Task 4.2: Optimize and manage ML and AI infrastructure costs and performance.

• ALarger general purpose instances with more vCPUs

  Incorrect

  More CPU capacity can raise throughput, but general purpose CPUs are not built for deep learning inference, so cost per inference rarely drops this way.

• BBurstable instances, to pay less for idle time

  Incorrect

  Burstable instances provide a baseline CPU level with the ability to burst, for workloads with low-to-moderate CPU use. A high, steady inference volume needs sustained performance.

• CInf2 instances, based on AWS Inferentia2

  Correct

  Inferentia chips are designed for high performance at the lowest cost in Amazon EC2 for deep learning and generative AI inference, and Inf2 instances are built on Inferentia2. The Neuron SDK integrates with PyTorch to deploy models on them.

• DThe same instance type with SageMaker AI managed warm pools

  Incorrect

  Warm pools keep training infrastructure ready between training jobs. They do not apply to endpoint inference cost.

For steady, high-volume deep learning inference, purpose-built accelerators such as Inferentia are a main cost lever. Use load tests, for example with Inference Recommender, to confirm the choice.

• AA Compute Savings Plan

  Incorrect

  Compute Savings Plans apply to EC2 instance usage, Fargate and Lambda. They do not cover SageMaker AI instance usage.

• BA SageMaker AI Savings Plan

  Correct

  SageMaker AI Savings Plans offer up to 64% off On-Demand rates and apply automatically to SageMaker AI usage regardless of instance family, size, Region and component, such as notebooks or training.

• CAn EC2 Instance Savings Plan for the m5 family

  Incorrect

  EC2 Instance Savings Plans apply to one EC2 instance family in one Region. They do not cover SageMaker AI usage, and they would not follow a move to c5 or another Region.

• DManaged Spot capacity for every workload

  Incorrect

  Managed Spot training can cut training cost, but Spot capacity can be interrupted, which does not suit endpoints that must stay up.

Commitment discounts are service-specific: SageMaker AI Savings Plans for SageMaker AI usage, Compute and EC2 Instance Savings Plans for EC2 and related compute.

• AAn AWS Budgets budget action that applies an IAM policy or SCP

  Correct

  AWS Budgets can run an action when a budget exceeds a cost or usage threshold, such as applying an IAM policy or a service control policy that stops new resources from being provisioned.

• BCost allocation tags on every SageMaker AI resource

  Incorrect

  Activated cost allocation tags organize costs in Cost Explorer and reports. They describe spending but do not stop it.

• CAWS Cost Explorer forecasts, reviewed by the team each week

  Incorrect

  Cost Explorer shows and forecasts costs. A weekly review is manual and does not block anything when the limit is reached.

• DAWS Compute Optimizer recommendations for SageMaker AI

  Incorrect

  Compute Optimizer recommends better-sized resources. It does not enforce a spending cap.

To turn a budget into a control rather than a notification, attach a budget action that applies an IAM policy or SCP when the threshold is crossed.

• APrompt caching for the repeated policy document

  Correct

  Prompt caching lets supported models reuse long, repeated prompt prefixes across requests, which reduces input token cost and latency. Static content should come first in the prompt.

• BThe Priority service tier for the nightly classification job

  Incorrect

  Priority delivers faster responses for a price premium. The nightly job does not need speed, so this raises cost.

• CNo-commitment Provisioned Throughput for the assistant

  Incorrect

  Provisioned Throughput is billed hourly for the capacity you buy, whether or not requests arrive. It adds throughput; it does not lower the token cost of a repeated prompt.

• DMoving the policy document to the end of each prompt

  Incorrect

  Caching works on prompt prefixes, so static content belongs at the beginning. Moving it to the end makes reuse less likely.

• EA Bedrock batch inference job for the nightly classification

  Correct

  Batch inference processes many prompts asynchronously from S3, and Bedrock offers select models for batch inference at a 50% lower price than on-demand inference.

Token cost falls by sending less new input (prompt caching for repeated prefixes) and by paying less per token for work that can wait (batch inference).

• ARun an Amazon SageMaker Inference Recommender job on the model

  Correct

  Inference Recommender automates load testing across SageMaker AI instance types and configurations and recommends the deployment that delivers the best performance at the lowest cost.

• BDeploy on the largest GPU instance and scale down later

  Incorrect

  Starting with the largest instance guarantees high cost and still requires manual load testing to find the right size.

• CUse AWS Compute Optimizer recommendations before the first deployment

  Incorrect

  Compute Optimizer bases its recommendations on observed utilization, so it has nothing to analyze for a model that has never run.

• DCheck AWS Cost Explorer for the cheapest instance used last month

  Incorrect

  Past spending on other workloads says nothing about whether an instance meets this model's latency target.

Right-size before launch with benchmark data. Inference Recommender runs the load tests and compares instance types and configurations for you.

• ABuy a 6-month commitment term

  Correct

  Provisioned Throughput is billed hourly, and the longer the commitment term (no commitment, 1 month or 6 months), the more discounted the hourly price.

• BDelete the Provisioned Throughput each night

  Incorrect

  Traffic runs around the clock, so deleting the Provisioned Throughput each night would take the model offline while requests still arrive.

• CCut the number of model units in half

  Incorrect

  Model units set the throughput level, so halving them reduces capacity.

• DKeep no commitment and add a budget alert

  Incorrect

  A budget alert reports spend; it does not lower the hourly price.

For steady, long-lived Provisioned Throughput, a longer commitment term is the main price lever.

• AS3 Glacier Deep Archive for all of the datasets

  Incorrect

  Deep Archive objects must be restored before they can be read, which breaks the need for immediate reads.

• BS3 Intelligent-Tiering without the archive tiers

  Correct

  Intelligent-Tiering moves objects between frequent, infrequent and Archive Instant Access tiers based on access, without performance impact; the optional archive tiers are the ones that need a restore.

• CS3 Intelligent-Tiering with Deep Archive Access on

  Incorrect

  Objects in the Deep Archive Access tier must be restored first, so turning it on breaks the need for immediate reads.

• DS3 Standard with the files compressed

  Incorrect

  Compression can help, but data that is rarely read still stays in the most expensive storage class.

Use Intelligent-Tiering for unknown or changing access patterns, and turn on its archive tiers only when a restore delay is acceptable.

• AA cross-Region inference profile

  Incorrect

  Cross-Region inference spreads requests for one model across Regions; it does not choose between models.

• BDistillation into a new, smaller model

  Incorrect

  Distillation trains a new model, and the team would still need its own logic to decide which model gets each request.

• CAmazon Bedrock intelligent prompt routing

  Correct

  Intelligent prompt routing provides a single serverless endpoint that routes requests between models in the same family, predicting the response quality of each to optimize for quality and cost.

• DProvisioned Throughput for the large model

  Incorrect

  Provisioned Throughput adds capacity for one model at a fixed cost; every request still goes to the large model.

Routing simple requests to smaller models in the family cuts cost while keeping quality where it is needed.

• AAWS Budgets with a fixed monthly amount

  Incorrect

  Budgets alert against amounts you set, which is the fixed-threshold approach the team wants to avoid.

• BCloudWatch alarms on each job metric

  Incorrect

  Job metrics show resource use, not spend, and per-job alarms would miss new resources.

• CCost allocation tags on all resources

  Incorrect

  Tags organize costs in reports; they do not raise alerts on their own.

• DAWS Cost Anomaly Detection

  Correct

  Cost Anomaly Detection uses machine learning models to detect and alert on anomalous spend, and ranks root causes by service, account, Region and usage type.

Combine budgets for planned limits with anomaly detection for surprises such as runaway jobs.

• ARebuild the index with 256-dimension embeddings

  Correct

  Titan Text Embeddings V2 can output 1,024, 512 or 256 dimensions, and the vector field must match the model's output, so the data must be embedded again at 256 into a matching index.

• BCut the stored vectors down to their first 256 values

  Incorrect

  The index is set up for the configured number of dimensions; the documented way to get smaller vectors is to request a smaller output size from the model.

• CReturn more chunks for each query

  Incorrect

  The number of results changes retrieval, not the size of the stored vectors.

• DStore the vectors as text in the metadata field

  Incorrect

  Vectors belong in the vector field; storing them as text would break similarity search.

Smaller embedding dimensions cut vector storage and search cost; test retrieval quality before you switch.