GPU Resources for Machine Learning

Table of Contents

Free Resources

University Laboratory

  • Pros:
    • It’d be the easiest and most flexible way to use.
    • There’d be a manual, member’s support or something in laboratory.
  • Cons:
    • The GPU quality depends on the laboratory’s budget.
    • You may need to wait until resources become available during busy season.
    • The project would be less reproducible. Documentation must be carefully prepared when source code is to be public.

Google’s TPU Research Cloud

TPU Research Cloud offers cloud machine resources to researchers for free.

  • Pros:
    • Tensor Processing Unit (TPU) is available free of charge.
  • Cons:
    • Free 30-days access to TPU requires application and approval.
    • TPUs are free, but other Google Cloud Platform services such as storage for dataset are charged.
    • Pytorch (by Meta) can be used, but it is said that TensorFlow (by Google) is better than it, because tensorflow natively supports TPUs.

Google Colab (Free Plan)

Google Colab.

  • Pros:
    • High reproducibility of your project.
  • Cons:
    • Free plan has poor machine power and time limits.

My Own GPU

  • Pros:
    • It’s free to use except for electricity once you purchase a GPU.
    • Always available.
  • Cons:
    • GPUs are easily outdated because GPUs evolve so quickly.
    • Low GPU utilization is less cost-effective than on-demand use of cloud resources.
    • GPU can be broken.
    • Low reproducibility of your source code.

Cloud Services

Consider costs.