Berkelium Cluster Specifications

128+

Compute Nodes

96x

Enterprise GPUs (A100/H100/L40S)

4.5 PB

Ceph & NVMe Storage

100 Gbps

ESnet / Interconnect Fabric

Compute & Acceleration Node Pools

GPU Accelerated

Dedicated and MIG-sliced GPU nodes optimized for PyTorch, TensorFlow, Jax, and Cryo-EM structural simulations.

High-Memory CPU

Large memory footprint nodes for genome assembly, graph analytics, in-memory databases, and large batch pipelines.

Interactive & Web

General application hosting, JupyterHub instances, web APIs, scientific portals, and pipeline orchestrators.

Berkelium provides low-latency, scalable storage classes configured specifically for massive scientific data formats (HDF5, NetCDF, Parquet, TIFF):

cephfs-replicated (ReadWriteMany)

Triple-replicated Ceph filesystem. Allows simultaneous read/write access across dozens of pods across all nodes.

nvme-scratch-local (ReadWriteOnce)

Ultra-fast direct PCIe NVMe storage for temporary intermediate model training data and high-IOPS shuffles.

LBNL Science Data Bridge

Read-only NFS/POSIX mounts connecting directly to central laboratory instrument repositories and NERSC archives.

Engineered from the ground up for zero-trust scientific multi-tenancy compliant with DOE and UC Berkeley standards:

LBNL OneID / OIDC Authentication

Direct integration with Berkeley Lab identity systems. No manual key exchange; authorization based on division and group membership.

Strict Namespace Isolation & Quotas

Granular ResourceQuotas and LimitRanges prevent noisy-neighbor issues and cap GPU consumption according to grant allocations.

Automated Secret Management

HashiCorp Vault Agent sidecars automatically inject sensitive credentials into container memory without storing secrets on disk.

Provision a dedicated Kubernetes namespace with custom CPU, memory, and GPU quotas through our self-service portal.