Berkelium Cluster Specifications

Cluster Hardware Inventory (Cached Snapshot)• Refreshed 2:57:39 AM

Live capacity and free resources dynamically calculated from active pods and allocatable node limits.

11 / 11
Ready Nodes
100% Operational
695 / 736
Free CPU Cores
94% Schedulable
3.77 TiB
Free RAM
4.34 TiB Total Capacity
14 / 32
Free GPUs Right Now
18 in Active Use
~40 TiB
Local NVMe Scratch
High IOPS per Node

What Can I Submit Right Now?

Live scheduling guidance based on unallocated, instantly available cluster resources.

Immediate Dispatch (Zero Wait)
AMD ROCm Pool12 of 12 Free

All 12x AMD Instinct MI100 GPUs across k8s-0001, k8s-0003, and k8s-0004 are idle and schedulable immediately.

Max instant single pod:
4 GPUs • 62 Cores • 500 GiB RAM
resources.requests["amd.com/gpu"]: 4
NVIDIA A40 Pool2 of 8 Free

Node k8s-0006 has 2x NVIDIA A40 (48GB GDDR6) ready for instant dispatch for beamline processing and LLM inference.

Max instant single pod on k8s-0006:
2 GPUs • 60 Cores • 459 GiB RAM
resources.requests["nvidia.com/gpu"]: 2
H100 & A100 StatusQueueing Active

All 8x H100 and 4x A100 GPUs are currently running active research jobs. Submissions will queue and schedule automatically as pods terminate.

Active flagship utilization:
8/8 H100 (100%) • 4/4 A100 (100%)
Pods enter Pending phase until slots free

Live GPU Accelerator Pools

100% In Use

NVIDIA-H100-80GB-HBM3

80 GiB VRAM per GPU • 8 Total GPUs
Allocated: 8/8100%
Available Free:0 GPUs
100% In Use

NVIDIA-A100-80GB-PCIe

80 GiB VRAM per GPU • 4 Total GPUs
Allocated: 4/4100%
Available Free:0 GPUs
2 Free Now

NVIDIA-A40

48 GiB VRAM per GPU • 8 Total GPUs
Allocated: 6/875%
Available Free:2 GPUs
12 Free Now

AMD Instinct MI100

32 GiB VRAM per GPU • 12 Total GPUs
Allocated: 0/120%
Available Free:12 GPUs

Dynamic Node Inventory & Free Resources

Inspect per-node allocatable limits, running pod reservations, and remaining headroom.

NodeStatusFree CPU / AllocatableFree Memory (RAM)NVMe ScratchFree GPUsAccelerator ModelActive PodsDetails
k8s-0000Ready
61.3 / 64 Cores
118 GiB
of 125 GiB
750 GiB
of 800 GiB
No GPUs
CPU Compute Node
General CPU Worker & Ingress (64 Cores, 128GB RAM)
18 pods
k8s-0001Ready
50.9 / 64 Cores
453 GiB
of 503 GiB
6.15 TiB
of 6.35 TiB
4 / 4 Free
AMD Instinct MI100
4x AMD Instinct MI100 (ROCm / HIP)
25 pods
k8s-0002Ready
49.6 / 64 Cores
289 GiB
of 503 GiB
5.96 TiB
of 6.35 TiB
0 / 4 (Full)
NVIDIA-A40
4x NVIDIA A40 (48GB GDDR6)
20 pods
k8s-0003Ready
58.5 / 64 Cores
483 GiB
of 503 GiB
6.25 TiB
of 6.35 TiB
4 / 4 Free
AMD Instinct MI100
4x AMD Instinct MI100 (ROCm / HIP)
34 pods
k8s-0004Ready
62.9 / 64 Cores
500 GiB
of 503 GiB
6.30 TiB
of 6.35 TiB
4 / 4 Free
AMD Instinct MI100
4x AMD Instinct MI100 (ROCm / HIP)
19 pods
k8s-0005Ready
63.5 / 64 Cores
437 GiB
of 503 GiB
6.20 TiB
of 6.35 TiB
0 / 4 (Full)
NVIDIA-A100-80GB-PCIe
4x NVIDIA A100 (80GB PCIe)
21 pods
k8s-0006Ready
60.8 / 64 Cores
459 GiB
of 503 GiB
6.23 TiB
of 6.35 TiB
2 / 4 Free
NVIDIA-A40
4x NVIDIA A40 (48GB GDDR6)
23 pods
k8s-0007Ready
223.4 / 224 Cores
877 GiB
of 1007 GiB
1.07 TiB
of 1.37 TiB
0 / 8 (Full)
NVIDIA-H100-80GB-HBM3
8x NVIDIA H100 (80GB HBM3 Flagship)
22 pods

Storage Architecture

Berkelium provides low-latency, scalable storage classes configured specifically for massive scientific data formats:

cephfs (ReadWriteMany)

Shared Ceph filesystem. Allows simultaneous high-bandwidth read/write access across all 8 worker nodes and dozens of concurrent research pods.

Local NVMe Scratch (~40 TiB Total)

High-IOPS local NVMe scratch on each worker node (up to 6.29 TiB per node) for intermediate model checkpoints, dataset untarring, and fast caching.

LBNL Science Data Bridge & ESnet

Direct high-speed POSIX and 100Gbps network connectivity to ALS beamlines, NERSC data repositories, and national scientific facilities.

Security & Multi-Tenancy

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.

Strict Namespace Isolation & Quotas

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

Need compute resources for your research group?

Provision a dedicated Kubernetes namespace with custom CPU, memory, and GPU quotas across our 8 worker nodes.

Create NamespaceExplore Documentation