Inference cost and capacity: MLA-C02 study notes and practice questions
What to know about Inference cost and capacity for the MLA-C02 exam, from the Operating, Monitoring, and Securing ML and AI Solutions domain (24% of the exam guide), then 2 free questions to test it. Part of the MLA-C02 study guide.
A dental chain serves an X-ray triage image model on GPU instances in SageMaker AI. Traffic is steady, and the model is a standard convolutional network trained and validated in TensorFlow. The finance team wants the inference cost cut without rewriting the model.
Inference latency must stay within the current service target, and the team is willing to compile the model for new target hardware as a one-time step.
Which change best meets these requirements?
Choose one answer.
AA. Move the endpoint to a larger GPU instance type, so that each request finishes sooner and the endpoint is billed for fewer instance hours.
BB. Replace the model with a rule-based image threshold that requires no model at all, since rules never cost money to serve.
CC. Turn off all logging and monitoring on the endpoint, which removes the cost of the logs without changing the model.
DD. Compile the model with SageMaker Neo for an AWS Inferentia instance type, and deploy the compiled model to an Inferentia-based endpoint that meets the latency target.
Show the answer and explanation
Answer: D
A. Wrong: A larger instance costs more per hour, and the total hours do not necessarily fall for a steady workload. The option raises cost rather than cutting it.
B. Wrong: Rules may not match the accuracy of the model for X-ray triage, and the stem asks to keep the model without rewriting it. The option changes the service.
C. Wrong: Logging is a small share of the cost, and removing monitoring weakens the operations the team relies on. The option does not address the instance cost that drives the bill.
Inferentia instances provide lower-cost inference for models that compile for them, and SageMaker Neo compiles a validated model for that hardware in one step. A larger GPU, a rule-based replacement or turning off monitoring each raise cost, change the service or weaken operations.
A solar installer runs two kinds of ML workloads on AWS. Nightly model training jobs run for about four hours each and can be restarted if interrupted. A real-time forecasting endpoint runs all day at a steady load on the same instance type for the whole year.
The finance team wants lower cost for both workloads without changing the models, the instance families they are built for, or the forecasting service targets.
Which TWO actions meet these requirements? (Choose two.)
Choose two answers.
AA. Use managed Spot training for the nightly jobs, with checkpoints saved to Amazon S3 so that an interrupted job resumes.
BB. Purchase a SageMaker AI Savings Plan for the steady, all-year inference usage, so the hourly rate for the endpoint is lower.
CC. Move the nightly jobs to On-Demand capacity with larger instance types, so that each job finishes in one hour and costs less overall.
DD. Reduce the endpoint's minimum instance count to zero during the day, so the endpoint stops running when traffic is low.
EE. Turn off the training jobs entirely and use the previous quarter's model for all forecasts, which removes the training cost.
B. Right: Correct. Savings Plans give a lower rate in exchange for a steady commitment, which fits an endpoint that runs all year on the same type. The endpoint and its instance type do not change. Learn more: What are Savings Plans? (opens in a new tab)
C. Wrong: Larger On-Demand instances cost more per hour, and the stem asks for lower cost. The option pays more for the same training work.
D. Wrong: The stem describes steady traffic all day, so scaling to zero would break the forecasting service targets. The option conflicts with the stated requirement.
E. Wrong: The stem requires the models to keep their current training. Skipping retraining would also let the forecasts drift, which the team does not want.
Explanation
Managed Spot training with checkpoints lowers the cost of restartable nightly training, and a Savings Plan lowers the rate for steady, all-year endpoint usage. Larger On-Demand instances, scaling to zero during steady traffic and skipping retraining each raise cost or break a stated requirement.
Want more on Inference cost and capacity? The 5 full-length practice sets have 13 more Inference cost and capacity questions, 325 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →