Available GPUs
| Type | VRAM | $/hr | Best for |
|---|---|---|---|
| H200 | 141GB | $3.59 | Frontier model training |
| H100 SXM | 80GB | $2.69 | Large model training |
| H100 PCIe | 80GB | $1.99 | Large model training |
| A100 SXM | 80GB | $1.39 | General training |
| A100 PCIe | 80GB | $1.19 | General training |
| L40S | 48GB | $0.79 | Inference, fine-tuning, video gen |
| L40 | 48GB | $0.69 | Inference, fine-tuning |
| RTX 6000 Ada | 48GB | $0.74 | Workstation workloads |
| A40 | 48GB | $0.35 | Cost-effective large models |
| L4 | 24GB | $0.44 | Efficient inference, fast iteration |
| A10G | 24GB | $0.38 | Good all-rounder |
| RTX 4090 | 24GB | $0.34 | Fast iteration |
| RTX 3090 | 24GB | $0.22 | Budget training |
Selecting a GPU
cassian init shows live availability so you only pick what is ready.
Switching GPUs
Auto-stop
Instances automatically stop after 30 minutes of idle time (no GPU or CPU activity). You won’t get charged for forgetting to runcassian down.
Billing
- Credits are deducted per minute while your instance is running.
- If your balance reaches $0, the instance is automatically stopped.
- If a container crashes, billing stops within 60 seconds.
- Check your balance anytime with
cassian statusor in the dashboard.