Compute Resources
Podstack offers multiple ways to deploy compute workloads, from containerized applications to dedicated GPU instances.
Compute Options
Pods (Containers)
Pods are containerized workloads running on Kubernetes. They offer:
- Fast deployment from Docker images
- Fractional or whole GPU allocation
- Web terminal and SSH access
- Jupyter notebook integration
- Auto-scaling with replicas
Best for: ML training, inference, Jupyter notebooks, containerized applications
GPU Marketplace (Baremetal)
Reserve dedicated GPU instances from the marketplace:
- Browse available inventory across multiple GPU types
- Dedicated hardware with no virtualization overhead
- Ideal for large-scale training jobs
Best for: Maximum GPU performance, dedicated resources
Comparing Options
| Feature | Pods | Baremetal |
|---|---|---|
| Deployment Speed | Fast (seconds) | Varies |
| GPU Sharing | Fractional supported | Dedicated |
| OS Customization | Container image | Full OS |
| Billing Granularity | Per-second | Per-hour |
| Best For | Dev/ML | Production training |
GPU Types Available
Podstack supports various NVIDIA GPUs:
| GPU | Memory | Best For |
|---|---|---|
| A100 | 40GB/80GB | Large model training |
| H100 | 80GB | Latest generation training |
| H200 | 141GB | Memory-intensive workloads |
| V100 | 16GB/32GB | Cost-effective training |
| L40S | 48GB | Inference and training |
| T4 | 16GB | Budget inference |
Availability varies by region and demand.
Next Steps
- Create a Pod for quick container deployment
- Browse GPU Marketplace for dedicated instances