On the same platform and with the same GPU, spot instance prices can drop to 30–50% of on-demand rates (e.g., Gpushare’s RTX 3090 spot at ¥0.3/hour; Vast.ai’s RTX 3090 spot at ~¥0.5/hour). This discount stems from:You’re using idle resources temporarily released by others—platforms may reclaim them instantly during peak demand。
Tasks suitable for spot instances
· Batch inference / batch image generation (tasks are splittable; restarts cost nothing if interrupted).
· Training with checkpointing (auto-saved every 15–30 minutes);
· Experiments with flexible deadlines.
Tasks you should never run
· 6-hour fine-tuning without checkpointing (if preempted = restart from scratch);
· Report due tomorrow / Night before defense;
· Inference API for external services (stability is critical).
Best Practices
Add periodic checkpointing to your training script (e.g., PyTorch checkpoints or SD’s auto-saved output directories), and design tasks assuming they may fail at any moment. This way, spot instances become pure savings—50–70% off. Spot pricing is separately labeled in our price comparison tables and excluded from lowest-price recommendations—cheaper, yes, but confirm your workload can tolerate interruptions first.