🐾SageMaker Studio: cheaper than you think🐾

🤓 I frequently hear the opinion that SageMaker is too expensive for the ML team to use. Let’s dive deeper into charging policies and discover how much you will pay for your ML workload. Little spoiler — you're in for a pleasant surprise.

SageMaker Studio domain

Let’s start with SageMaker Studio domain creation. While creating a domain, you should consider which network option to choose:

• No customer VPC is attached.

• Customer VPC is attached with internet access enabled via NAT

• Customer VPC is attached without internet access enabled

In the first two cases, you won’t have any additional network charges. For the third option, you should be aware of the cost of interface endpoints you need to attach to VPC’s subnets. You need at least two endpoints: com.amazonaws.region.sagemaker.api and com.amazonaws.region.sagemaker.runtime.

After your domain is created, you will be charged for the EFS volume attached to it. There are no other charges applied. The default app for each user doesn't incur any cost as well.

SageMaker Studio notebooks and Code Editor

While working in SageMaker Studio you use Notebooks and launch kernels for them. Kernels are charged based on the On-demand price of the chosen instance. In case you are afraid that the ML team can forget to shut down the instance after work is finished, you can use the auto shutdown extension.

✅ Leverage SageMaker Pipeline steps to run training and preprocessing steps on larger instances, but choose smaller ones for running notebooks.

SageMaker Model serving

SageMaker has several options for model serving:

• Real-time inference

• Asynchronous inference

• Serverless inference

• Batch Transform jobs

Real-time and asynchronous endpoints are charged based on the time endpoint is running, even if it doesn’t make predictions. Serverless endpoints and Batch Transform jobs are charged only for the computing time. All prices are based on the On-Demand price of the chosen instance.

Other SageMaker services

There are some other services such as Data Wrangler, JumpStart are charged based on On-Demand price of the chosen instance.

⚠️ Don’t confuse the price of m5.xlarge with the price of ml.m5.xlarge in your calculations. You can refer to the SageMaker pricing for calculation.

🦊 The right choice of resources will make SageMaker Studio one of your best investments.

---

Thank you for reading, let’s chat 💬

💬 What do you think about SageMaker Studio as a development environment?
💬 Which tools do you use for SageMaker cost optimisation?
💬 Do you use Code Whisperer extension in notebooks?

I love hearing from readers 🫶🏻 Please feel free to drop comments, questions, and opinions below👇🏻