Streamlining ML Workflows: Lessons from Nielsen Sports' 75% Cost Reduction

A recent AWS Machine Learning blog post co-authored by Eitan Sela, a Generative AI and Machine Learning Specialist at AWS, highlights how Nielsen Sports achieved a remarkable 75% cost reduction in their video analysis workflows. They modernized their system, which runs thousands of different machine learning models, by leveraging Amazon SageMaker multi-model endpoints (MMEs) powered by the NVIDIA Triton Inference Server.

Nielsen Sports faced the challenge of scaling a massive computer vision system that identifies over 120 million brand impressions monthly across thousands of TV channels. Their legacy architecture suffered from low GPU utilization (30-40%) and a slow, cumbersome process for deploying new models, which could take over a month.

By re-architecting their system to use SageMaker MMEs, Nielsen was able to host multiple models on a single endpoint. This led to significant improvements:

• Dramatically Lower Costs: A 75% reduction in operational and financial costs.
• Increased Efficiency: GPU utilization soared to over 80%, and the overall pipeline runtime was cut by 33%.
• Greater Agility: The time to deploy new ML models plummeted from over a month to under a week.

This case study is a powerful example of how adopting a multi-tenant architecture with tools like SageMaker MMEs can lead to substantial cost savings, performance gains, and increased productivity for machine learning teams.

For a deeper dive into their technical solution and results, you can read the full post here: Nielsen Sports sees 75% cost reduction in video analysis with Amazon SageMaker multi-model endpoints.