Compute¶
Nodes¶
| Role | Minimum | Purpose |
|---|---|---|
| CPU | 1 node - 4 vCPU, 16 GB RAM, 100 GB disk | Registry, gateway, application layer |
| GPU | 1 node with NVIDIA GPUs | Model inference |
All target nodes must report Ready. Nodes are selected by label, so workload placement
labels must be applied before installation.
GPU¶
GPUs must be advertised to Kubernetes as the nvidia.com/gpu extended resource, via
either:
- the NVIDIA GPU Operator - typical on-prem, EKS, AKS
- the NVIDIA device plugin - typical on GKE, often provider-installed
Both Capacity and Allocatable must show a non-zero nvidia.com/gpu count.
Drivers¶
NVIDIA drivers must be installed on every GPU node before installation. shaide does not install drivers - they are tied to your node image and kernel.
Reference: NVIDIA GPU Operator platform support.
VRAM sizing¶
VRAM per replica must cover the model weights plus its KV cache.
| Model size | Precision | VRAM before KV cache |
|---|---|---|
| ~20B | FP8 | ~24 GB |
| ~30B | FP8 | ~35 GB |
| ~30B | BF16 | ~70 GB |
See Model catalog for validated models.
Per-target notes¶
| Target | Note |
|---|---|
| EKS | GPU AMIs require the device plugin |
| GKE | GPU node pools require driver installation enabled |
| AKS | GPU VM sizes are quota-limited per region |
| RKE2 | Drivers and the GPU Operator are your responsibility |