Online Compute Resource Tips

Introduction

This page collects various resources to help mainly ML/DL practitioners make more efficient use of available online compute resources for various projects.

Contents

Free Online Compute

Free online compute resources do not cost anything and do not require you to add in payment details. They also usually come with paid upgrades for more resources.

Google Colaboratory gives you free access to custom cloud-based Jupyter notebook environment with pre-installed ML/DL packages.

  • 2 vCPU + 12GB RAM
  • GPU (K80, P100, T4, V100 possible)
  • Cloud TPU v2-8
  • Up to 12h session (preemptible)

Paperspace Gradient lets you run public notebooks on free GPU instances and select from a variety of preconfigured environments.

  • 8 vCPU + 30GB RAM
  • GPU (M4000)
  • Up to 6h session

Kaggle gives you free access to custom cloud-based Jupyter notebook environment with pre-installed ML/DL packages.

  • 2 vCPU + 15GB RAM
  • GPU (P100)
  • TPU v3-8
  • 30h or more, resets per week (subject to demand), up to 9h session

BlazingSQL Notebooks gives you free access to T4 GPU instances with access to the JupyterLab interface. The GPU is shared which means multiple users might be using the same GPU.

  • 4 vCPU + 15GB RAM
  • GPU (T4) – shared
  • Session terminates after you disconnect

Coiled.io gives you free access to Dask clusters with access to the JupyterLab interface. Dask allows you to scale numerical computation (Numpy etc.) and data science (Pandas, Scikit-learn etc.) workloads to many CPU cores.

  • 100 vCPU Dask cluster
  • Session terminates after you disconnect

Cheap Online Compute

Cheap online compute resources here are gathered to help facilitate DL work at a cost of less than USD$0.50 per hour of usage with flexible billing options. This list is organised by the type of compute source along with the recommended type of usage.

NVIDIA T4 GPU

8 FP32 TFLOPS
65 FP16 TLFOPS
16GB 300GB/s DRAM

Training Performance:
400+ img/s RN50
10+ seq/s BERTLARGE

NVIDIA T4 GPUs are the most cost effective GPU resource for training DL models that can take advantage of mixed precision (mixed FP32+FP16) training.

  • Google Cloud Platform
    n1-standard-4 (4 vCPU + 16GB RAM) + T4 GPU
    Preemptible: USD$0.16/h
    On-demand: USD$0.38/h
  • AWS EC2
    g4dn.xlarge (4 vCPU + 16GB RAM)
    Spot: USD$0.16/h
Footnotes
  • T4 Benchmark performance shown is obtained from this page.