OVHcloud GPU benchmark

As part of our commitment to helping customers optimize their AI workflows in the cloud, we have conducted a comprehensive benchmarking exercise to compare the performance of various cloud GPU VMs from top providers such as Amazon Web Services (AWS), Google Cloud, Microsoft Azure, and OVHcloud. Our goal is to provide actionable insights that can help developers and data scientists make informed decisions when selecting the right cloud infrastructure for their AI projects.

In this report, we present the results of our benchmarking exercise, which focused on evaluating the performance of various cloud GPU VMs from the aforementioned providers. We used a range of popular AI workloads, including LLM (Large Language Model) performance, Speech-to-Text processing with Whisper, and computer vision using YOLO (You Only Look Once). Our analysis compares the performance of different cloud GPU VMs across these workloads, highlighting strengths and weaknesses in each provider's offering.

By reading this project, you will gain a deeper understanding of the relative performance characteristics of various cloud GPU VMs and be better equipped to make informed decisions about your own AI projects. Whether you're a developer, data scientist, or IT professional, we hope that this project will provide valuable insights into the world of cloud-based AI computing.

Tooling

To make our methodology completly reproducible, we bas our GPU performance evaluation on several open-source tools specialized in their field:

• nvbandwidth: Throughput between host and GPU as GB/sec
• ollama-benchmark: LLM model performance as token/sec
• whisper-benchmark: Speech-to-text performance frame/sec
• yolo-benchmark: Computer vision performance inference latency

From these tools we elaborated different scenarii exploring the the deep learning capacity of NVIDIA hardware through model  sizes and precision format (FP16 and FP32).