Using cutlass to compare cloud based GPU offerings

[bra-fsn]

Hello,

I’m looking to benchmark cloud-available GPUs across different platforms (e.g., AWS, GCP) for common use cases like machine learning, deep learning inference etc. I’m considering using Cutlass for this purpose, but I’m unsure about its suitability for benchmarking in these environments.

• Is Cutlass a suitable tool for general benchmarking tasks, especially for cloud-based GPUs (e.g., comparing AWS A10 GPUs with GCP A10 GPUs or different types of GPUs in different instance types, like to see how an L4 GPU performs in different AWS instances, considering thermal/power/bandwidth limitations)?
• What specific tests and parameters would you recommend running to generate meaningful and relevant comparisons for the most common GPU use cases?
• Are there any pre-defined configurations in Cutlass that would work for this type of benchmarking, or would it be necessary to define custom benchmarks?

Thank you in advance for your help!