Kaggle Free GPU & TPU - 30 Hours/Week
Source: https://www.kaggle.com/code
Description
Create account to comment on specific lines or Sign in
+ 1 Kaggle (owned by Google) gives every verified account holder free weekly access to NVIDIA GPUs and Google TPU v3-8 accelerators through browser-based Jupyter notebooks. No credit card is required. You get 30 hours/week of GPU time and a separate 20–30 hours/week of TPU time (Kaggle uses a floating quota system that can vary with demand). Sessions run up to 9 hours for GPU/TPU notebooks and up to 12 hours for CPU-only notebooks, with background execution so training continues after you close the browser tab.
No comments on this line yet.
+ 2
No comments on this line yet.
+
3
No comments on this line yet.
+ 4
No comments on this line yet.
+ 6
No comments on this line yet.
+ 7 1. Go to kaggle.com and click Register
No comments on this line yet.
+ 8 2. Sign up with a Google account, email, or other supported method — completely free
No comments on this line yet.
+ 9 3. Navigate to your profile settings (click your avatar → Settings)
No comments on this line yet.
+ 10 4. Scroll to Phone Verification and verify your phone number via SMS code
No comments on this line yet.
+ 11 5. Phone verification is required to unlock GPU and TPU accelerators — without it you only get CPU notebooks
No comments on this line yet.
+ 12 6. Once verified, open or create any notebook at kaggle.com/code, click Settings in the right sidebar, and select your accelerator under the Accelerator dropdown (GPU P100, GPU T4 x2, or TPU v3-8)
No comments on this line yet.
+ 13
No comments on this line yet.
+ 14 Important:
No comments on this line yet.
+ 15 • One phone number per account — Kaggle enforces this to prevent abuse
No comments on this line yet.
+ 16 • Some users report difficulties with phone verification (certain carriers or VoIP numbers may not work)
No comments on this line yet.
+ 17 • No billing account or credit card is ever required
No comments on this line yet.
+ 18 • You can start using CPU notebooks immediately without phone verification
No comments on this line yet.
+ 19
No comments on this line yet.
+
20
No comments on this line yet.
+ 21
No comments on this line yet.
+ 23
No comments on this line yet.
+ 24 AcceleratorVRAM / MemoryDetails
No comments on this line yet.
+ 25 NVIDIA Tesla P10016 GB HBM23,584 CUDA cores; strong FP32 performance; best for standard precision training
No comments on this line yet.
+
26
NVIDIA T4 x2 (beta)2 × 16 GB GDDR6 (32 GB total)Tensor Cores optimized for FP16 / mixed precision; use DataParallel for distributed training
No comments on this line yet.
+ 27 TPU v3-8128 GB HBM (8 cores)Google's custom ASIC; excellent for large-batch TensorFlow and JAX workloads
No comments on this line yet.
+ 28
No comments on this line yet.
+ 30
No comments on this line yet.
+ 31 • P100 is faster for pure FP32 workloads — roughly 1.5× faster than dual T4 in FP32 due to superior memory bandwidth
No comments on this line yet.
+ 32 • T4 x2 shines with mixed precision (AMP / FP16) training — Tensor Cores go unused without AMP enabled
No comments on this line yet.
+ 33 • Dual T4 can suffer from CPU bottlenecks (CPU must decode data for both GPUs) and inter-GPU communication overhead
No comments on this line yet.
+ 34 • For most deep learning tasks with AMP enabled, T4 x2 offers more total VRAM (32 GB vs 16 GB)
No comments on this line yet.
+ 35
No comments on this line yet.
+
36
No comments on this line yet.
+ 37
No comments on this line yet.
+ 39
No comments on this line yet.
+ 40 ResourceLimit
No comments on this line yet.
+ 41 GPU quota30 hours/week (shared across P100 and T4 x2)
No comments on this line yet.
+ 42 TPU quota20–30 hours/week (floating quota, varies with demand)
No comments on this line yet.
+ 43 GPU/TPU session max9 hours per session
No comments on this line yet.
+ 44 CPU-only session max12 hours per session
No comments on this line yet.
+ 45 CPU weekly quotaUnlimited (no weekly cap, only per-session limit)
No comments on this line yet.
+ 46 Quota resetRolling weekly window
No comments on this line yet.
+ 47
No comments on this line yet.
+ 48 Notes:
No comments on this line yet.
+ 49 • Kaggle uses a "floating" quota system — your weekly hours may vary slightly based on overall platform demand
No comments on this line yet.
+ 50 • If you exhaust your GPU quota mid-week, you can still run CPU-only notebooks without restriction
No comments on this line yet.
+ 51 • Interactive sessions (in the browser editor) will prompt "Are you still there?" after inactivity and may terminate if not confirmed
No comments on this line yet.
+ 52
No comments on this line yet.
+
53
No comments on this line yet.
+ 54
No comments on this line yet.
+ 56
No comments on this line yet.
+ 57 Kaggle supports background execution via the "Save & Run All (Commit)" feature:
No comments on this line yet.
+ 58
No comments on this line yet.
+ 59 1. Click Save Version → select Save & Run All (Commit)
No comments on this line yet.
+ 60 2. A separate background kernel is launched that runs independently of your browser
No comments on this line yet.
+ 61 3. You can close the tab, and the notebook continues executing in the background
No comments on this line yet.
+
62
4. Output files are saved to /kaggle/working/ and become downloadable once the run completes
No comments on this line yet.
+ 63 5. The same session time limits apply (9 hours for GPU/TPU, 12 hours for CPU)
No comments on this line yet.
+ 64
No comments on this line yet.
+ 65 This is the recommended approach for long training runs — interactive sessions will time out if you walk away.
No comments on this line yet.
+ 66
No comments on this line yet.
+
67
No comments on this line yet.
+ 68
No comments on this line yet.
+ 70
No comments on this line yet.
+ 71 ResourceAmount
No comments on this line yet.
+ 72 CPU cores4 vCPUs
No comments on this line yet.
+ 73 RAM (GPU notebooks)~29 GB
No comments on this line yet.
+ 74 RAM (CPU notebooks)~16 GB
No comments on this line yet.
+
75
Disk (working directory)20 GB (/kaggle/working/)
No comments on this line yet.
+ 76 Output file limit500 files max in output directory
No comments on this line yet.
+ 77 Internet accessAvailable (must be enabled in Settings; disabled by default in some competitions)
No comments on this line yet.
+ 78
No comments on this line yet.
+
79
No comments on this line yet.
+ 80
No comments on this line yet.
+ 82
No comments on this line yet.
+ 83 • /kaggle/input/ — read-only mount for attached datasets (does not count against your 20 GB disk)
No comments on this line yet.
+ 84 • /kaggle/working/ — 20 GB writable workspace; output files saved here persist after commit
No comments on this line yet.
+ 85 • Kaggle Datasets — access 50,000+ community datasets; attach any dataset to your notebook with one click via "Add Data"
No comments on this line yet.
+ 86 • Kaggle Models — 200+ pre-trained model weights available for direct import
No comments on this line yet.
+ 87 • Private datasets — upload your own data (up to 100 GB per dataset) and attach to notebooks
No comments on this line yet.
+ 88 • Tip: compress large outputs with zip/tar to stay within the 500-file and 20 GB limits
No comments on this line yet.
+ 89
No comments on this line yet.
+
90
No comments on this line yet.
+ 91
No comments on this line yet.
+ 93
No comments on this line yet.
+ 94 Kaggle notebooks come with a comprehensive pre-configured environment:
No comments on this line yet.
+ 95
No comments on this line yet.
+ 96 • Python 3.10+ with NumPy, Pandas, Scikit-learn, Matplotlib, Seaborn
No comments on this line yet.
+ 97 • Deep Learning: TensorFlow, PyTorch, Keras, JAX, XGBoost, LightGBM, CatBoost
No comments on this line yet.
+ 98 • NLP: HuggingFace Transformers, Tokenizers, Datasets
No comments on this line yet.
+ 99 • Computer Vision: OpenCV, Albumentations, torchvision
No comments on this line yet.
+ 100 • R notebooks also supported
No comments on this line yet.
+
101
• Install additional packages with !pip install (internet must be enabled)
No comments on this line yet.
+ 102
No comments on this line yet.
+
103
No comments on this line yet.
+ 104
No comments on this line yet.
+ 106
No comments on this line yet.
+ 107 • Fine-tuning LLMs — 16–32 GB VRAM is enough for QLoRA fine-tuning of 7B–13B parameter models
No comments on this line yet.
+ 108 • Training image classifiers — P100 or T4 handles standard ResNet/EfficientNet training comfortably
No comments on this line yet.
+ 109 • Kaggle competitions — the platform is purpose-built for this; many competition winners train exclusively on Kaggle GPUs
No comments on this line yet.
+ 110 • Prototyping and experimentation — zero setup, just open a notebook and start coding
No comments on this line yet.
+ 111 • Running inference on large models — load HuggingFace models directly from Kaggle Models
No comments on this line yet.
+ 112
No comments on this line yet.
+
113
No comments on this line yet.
+ 114
No comments on this line yet.
+ 116
No comments on this line yet.
+
117
• No persistent VM — each session starts fresh; only /kaggle/working/ outputs survive via commits
No comments on this line yet.
+ 118 • No SSH or terminal access — you work through the notebook interface only
No comments on this line yet.
+ 119 • No deployment tools — Kaggle is for experimentation, not production serving
No comments on this line yet.
+ 120 • Queue wait times — during peak demand, GPU/TPU allocation may be delayed
No comments on this line yet.
+ 121 • Older GPUs — P100 (2016) and T4 (2018) are capable but not cutting-edge; no A100 or H100 access
No comments on this line yet.
+ 122 • Session time limits — 9-hour GPU sessions may not be enough for very large training runs
No comments on this line yet.
+ 123 • Floating quota — your actual weekly hours may be less than 30 during high-demand periods
No comments on this line yet.
+ 124
No comments on this line yet.
+
125
No comments on this line yet.
+ 126
No comments on this line yet.
+ 128
No comments on this line yet.
+ 129 • Use background execution (commit) for all training runs — don't rely on keeping a browser tab open
No comments on this line yet.
+ 130 • Enable mixed precision (AMP) when using T4 GPUs to leverage Tensor Cores
No comments on this line yet.
+
131
• Save checkpoints frequently to /kaggle/working/ so you can resume across sessions
No comments on this line yet.
+ 132 • Chain notebooks — save model checkpoints as a Kaggle dataset, then load them in a new notebook to continue training
No comments on this line yet.
+
133
• Use the Kaggle API (kaggle kernels push) to automate notebook execution from your local machine or CI/CD
No comments on this line yet.
+ 134 • Combine with Google Colab — use KaggleHub to share datasets between Kaggle and Colab for additional free GPU hours
No comments on this line yet.
+ 135 • Monitor your quota — check remaining GPU/TPU hours in the notebook Settings panel
No comments on this line yet.
+ 136
No comments on this line yet.
+
137
No comments on this line yet.
+ 138
No comments on this line yet.
+ 140
No comments on this line yet.
+ 141 Kaggle allows you to connect your GitHub repository to upload and sync notebooks, making it easy to version-control your work outside the platform.
No comments on this line yet.
+ 142
No comments on this line yet.
+
143
No comments on this line yet.
+ 144
No comments on this line yet.
+ 145 Sources:
No comments on this line yet.
+ 146 • Kaggle Notebooks
No comments on this line yet.
+ 147 • Kaggle Notebooks Documentation
No comments on this line yet.
+ 148 • Kaggle TPU Documentation
No comments on this line yet.
+ 149 • How to Use GPU in Kaggle — GeeksforGeeks
No comments on this line yet.
+ 150 • Kaggle GPU Resources Guide — Oreate AI
No comments on this line yet.
+ 151 • Kaggle 29GB RAM GPU Update
No comments on this line yet.
+ 152 • Gradient vs Kaggle Comparison — Paperspace
No comments on this line yet.
+ 153 • Free Cloud GPU Comparison 2026 — AI Multiple
No comments on this line yet.
+ 154 • Free GPU Cloud Trials 2026 — GMI Cloud
No comments on this line yet.
+ 155 • T4 vs P100 When to Choose — Medium
No comments on this line yet.