NCP-AII Question 100
Select 3You are designing an AI infrastructure cluster to support both training and inference workloads. The cluster will utilize NVIDIA GPUs and requires a storage solution that can handle high throughput for training data and low latency for model inference. Which storage considerations should be prioritized to meet these requirements?
- A
Implement a high-bandwidth parallel file system to optimize data throughput for training workloads.
- B
Use object storage for inference to ensure high scalability and low latency.
- C
Deploy NVMe-based storage for frequently accessed training and inference datasets to minimize latency.
- D
Rely solely on traditional hard disk drives (HDDs) to maximize cost efficiency.
- E
Ensure the storage solution supports data caching to improve performance for repetitive data access.
Show answer and explanation
Correct answers: A, C, E
Explanation
AI workloads, especially in training and inference, require a storage solution designed for both high throughput and low latency. A high-bandwidth parallel file system ensures sufficient data flow for training, while NVMe storage and caching address the latency requirements for both training and inference. Object storage and HDDs, while beneficial in specific use cases, do not meet the performance demands of AI workloads in this scenario.
- A. Correct.
A high-bandwidth parallel file system, such as Lustre or GPFS, is essential for training workloads that require massive data throughput to feed GPUs at scale.
- B. Incorrect.
Object storage, while scalable, is not typically optimized for low-latency access, making it unsuitable for inference workloads that demand quick response times.
- C. Correct.
NVMe-based storage provides the low latency and high-speed data access necessary for both training and inference, especially for frequently accessed datasets.
- D. Incorrect.
Traditional HDDs, while cost-efficient, do not meet the performance requirements for AI workloads, particularly in high-throughput or low-latency scenarios.
- E. Correct.
Data caching enables faster access to frequently used data, reducing latency and improving overall storage performance, which is critical for AI workloads.