New NVIDIA H100 and L40S GPU Pools Added to Berkelium

We are excited to announce a major hardware expansion on Berkelium. Science IT has deployed high-density GPU nodes equipped with NVIDIA H100 Tensor Core GPUs and NVIDIA L40S accelerators, available immediately to all research groups with GPU quota allocations.

Key Capabilities

• Hopper Architecture & FP8 Precision: H100 nodes offer up to 6x higher compute throughput for large transformer training and molecular dynamics simulations compared to previous generation cards.
• MIG (Multi-Instance GPU) Partitioning: L40S and H100 nodes support flexible MIG configurations, allowing lightweight development pods to share GPU hardware without contention.
• NVLink & Fast Host Interconnect: High-bandwidth inter-GPU communication enables scalable multi-GPU PyTorch distributed training within a single node.

# Example pod resource request for H100 GPU
apiVersion: v1
kind: Pod
metadata:
  name: pytorch-train-h100
spec:
  containers:
  - name: cuda-container
    image: nvcr.io/nvidia/pytorch:26.04-py3
    resources:
      limits:
        nvidia.com/gpu: 2

To request access to the H100 node pool for your project namespace, submit a quota adjustment request via the Namespace Portal.