Kaggle Free GPU and TPU Notebooks

Overview

Kaggle provides free browser-based Jupyter notebooks with GPU and TPU accelerators. Compute is allocated through a weekly quota instead of a paid plan, and notebooks come with a pre-configured Python or R environment.

Free Tier Quotas & Limits

Kaggle's documentation states:

• GPU: up to 30 hours per week; the quota resets weekly and is "sometimes higher depending on demand and resources".
• TPU: up to 20 hours per week on a TPU v3-8, with a maximum of 9 hours in a single session.
• Session limits: 12 hours execution time for CPU and GPU notebook sessions, 9 hours for TPU notebook sessions.
• Accelerators: one NVIDIA Tesla P100 (16 GB), two NVIDIA T4 GPUs (2 x 16 GB), or a TPU v3-8 with 8 cores and 128 GB of high-bandwidth memory.
• Per-session resources: 4 CPU cores with 29 GB RAM on GPU notebooks, 30 GB RAM on CPU notebooks, and 20 GB of auto-saved disk in /kaggle/working.
• Internet access can be enabled from the notebook settings pane. Free accelerator availability is limited, so busy periods can place sessions in a queue.

How to Get Started

Register for a free account at kaggle.com, then open or create a notebook at kaggle.com/code and pick an accelerator under Settings, Accelerator. For runs that outlast the browser, use "Save & Run All" so the notebook executes in a background session. Datasets can be attached read-only from /kaggle/input, and the Kaggle API allows notebooks to be pushed and executed from the command line.