MLS-C01 Question 325
Single answerYou are building a machine learning model that requires high-performance GPU resources for training. The training process is expected to run for several days, and cost optimization is a priority. Which instance type and purchasing option should you choose to minimize costs while ensuring the training completes successfully?
- A
On-demand p3.2xlarge instances
- B
Spot p3.2xlarge instances with checkpointing enabled
- C
Reserved p3.2xlarge instances for a 1-year term
- D
On-demand g4dn.xlarge instances
Show answer and explanation
Correct answer: B
Explanation
Spot instances are the most cost-effective option for GPU-intensive machine learning training workloads. Enabling checkpointing ensures that the training process can resume from the last saved state in case of instance interruptions, making it a reliable and economical choice for this scenario.
- A. Incorrect.
On-demand p3.2xlarge instances are suitable for GPU-intensive workloads, but they are the most expensive option and not cost-efficient for multi-day training.
- B. Correct.
Spot p3.2xlarge instances are highly cost-effective for training as they are up to 90% cheaper than on-demand instances. By enabling checkpointing, you can safely resume training if the instance is interrupted, ensuring successful completion of the job.
- C. Incorrect.
Reserved p3.2xlarge instances offer savings over on-demand instances, but they require a commitment of at least 1 year. This option is not ideal for a one-time, multi-day training task.
- D. Incorrect.
On-demand g4dn.xlarge instances are less expensive than p3 instances, but they provide lower GPU performance, which might significantly increase training time and overall costs for a resource-intensive task.