As businesses and organizations continue to leverage the power of artificial intelligence and machine learning, cloud service providers like Amazon Web Services (AWS) are relentlessly innovating and introducing new services to support these modern needs. In this post we'll be pulling back the curtain on two of AWS´s centerpieces AI/ML services: the preview-stage Amazon Bedrock and the well-established Amazon SageMaker.
We're going to pit these two contenders against each other, in a friendly way of course. We’ll weigh them up, focusing on the heavy hitters: general differences, data protection and security, setup efforts, customizability, and potential use cases.
It's only fair to tell you that Amazon Bedrock is not yet available to the masses. It's currently only presenting itself to a select audience of customers and partners. But don't worry, we've gathered enough information to send it into the ring with SageMaker.
General Differences
Amazon Bedrock
Amazon Bedrock is a fully managed service that provides access to pre-trained foundation models from Amazon/AWS and prominent AI startups. Though not yet generally available, Bedrock promises a number of impressive features. Bedrock eliminates the need to manage the underlying infrastructure and integrates seamlessly into the AWS service landscape.
Amazon Bedrock features summarized:
Diverse Foundation Models: Amazon Bedrock provides access to a range of pre-trained foundation models, offering a significant multifunction advantage compared to other AI services that usually come with only text foundation models. Some of these models include Amazon Titan for text summarization and generation, Jurassic-2 for multilingual text generation, Claude 2 for thoughtful dialogue and content creation, Command and Embed for text generation for business applications, and Stable Diffusion for image generation.
Agents for Amazon Bedrock: Agents allow developers to build generative AI systems capable of incorporating data and integrating with internal APIs via Foundation Models (FMs), privately and securely, without the need to train the models. The FMs can complete complex tasks, in addition to generating text or chatting.
Serverless Experience: Bedrock is serverless, meaning there are no servers or infrastructure to manage. Users only need to interact with a simple API: provide an input and specify the model to use, and Bedrock will provide an output. It's as simple as that.
Easy Model Customization: Bedrock allows easy model customization through fine-tuning. Customers only need to point Bedrock to a few labeled examples in an S3 bucket, and the service can fine-tune the model for a specific task. You just need to provide a couple of dozen prompt/response examples, and you're good to go.
Data Privacy: None of the user data is used to train the underlying foundational models. All data is encrypted and does not leave the user's Virtual Private Cloud (VPC). This feature is a significant milestone for the real-world, production use of foundational models.
Service Integration: Bedrock integrates seamlessly with other AWS services, including SageMaker Experiments (for testing different models) and Pipelines (for managing Foundation Models at scale).
Amazon SageMaker
On the other hand, Amazon SageMaker is a comprehensive service that allows data scientists and developers to build, train, and deploy machine learning models for extended use cases. SageMaker supports the complete machine learning lifecycle, providing tools to label and prepare your data, choose an algorithm, train the model, tune and optimize it for deployment, make predictions, and take action.
Amazon SageMaker comes with a capability called "JumpStart" that essentially does exactly what it says. It's a jump start that speeds up machine learning projects by providing pre-built solutions and trained models, making it easier for users to get their projects off the ground. This feature is a blessing for developers who want to get started quickly without having to reinvent the wheel.
Amazon SageMaker features summarized:
Comprehensive Machine Learning Lifecycle Support: Amazon SageMaker supports every step of the machine learning process, from data labeling and preparation to model deployment and monitoring.
Built-in Algorithms and Frameworks: Amazon SageMaker provides various built-in algorithms and frameworks, allowing users to select the most suitable one for their needs without requiring them to build everything from scratch.
Automatic Model Tuning: Amazon SageMaker automatically tunes models by adjusting thousands of different combinations of algorithm parameters, leading to the most optimal model.
Training and Inference with SageMaker Studio: The SageMaker Studio provides a single, web-based visual interface where you can perform all ML development steps, making it easier to build, train, and tune machine learning models.
Managed Spot Training: Amazon SageMaker Managed Spot Training allows users to use Amazon EC2 Spot instances for training ML models, resulting in cost savings of up to 90% compared to on-demand instances.
Data Protection and Security Requirements
Both Bedrock and SageMaker offer robust security features inherent in the AWS ecosystem. However, the services differ in terms of data management as their underlying architectures bring significant differences that could potentially impact your use case.
With Amazon SageMaker, customers have complete control over their data and the underlying infrastructure. This means that users have full authority over where the data is processed. They have the ability to encrypt data both at rest and in transit, manage data access through Identity and Access Management (IAM) roles, and comply with regulations through AWS’s robust compliance offerings. Users can also use SageMaker in their Virtual Private Cloud (VPC) to have network level control. For customers with stringent data security requirements, this level of control is paramount.
On the other hand, Amazon Bedrock, being a managed service, processes data within the confines of the AWS environment. While this means users won't have to worry about infrastructure management, it also means that they have less direct control over where the data is processed. Nonetheless, Bedrock ensures that none of the user data is used to train the underlying foundational models and all data is encrypted and does not leave the user's VPC. However, organizations with extremely high-security requirements might need to evaluate these considerations carefully.
While both services provide robust security features, the decision will likely come down to how much direct control over data and infrastructure your organization requires.
Effort to Setup
Bedrock's setup, being fully managed, is expected to require less effort compared to SageMaker. Users simply select an appropriate pre-trained foundation model, customize it with their data, and start using it.
Conversely, despite its powered experience, SageMaker requires more setup effort due to its large feature set. Users have to prepare data, select or create an algorithm, train the model, and then deploy it. In addition, using SageMaker effectively requires more technical experience and additional infrastructure management compared to Bedrock.
Customizability
SageMaker excels in customizability as it allows you to use your own algorithms, built-in ones, or select from many available on the AWS Marketplace or third party. You can fine-tune these models to meet specific requirements. This means that requirements delving deeper into detail are better suited with SageMaker, as you can use any available open-source Large Language Model (LLM) and train it to your preference.
In contrast, Bedrock's customizability appears less flexible than SageMaker. While it allows users to customize foundation models with their data, Bedrock only comes with default Foundation Language Models, although with fine-tuning capabilities.
Use Cases
Bedrock is ideal for scenarios where organizations need advanced AI capabilities quickly without having to deal with building models or managing infrastructure. Use cases include creative content generation, dialog system creation, text summarization, multilingual text creation, and advanced image generation tasks.
SageMaker is suitable for a broader range of machine learning tasks that require detailed control over the process of model creation, training and deployment. It is ideal for predictive analytics, recommendation systems, anomaly detection, or any other task that requires a customized machine learning model.
Costs
When it comes to cost, Amazon Bedrock could potentially have an advantage over Amazon SageMaker. Since Bedrock is a fully managed service, it abstracts away much of the infrastructure management, potentially reducing operational costs. Additionally, given that it's serverless, you only pay for the resources you actually use. This contrasts with SageMaker, where you may need to maintain dedicated resources even when not in use.
Specifically, Amazon Bedrock's pricing model is expected to be based on usage, with charges for the number of tokens processed by the foundational models and the compute time used for fine-tuning. This means that the cost directly scales with your needs.
Conclusion
Both Amazon Bedrock and Amazon SageMaker are powerful tools in the AWS AI/ML service landscape. Choosing between the two depends on your specific needs. If you need to quickly integrate advanced AI capabilities without much customization, Bedrock is the way to go. But if your use case demands deep customizability and you're willing to invest more effort into setting up and managing the model training process, SageMaker would be a better fit.
Top comments (0)