> ## Documentation Index
> Fetch the complete documentation index at: https://docs.vast.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Rental Types FAQ

> Understanding on-demand vs interruptible instances

## Rental Type Overview

We currently offer three rental types:

### On-Demand (High Priority)

* Fixed price set by the host
* Runs as long as you want
* Cannot be interrupted
* More expensive but reliable

### Reserved (High Priority)

* Discounted rates with pre-payment
* Same priority as on-demand
* Convert from existing on-demand instances
* Up to 50% discount based on commitment length

For detailed information, see [Reserved Instances](/guides/instances/choosing/reserved-instances).

### Interruptible (Low Priority)

* You set a bid price
* Can be stopped by higher bids
* Saves 50-80% on costs
* Good for fault-tolerant workloads

## How do interruptible instances compare to AWS Spot?

**Similarities:**

* Both can be interrupted
* Both offer significant savings

**Differences:**

* Vast.ai uses direct bidding (you control your bid price)
* AWS uses market pricing
* No 24-hour limit like GCE preemptible instances
* Vast.ai instances can run indefinitely if not outbid

## What happens when my interruptible instance loses the bid?

Your instance is stopped (killing running processes). Important considerations:

* **Save work frequently** to disk
* **Use cloud storage** for backups
* **Instance may wait long** to resume
* **Implement checkpointing** for long jobs

When using interruptible instances, always design your workload to handle interruptions gracefully.

## DLPerf Scoring

### What is DLPerf?

DLPerf (Deep Learning Performance) is our scoring function that estimates performance for typical deep learning tasks. It predicts iterations/second for common tasks like training ResNet50 CNNs.

**Example scores:**

* V100: \~21 DLPerf
* 2080 Ti: \~14 DLPerf
* 1080 Ti: \~10 DLPerf

A V100 (21) is roughly 2x faster than a 1080 Ti (10) for typical deep learning.

### Is DLPerf accurate for my workload?

DLPerf is optimized for common deep learning tasks:

* ✅ CNN training (ResNet, VGG, etc.)
* ✅ Transformer models
* ✅ Standard computer vision
* ⚠️ Less accurate for unusual compute patterns
* ⚠️ Not optimized for non-ML workloads

For specialized workloads, benchmark on different GPUs yourself. While not perfect, DLPerf is more useful than raw TFLOPS for most ML tasks.