AWS - SageMaker GPU Setup

1. Purpose

   • To use GPU on SageMaker Notebook.

2. Solution

   1. Request Amazon to open 2 restrictions,

      Limit type	Region	Resorce Type	Limit	New limit value
      SageMaker	(your region)	SageMaker Training	ml.p2.xlarge instance	1
      SageMaker	(your region)	SageMaker Hosting	ml.p2.xlarge instance	1

      • From AWS Support page, create case to issue above.

   2. Re-make Notebook instance on ml.p2.xlarge.

3. Confirmation

   • Type the codes in Notebook cell.

     ```python import torch

     print(torch.cuda.is_available())# True if available. print(torch.cuda.get_device_name(0))# Tesla K80 etc. ```

4. Costs / Notes

   • Take care that ml.p2.xlarge costs some, for Notebook Instance ml.p2.xlarge per hour.

   • Notebook Instances which I have used.

     Instance Type	vCPU	GPU	Memory(GiB)	GPU Memory(GiB)	Network Performance	Note
     ml.t2.medium	2	-	4	-	Low-Middle	SageMaker with CPU.
     ml.p2.xlarge	4	1xK80	61	12	High	SageMaker with GPU. This costs by hourly fee.
     ml.m4.xlarge	4	-	16	-	High	Deploy the model.