Exam AWS Certified Machine Learning - Specialty topic 1 question 18 discussion

https://docs.aws.amazon.com/sagemaker/latest/dg/sagemaker-dg.pdf page 55: If you plan to use GPU devices, make sure that your containers are nvidia-docker compatible. Only the CUDA toolkit should be included on containers. Don't bundle NVIDIA drivers with the image. For more information about nvidia-docker, see NVIDIA/nvidia-docker. So the answer is B

Answer is B. below is from AWS documentation, If you plan to use GPU devices for model training, make sure that your containers are nvidia-docker compatible. Only the CUDA toolkit should be included on containers; don't bundle NVIDIA drivers with the image. https://docs.aws.amazon.com/sagemaker/latest/dg/your-algorithms-training-algo-dockerfile.html

When using Amazon SageMaker with GPU-based EC2 instances (e.g., P3 instances), you must ensure that your custom Docker container can leverage NVIDIA GPUs. NVIDIA-Docker (now part of Docker with nvidia-container-runtime) allows containers to access GPU resources without needing to bundle NVIDIA drivers inside the container. To make a custom Docker container GPU-compatible, the Machine Learning Specialist should: Use NVIDIA CUDA and cuDNN in the Dockerfile. Ensure the container is built using the NVIDIA Container Toolkit (nvidia-docker). Use nvidia-container-runtime as the runtime.

To leverage the NVIDIA GPUs on Amazon EC2 P3 instances for training with Amazon SageMaker, the Docker container must be built to be compatible with NVIDIA-Docker. NVIDIA-Docker is a wrapper around Docker that makes it easier to use GPUs in containers by providing GPU-aware functionality. To build a Docker container that is compatible with NVIDIA-Docker, the Specialist should install the NVIDIA GPU drivers in the Docker container and install the NVIDIA-Docker runtime on the EC2 instances.

NVIDIA-Docker is a Docker container runtime plugin that allows the Docker container to access the GPU resources on the host machine. By building the Docker container to be NVIDIA-Docker compatible, the Docker container will have access to the NVIDIA GPU resources on the Amazon EC2 P3 instances, allowing for accelerated training of the ResNet model.

The reason for this choice is that NVIDIA-Docker is a tool that enables GPU-accelerated containers by automatically configuring the container runtime to use NVIDIA GPUs1. NVIDIA-Docker allows you to build and run Docker containers that can fully access the GPUs on your host system. This way, you can run GPU-intensive applications, such as deep learning frameworks, inside containers without any performance loss or compatibility issues.

A. NO - the drivers are not necessary (https://docs.aws.amazon.com/sagemaker/latest/dg/your-algorithms-training-algo-dockerfile.html) B. YES - it is about using the CUDA library, need to use proper base image (https://medium.com/@jgleeee/building-docker-images-that-require-nvidia-runtime-environment-1a23035a3a58) C. NO - file structure irrelavant to GPU D. NO - SageMaker config, irrelevant to Docker

https://docs.aws.amazon.com/sagemaker/latest/dg/sagemaker-dg.pdf page 55

Answer B Load the CUDA toolkit only, not the drivers. Ref GPU section : https://docs.aws.amazon.com/sagemaker/latest/dg/studio-byoi-specs.html

I think it should be b

B is correct!

As per aws documentation, answer is B, and A is even explicitly not recommended

As referred in other comments ans is B

ANS B As mentioned byi other users

As per me answer is B

The answer is for sure B - as mentioned by others. And this is clearly stated in the docs

Ans. is B.