# Perlmutter: every supplier in the machine

NERSC (Lawrence Berkeley National Laboratory) · US · system record on Compute Atlas: https://computeatlas.net/supercomputers/systems/perlmutter

## Specs

- **accelerator:** NVIDIA A100 (SXM4), 4 per GPU node; 40 GB HBM2 on 1,536 nodes and 80 GB HBM2e on 256 nodes; TSMC 7nm
- **host cpu:** GPU nodes: 1 x AMD EPYC 7763 (64 cores); CPU nodes: 2 x AMD EPYC 7763 (64 cores each)
- **integrator:** HPE (HPE Cray EX, Shasta platform; TOP500 lists the system as HPE Cray EX 235n)
- **nodes:** 1,792 GPU nodes (1,536 with A100 40 GB and 256 with A100 80 GB) and 3,072 CPU-only nodes, plus 40 login nodes
- **memory:** GPU node: 256 GB DDR4 plus 4 x A100 HBM; CPU node: 512 GB DDR4; aggregate 328 TB (GPU partition) and 1,536 TB (CPU partition) per NERSC's partition table
- **interconnect:** HPE Slingshot 11, 200 Gb/s Cassini NICs (4 per GPU node, 1 per CPU node), dragonfly topology
- **storage:** Perlmutter scratch: all-flash Lustre on HPE Cray ClusterStor E1000, 44 PB, with 16 metadata servers and 370 I/O servers; also mounts NERSC's Community File System
- **cooling:** Direct liquid cooling in HPE Cray EX liquid-cooled cabinets, with switch blades and ports also liquid cooled
- **power mw:** 2.9 MW (2,945 kW, power measurement level 3, TOP500 June 2026 list)
- **rmax pflops:** 79.23 PFlop/s (TOP500 June 2026 list)
- **rpeak pflops:** 113.00 PFlop/s (TOP500 June 2026 list)
- **status note:** Phase 1 GPU partition unveiled 27 May 2021 on Slingshot 10; the 11/2023 TOP500 entry is the first with Slingshot-11 and the full 888,832-core configuration; TOP500 gives the installation year as 2021; a DOE Office of Science user facility

Unmarked companies come from the operator's or a vendor's own pages (docs.nersc.gov, nersc.gov, cug.org paper by NERSC staff, nvidia.com, hpe.com, top500.org); reported companies come from named trade press or analyst reporting whose article body was read in full, not just the headline. This is a map of who plays where, not a confirmed bill of materials: NERSC documentation (docs.nersc.gov), NERSC and Berkeley Lab news pages, a NERSC-authored Cray User Group 2021 paper on Perlmutter's all-flash file system, NVIDIA's A100 architecture write-up, an HPE cabinet brochure and the TOP500 system entry (June 2026 list), researched 2026-10-10

## Lines

1. System design and cabinets (HPE Cray EX liquid-cooled cabinets, eight compute chassis each; GPU cabinets and CPU-only cabinets)
   - parts: Rack-scale systems (4.01)
   - companies: HPE (system integrator and cabinet designer (HPE Cray EX, Shasta platform))
   - sources: https://docs.nersc.gov/systems/perlmutter/architecture/ https://assets.ext.hpe.com/is/content/hpedam/a50002389enw
2. GPU compute blade (HPE Cray EX 235n: two nodes per blade, each with one EPYC 7763, four A100 and four Slingshot NICs; 1,792 nodes)
   - parts: GPU baseboards & compute trays (4.03)
   - companies: HPE (Cray EX compute blade and node design)
   - not confirmed: maker of the node's DDR4 DIMMs (256 GB per node) — not named in any source read; maker of the A100 SXM4 baseboard and the NVLink board — not named in any source read
   - sources: https://docs.nersc.gov/systems/perlmutter/architecture/ https://www.top500.org/system/179972/
3. NVIDIA A100 accelerator (Ampere GA100, 40 GB HBM2 or 80 GB HBM2e, fabricated on TSMC 7nm; four per GPU node)
   - parts: Merchant GPUs / accelerators (1.01); Leading-edge foundry (1.09)
   - companies: Nvidia (designer of the A100 GPU); TSMC (7nm N7 process for the GA100 die, per NVIDIA's Ampere architecture write-up)
   - not confirmed: maker of the HBM2 and HBM2e stacks — NVIDIA and NERSC give capacity and bandwidth but name no memory supplier
   - sources: https://docs.nersc.gov/systems/perlmutter/architecture/ https://developer.nvidia.com/blog/nvidia-ampere-architecture-in-depth/
4. GPU-node host CPU (one 64-core AMD EPYC 7763 "Milan" per node, PCIe 4.0 to the GPUs and NICs)
   - parts: Server CPUs (1.05)
   - companies: AMD
   - not confirmed: foundry for the EPYC 7763 — no source read names it
   - sources: https://docs.nersc.gov/systems/perlmutter/architecture/
5. CPU partition (3,072 nodes, four per compute blade, each with two 64-core AMD EPYC 7763 and 512 GB DDR4)
   - parts: GPU baseboards & compute trays (4.03); Server CPUs (1.05)
   - companies: HPE (Cray EX CPU compute blade, four nodes per blade); AMD (two EPYC 7763 "Milan" CPUs per node)
   - not confirmed: maker of the node DDR4 memory (512 GB per node) — not named in any source read
   - sources: https://docs.nersc.gov/systems/perlmutter/architecture/ https://www.nersc.gov/news-and-events/news/berkeley-lab-deploys-next-generation-supercomputer-perlmutter-bolstering-u-s-scientific-research
6. Network interface (200 Gb/s HPE Slingshot "Cassini" NIC; four per GPU node, one per CPU node)
   - parts: NICs / SuperNICs (1.06)
   - companies: HPE (Slingshot NICs)
   - sources: https://docs.nersc.gov/systems/perlmutter/architecture/
7. Slingshot switches and dragonfly fabric (switch blades at the rear of each cabinet; 16-switch groups; copper inside groups, optical between)
   - parts: Scale-out switch ASICs (3.02); Switch systems & white-box ODMs (3.03); Leading-edge foundry (1.09)
   - companies: HPE (designs and supplies the Slingshot switches); TSMC (Rosetta switch ASIC process, per a paper co-authored by HPE engineers)
   - not confirmed: maker of the copper and optical cables and transceivers — NERSC describes copper L0/L1 cables and optical L2 cables but names no cable maker
   - sources: https://docs.nersc.gov/systems/perlmutter/architecture/ https://arxiv.org/abs/2008.08886
8. Perlmutter scratch file system (all-flash Lustre on HPE Cray ClusterStor E1000 enclosures, 44 PB; servers sit directly on the Slingshot network)
   - parts: AI storage systems (2.10)
   - companies: HPE (ClusterStor E1000 enclosures (two server canisters and 24 NVMe bays each))
   - sources: https://cug.org/proceedings/cug2021_proceedings/includes/files/pap120s2-file1.pdf https://docs.nersc.gov/filesystems/perlmutter-scratch/
9. Scratch SSDs (15.36 TB Samsung PM1733 NVMe drives in every E1000 enclosure of the original 35 PB build)
   - parts: Enterprise SSDs & controllers (2.08)
   - companies: Samsung Electronics (PM1733 NVMe SSDs, 15.36 TB each, per NERSC's CUG 2021 paper)
   - not confirmed: maker of the drives added in the expansion from 35 PB to the 44 PB the current documentation lists — not named in any source read
   - sources: https://cug.org/proceedings/cug2021_proceedings/includes/files/pap120s2-file1.pdf
10. Scratch storage servers (single-socket AMD EPYC 7502 in each E1000 server canister, with two Slingshot NICs per server)
   - parts: Server CPUs (1.05)
   - companies: AMD (EPYC 7502 "Rome" CPU in each Lustre OSS and MDS)
   - sources: https://cug.org/proceedings/cug2021_proceedings/includes/files/pap120s2-file1.pdf
11. External network gateway (two Arista 7804 routers bridging the Slingshot fabric to NERSC's Ethernet networks and archive)
   - parts: Switch systems & white-box ODMs (3.03)
   - companies: Arista (7804 routers, 400 Gb/s per port, in Perlmutter's service group)
   - sources: https://cug.org/proceedings/cug2021_proceedings/includes/files/pap120s2-file1.pdf
12. In-cabinet direct liquid cooling (cold plates on the compute blades, liquid-cooled switch blades and ports, coolant distribution units)
   - parts: Direct-to-chip cold plates (5.01); Coolant distribution units (5.05)
   - companies: HPE (cabinet, blade cold plates and CDUs are part of the HPE Cray EX design)
   - not confirmed: makers of the cold plates, quick disconnects and CDU internals — not named in any source read; Perlmutter's own CDU count, coolant and loop temperatures — not found; the HPE brochure figures are generic Cray EX values
   - sources: https://www.nersc.gov/systems/perlmutter/ https://assets.ext.hpe.com/is/content/hpedam/a50002389enw

* reported: named trade press or analyst reporting rather than the operator or vendor.

## What could not be confirmed

No source read names the maker of the DDR4 DIMMs, the HBM2 and HBM2e stacks on the A100 modules, the A100 SXM4 baseboard, the node-local drives, the Slingshot cables and optics, or the cold plates, quick disconnects, rectifiers and coolant distribution units inside the cabinets. NERSC describes Perlmutter as direct liquid cooled and, when DOE announced the system in 2018, said Berkeley Lab was launching a power and cooling upgrade in Shyh Wang Hall to prepare the building; no source read names who built the electrical or cooling plant, so no facility line is listed. The scratch file system's current 44 PB expansion beyond the original 35 PB build is described by NERSC without naming the added drives' maker. NERSC's pages disagree on several scratch figures, as noted in the specs.