NVIDIA GPU Accelerated Pods

Berkelium hosts high-density GPU nodes equipped with NVIDIA H100 80GB (SXM5) with NVLink interconnects (900 GB/s), A100 80GB (NVLink), and L40S 48GB inference accelerators.

Requesting a GPU in a Pod

To schedule a pod on a GPU node, request the nvidia.com/gpu resource limit:

apiVersion: v1
kind: Pod
metadata:
  name: pytorch-gpu-job
  namespace: sci-myproject
spec:
  containers:
    - name: cuda-container
      image: nvcr.io/nvidia/pytorch:24.04-py3
      command: ["python", "-c", "import torch; print(f'CUDA available: {torch.cuda.is_available()}, Devices: {torch.cuda.device_count()}, Device name: {torch.cuda.get_device_name(0)}')"]
      resources:
        limits:
          cpu: "16"
          memory: "64Gi"
          nvidia.com/gpu: "1"  # Number of GPUs requested

---

Targeting Specific GPU Architectures

If your scientific workload requires specific GPU models (e.g. FP8 Transformer Engine on H100 vs. A100), use nodeSelector:

Target NVIDIA H100 80GB SXM5 Nodes

spec:
  nodeSelector:
    nvidia.com/gpu.product: "NVIDIA-H100-80GB-HBM3"
  containers:
    - name: train
      resources:
        limits:
          nvidia.com/gpu: "4"

Target NVIDIA A100 80GB Nodes

spec:
  nodeSelector:
    nvidia.com/gpu.product: "NVIDIA-A100-SXM4-80GB"
  containers:
    - name: train
      resources:
        limits:
          nvidia.com/gpu: "2"

---

Multi-Instance GPU (MIG) Slices

For smaller exploration workloads or inference jobs that do not need a full 80GB GPU, Berkelium supports MIG partitioning:

• nvidia.com/mig-1g.10gb (10GB VRAM slice)
• nvidia.com/mig-2g.20gb (20GB VRAM slice)
• nvidia.com/mig-3g.40gb (40GB VRAM slice)