 7 / 7

  • Manufacturer
  • Ranked core count
  • Theoretical peak (PFlop/s)
  • Measured power (kW)
  • Energy efficiency (GFlop/s per watt)
  • Operating system
  • Network topology

GPU · A100 40 GB — 18 / 22

  • System / node model
  • Compute nodes of this type
  • CPU model
  • CPU sockets per node
  • Cores per socket
  • Physical CPU cores
  • CPU note
  • Accelerator model
  • Physical accelerators per node
  • Host / unified memory
  • Accelerator memory
  • Accelerator ↔ accelerator link
  • On-node fabric
  • CPU ↔ accelerator link
  • Network interfaces per node
  • Interface model
  • Interface speed
  • Interface-to-accelerator topology
  • Aggregate scale-out per node
  • GPUDirect RDMA
  • Management / storage networking
  • Local storage
  • NVMe data disks
  • Max remote data disks
  • Uncached disk IOPS
  • Uncached disk throughput (MBps)
  • Form factor

GPU · A100 80 GB — 16 / 22

  • System / node model
  • Compute nodes of this type
  • CPU model
  • CPU sockets per node
  • Cores per socket
  • Physical CPU cores
  • CPU note
  • Accelerator model
  • Physical accelerators per node
  • Host / unified memory
  • Accelerator memory
  • Accelerator ↔ accelerator link
  • On-node fabric
  • CPU ↔ accelerator link
  • Network interfaces per node
  • Interface model
  • Interface speed
  • Interface-to-accelerator topology
  • Aggregate scale-out per node
  • GPUDirect RDMA
  • Management / storage networking
  • Local storage
  • NVMe data disks
  • Max remote data disks
  • Uncached disk IOPS
  • Uncached disk throughput (MBps)
  • Form factor

CPU only — 10 / 15

  • System / node model
  • Compute nodes of this type
  • CPU model
  • CPU sockets per node
  • Cores per socket
  • Physical CPU cores
  • CPU note
  • Accelerator model
  • Physical accelerators per node
  • Host / unified memory
  • Accelerator memory
  • Accelerator ↔ accelerator link
  • On-node fabric
  • CPU ↔ accelerator link
  • Network interfaces per node
  • Interface model
  • Interface speed
  • Interface-to-accelerator topology
  • Aggregate scale-out per node
  • GPUDirect RDMA
  • Management / storage networking
  • Local storage
  • NVMe data disks
  • Max remote data disks
  • Uncached disk IOPS
  • Uncached disk throughput (MBps)
  • Form factor

Sources

Browse all cluster architecture pages →

Data & sources