Nielsen Sports, a global leader in audience insights, data, and analytics, has recently reduced operational and financial cost by 75% through modernizing its system, with the use of Amazon SageMaker multi-model endpoints (MMEs). This system involved a vast number of machine learning models in production.
Nielsen Sports’ technology uses artificial intelligence, particularly computer vision, to track brand exposure accurately. This can identify over 100,000 brands across numerous channels and has built a database of brand impressions with over 6 billion data points.
However, sustaining this scale of operation, with over a thousand channels and tens of thousands of hours of video annually analyzed through automated segments, presented significant cost and scaling challenges. The diverse models for each specific channel characteristic were expensive, prone to errors, and slow to upgrade or alter.
The previous system, using RabbitMQ for the pipeline, was labor-intensive to change, test, or maintain due to the inter-dependence between components. Notably, running a single model on a machine resulted in just 30-40% GPU utilization, with inefficient pipeline runs and scheduling algorithms identified.
The team at Nielsen Sports thus embarked on a journey to build a new multi-tenant architecture, based on SageMaker to support dynamic batch sizes, run multiple models at a time, and thus address performance optimization.
The result’s of this new architecture resulted in a 33% decrease in overall pipeline runtime; upgrading of infrastructure to the use of newer AWS instances with GPU’s, an increase in a single machine and it’s GPU utilization from 40% to more than 80%; increased agility and productivity with 75% improvement in planning; and improved accuracy in ML models with SageMaker’s A/B testing capabilities.
In conclusion, the modernization of Nielsen Sports system by the use of SageMaker MME’s resulted in a significant reduction in operational and financial cost whilst increasing efficiency, productivity, and accuracy.
