4.2.6

Question	Answer & Explanation
### Question 1 - Purchasing Options for SageMaker Inference An e-commerce company runs a stable, year-round fraud detection model on Amazon SageMaker AI real-time endpoints. The baseline traffic is predictable and consistent 24/7. The ML engineering team wants to implement a purchasing option that provides the most cost-effective solution for this steady baseline usage. Which purchasing option should the team select to achieve this? Select ONE - A) Compute Savings Plans - B) EC2 Instance Savings Plans - C) SageMaker AI Savings Plans - D) Reserved Instances for Amazon EC2	Correct Answer: C Correct Answer Analysis: - C is Correct: SageMaker AI Savings Plans offer a flexible, commitment-based pricing model designed specifically for SageMaker ML instance usage. By committing to a consistent hourly spend over a 1-year or 3-year term, the team receives a significant discount (up to 64%) on the steady baseline inference usage compared to On-Demand pricing. Distractor Analysis: - A is Incorrect: Compute Savings Plans apply to Amazon EC2, AWS Fargate, and AWS Lambda, but they do not extend to Amazon SageMaker ML instance usage. - B is Incorrect: EC2 Instance Savings Plans apply exclusively to Amazon EC2 instances within a specific instance family and region, and cannot be applied to SageMaker endpoints (ml.* instance types). - D is Incorrect: Reserved Instances are an EC2 pricing feature and are not applicable to managed SageMaker endpoints. > Key Takeaway: For steady, predictable SageMaker ML workloads (training, inference, notebooks), SageMaker AI Savings Plans provide the correct commitment-based discount, whereas Compute Savings Plans do not cover SageMaker.
### Question 2 - Cost Optimization for High-Volume Bedrock Traffic A media streaming company uses an Amazon Bedrock foundation model to generate daily content summaries. The traffic is highly predictable, with a constant, high volume of requests running 24/7. The team is currently using On-Demand token pricing but wants to optimize infrastructure costs for this predictable workload. Which purchasing option provides the most cost-effective solution? Select ONE - A) Enable Amazon Bedrock Provisioned Throughput - B) Switch to Bedrock On-Demand with a smaller model - C) Purchase SageMaker AI Savings Plans - D) Use Spot Instance pricing for inference	Correct Answer: A Correct Answer Analysis: - A is Correct: For predictable, high-volume foundation model inference, Amazon Bedrock Provisioned Throughput allows you to reserve model units at a fixed hourly rate. This is more cost-effective than On-Demand token pricing for workloads with steady, continuous usage. Distractor Analysis: - B is Incorrect: While choosing a smaller model can reduce per-token costs, On-Demand pricing is still more expensive than Provisioned Throughput for a steady, high-volume 24/7 workload. The question asks for the best purchasing option. - C is Incorrect: SageMaker AI Savings Plans apply to SageMaker ML instances, not to Amazon Bedrock foundation model inference. - D is Incorrect: Spot Instance pricing applies to EC2 and SageMaker compute instances, not to serverless token-based inference on Bedrock. > Key Takeaway: In the Generative AI world, "baseline commitment" translates from SageMaker Savings Plans to Amazon Bedrock Provisioned Throughput, reserving model units at a fixed rate for predictable traffic.
### Question 3 - Purchasing Options for Interruptible Training Jobs A deep learning research team runs distributed training jobs on Amazon SageMaker that take several days to complete. The jobs can be checkpointed and safely restarted if interrupted. The team requires the absolute lowest cost for these training instances and can tolerate interruptions. Which purchasing option should the team use to optimize costs? Select ONE - A) SageMaker AI Savings Plans - B) Amazon Bedrock Provisioned Throughput - C) SageMaker Managed Spot Training - D) EC2 Instance Savings Plans	Correct Answer: C Correct Answer Analysis: - C is Correct: SageMaker Managed Spot Training uses spare EC2 capacity at a heavily discounted rate. Because the workload is interruptible and can be checkpointed, this is the most cost-effective purchasing option for this specific scenario. Distractor Analysis: - A is Incorrect: SageMaker AI Savings Plans require a 1-year or 3-year commitment to a steady hourly spend. They are not suited for sporadic, interruptible research jobs that only run for a few days. - B is Incorrect: Bedrock Provisioned Throughput is for foundation model inference, not for custom training jobs on SageMaker. - D is Incorrect: EC2 Instance Savings Plans require a long-term commitment and apply to EC2 instances, not directly to SageMaker Managed Spot Training. > Key Takeaway: For fault-tolerant, interruptible workloads like checkpointable training jobs, Managed Spot Training (spare capacity) provides the lowest cost without requiring a long-term commitment.
### Question 4 - Scope of SageMaker AI Savings Plans An enterprise is migrating all of its machine learning workloads to Amazon SageMaker AI. This includes Amazon SageMaker Studio notebooks for data preparation, Batch Transform jobs for feature engineering, and Real-Time Inference endpoints for production models. The team wants a single commitment-based purchasing option that will apply discounts across all of these SageMaker compute usages. Which option meets this requirement? Select ONE - A) Compute Savings Plans - B) SageMaker AI Savings Plans - C) EC2 Instance Savings Plans - D) AWS Free Tier	Correct Answer: B Correct Answer Analysis: - B is Correct: SageMaker AI Savings Plans apply automatically to all SageMaker ML instance usage, including SageMaker Studio notebooks, Batch Transform, Processing, Training, and Real-Time Inference endpoints. Distractor Analysis: - A is Incorrect: Compute Savings Plans apply to EC2, Fargate, and Lambda, but do not cover SageMaker instances. - C is Incorrect: EC2 Instance Savings Plans are restricted to specific EC2 instance families and do not apply to SageMaker workloads. - D is Incorrect: The AWS Free Tier provides short-term free trials, not a commitment-based discount for steady enterprise workloads. > Key Takeaway: SageMaker AI Savings Plans provide a flexible discount across the entire SageMaker suite (training, inference, notebooks, processing) as long as the compute is billed as SageMaker usage.
### Question 5 - Variable Traffic Purchasing Options on Bedrock A startup is building an AI-powered customer service agent using Amazon Bedrock. The application is in the beta phase, and traffic is highly variable, driven by unpredictable viral marketing campaigns. The team wants the most cost-effective purchasing model that avoids paying for idle capacity during quiet periods. Which purchasing option should the team select? Select ONE - A) Amazon Bedrock Provisioned Throughput - B) SageMaker AI Savings Plans - C) Amazon Bedrock On-Demand - D) Compute Savings Plans	Correct Answer: C Correct Answer Analysis: - C is Correct: Bedrock On-Demand pricing bills per token for input and output. It requires no upfront commitment and is ideal for variable or experimental traffic where you do not want to pay for idle capacity. Distractor Analysis: - A is Incorrect: Provisioned Throughput reserves model units at a fixed hourly rate. For highly variable traffic, this would result in paying for idle capacity during quiet periods, making it less cost-effective. - B is Incorrect: SageMaker AI Savings Plans are for SageMaker compute instances, not Bedrock token-based inference. - D is Incorrect: Compute Savings Plans are for EC2, Fargate, and Lambda, not Bedrock. > Key Takeaway: For variable, unpredictable Generative AI inference traffic, stay on Bedrock On-Demand to avoid paying for idle reserved model units. Use Provisioned Throughput only for steady, predictable baselines.
### Question 6 - Commitment Duration and Instance Flexibility A company is running a steady inference baseline on Amazon SageMaker AI. They want to purchase a SageMaker AI Savings Plan to optimize costs. However, they plan to re-architect their ML models and transition to a different instance family in 10 months. They need a commitment that provides a discount now but allows them to change instance families without penalty. Which action should the team take? Select ONE - A) Purchase a 3-year SageMaker Savings Plan for the maximum discount, as it automatically covers all instance families. - B) Purchase a 1-year SageMaker Savings Plan, as it provides the flexibility to change instance families while offering a discount for the 10-month period. - C) Purchase an EC2 Instance Savings Plan restricted to the current instance family. - D) Stay on On-Demand pricing for the 10 months to maintain maximum flexibility.	Correct Answer: B Correct Answer Analysis: - B is Correct: A 1-year SageMaker AI Savings Plan provides a discounted rate compared to On-Demand and automatically applies to any SageMaker ML instance family, allowing the team to re-architect and change instances within the 10-month window without penalty. Distractor Analysis: - A is Incorrect: A 3-year commitment locks the team into paying the Savings Plan rate for 3 years, which is inefficient if they re-architect in 10 months. - C is Incorrect: EC2 Instance Savings Plans lock you into a specific EC2 instance family and do not apply to SageMaker instances. - D is Incorrect: Staying on On-Demand forfeits the available cost savings for the steady baseline they are currently running. > Key Takeaway: SageMaker AI Savings Plans offer instance family flexibility. Choose a 1-year term for a steady workload with near-term architectural changes to balance discounts and flexibility.
### Question 7 - Spot Instances for Real-Time Inference An e-commerce platform hosts a real-time fraud detection model on a SageMaker endpoint. The business mandates 99.99% availability and guarantees sub-100ms latency for all transactions, especially during peak shopping events like Black Friday. To save costs, a junior engineer suggests using SageMaker Spot endpoints for this real-time inference workload. Why is this suggestion invalid? Select ONE - A) Spot capacity can be reclaimed by AWS with a two-minute warning, which violates the strict availability and latency requirements for critical real-time inference. - B) Spot Instances do not support GPU instance types, which are required for the model. - C) Spot Instances can only be used for Batch Transform jobs, not real-time endpoints. - D) Spot Instances require a 3-year upfront commitment, defeating the purpose of cost savings.	Correct Answer: A Correct Answer Analysis: - A is Correct: Spot Instances use spare AWS capacity, which can be reclaimed by AWS with only a two-minute warning. For a critical, real-time fraud detection endpoint requiring 99.99% availability and low latency, capacity reclamation would cause unacceptable service disruptions. Distractor Analysis: - B is Incorrect: Spot Instances do support GPU instance types (e.g., ml.p3, ml.g4dn). - C is Incorrect: While commonly used for training and batch transform, Spot can be used for real-time endpoints (with the risk of interruption), but it is a bad architectural choice for critical workloads. - D is Incorrect: Spot Instances are On-Demand spare capacity and do not require any upfront commitment. > Key Takeaway: Never use Spot capacity for real-time, critical inference endpoints that require high availability and low latency, as AWS can reclaim the capacity with minimal warning.
### Question 8 - Consolidating Inference Endpoints A data science team has deployed 50 small, distinct PyTorch models to production. Each model is hosted on its own ml.c5.large SageMaker real-time endpoint. Each model receives only a few inference requests per day. The company wants to optimize infrastructure costs while maintaining the PyTorch serving framework. Which approach provides the MOST cost-effective solution? Select ONE - A) Purchase a SageMaker AI Savings Plan covering the 50 ml.c5.large instances. - B) Convert the endpoints to use SageMaker Spot capacity. - C) Consolidate the models onto a single SageMaker Multi-Model Endpoint (MME) using shared compute instances. - D) Convert the real-time endpoints to Asynchronous Inference endpoints.	Correct Answer: C Correct Answer Analysis: - C is Correct: SageMaker Multi-Model Endpoints (MMEs) allow you to deploy multiple models to a single shared endpoint, loading them into memory on-demand. Since these 50 models share the same framework and are infrequently accessed, consolidating them onto one shared instance pool drastically reduces idle compute costs. Distractor Analysis: - A is Incorrect: While a Savings Plan provides a discount, you are still paying for 50 provisioned instances that sit mostly idle. Consolidating the compute is a much larger cost saving. - B is Incorrect: Spot capacity is not suitable for endpoints (risk of reclamation) and does not solve the underlying problem of fragmented, idle compute. - D is Incorrect: Asynchronous Inference is for large payloads or long processing times, not for consolidating multiple small models. > Key Takeaway: When managing many small, infrequently accessed models sharing the same framework, Multi-Model Endpoints (MMEs) eliminate idle compute costs by sharing a single instance pool, offering massive savings over dedicated endpoints.
### Question 9 - Billing Behavior of SageMaker Savings Plans A company has purchased a 1-year SageMaker AI Savings Plan with a commitment of $100/hour. During a typical hour, their SageMaker training and inference workloads accrue $80/hour of On-Demand equivalent usage. However, during a peak end-of-month processing window, the workloads spike to $150/hour. How will the usage be billed during the peak processing window? Select ONE - A) The entire $150/hour will be billed at the On-Demand rate because the commitment was exceeded. - B) $100/hour will be billed at the discounted Savings Plan rate, and the remaining $50/hour will be billed at the On-Demand rate. - C) The entire $150/hour will be billed at the discounted Savings Plan rate, and the plan will auto-expand. - D) The Savings Plan will be suspended, and a penalty rate will be applied to the $150/hour.	Correct Answer: B Correct Answer Analysis: - B is Correct: SageMaker AI Savings Plans apply a discount to the committed \(100/hour of usage. Any usage beyond the committed amount (\)50/hour in this case) is billed at the standard On-Demand rate. The plan does not automatically expand or cap. Distractor Analysis: - A is Incorrect: The Savings Plan commitment is utilized first. You do not lose the discount on the committed amount just because you exceeded it. - C is Incorrect: Savings Plans do not auto-expand to cover usage beyond the committed hourly spend with the discounted rate; excess usage reverts to On-Demand. - D is Incorrect: There are no penalty rates for exceeding a Savings Plan commitment. > Key Takeaway: Savings Plans provide a discounted rate up to your committed hourly spend. Any usage exceeding that commitment is billed at the standard On-Demand rate.
### Question 10 - Asynchronous Batch Workloads on Bedrock A financial services company needs to process 5 million historical loan documents through a large foundation model on Amazon Bedrock to extract structured entities. The workload is asynchronous, non-urgent, and runs once a quarter. The team wants the most cost-effective way to execute this inference without maintaining idle provisioned capacity year-round. Which purchasing option is the best fit for this workload? Select ONE - A) Purchase a 1-year Bedrock Provisioned Throughput commitment. - B) Use Amazon Bedrock On-Demand pricing for the quarterly token usage. - C) Host the model on SageMaker using Spot Instances. - D) Purchase a 1-year SageMaker AI Savings Plan.	Correct Answer: B Correct Answer Analysis: - B is Correct: For an asynchronous, intermittent workload that runs infrequently (once a quarter), Bedrock On-Demand token pricing is the most cost-effective choice. It avoids the high fixed costs of Provisioned Throughput, which would sit idle for the rest of the year. Distractor Analysis: - A is Incorrect: Provisioned Throughput requires a fixed hourly commitment. Paying for this 24/7/365 for a quarterly workload would be significantly more expensive than On-Demand. - C is Incorrect: You cannot natively host a Bedrock foundation model on SageMaker Spot Instances to avoid Bedrock token billing. - D is Incorrect: SageMaker Savings Plans do not apply to Amazon Bedrock token-based inference. > Key Takeaway: For infrequent, asynchronous Generative AI workloads, stick to On-Demand token pricing. Provisioned Throughput is only cost-effective for steady, continuous baselines.