Intelligent Delivery: Making Every Unit of Compute Serve Better Ads

What happened

Google has created a system called Intelligent Delivery that uses machine learning to decide how to distribute computing resources for ranking ads. The system has been in use since 2017 and is designed to improve the efficiency of ad delivery by dynamically allocating compute power where it is most effective. Intelligent Delivery predicts which ad requests will benefit most from extra computation and adjusts resources accordingly. This approach has led to a measurable increase in ads ranking performance across Google's platforms.

Key details

• Intelligent Delivery has delivered about a 10% improvement in ads ranking performance since 2017.
• The system uses deep learning models with 6 billion sparse parameters and 20 million dense parameters.
• It operates by predicting the marginal utility of compute for each ad request and reallocating resources in real time.
• The Resource Core Model is shared across the ad delivery funnel to inform allocation decisions.
• The system adapts to changes in user behavior, model updates, and hardware upgrades.

Why it matters

Advertisers and users of Google's apps and services are directly affected, as ads can be ranked more effectively without increasing hardware costs. Google can operate larger and more complex ad ranking models within existing compute limits, improving efficiency and potentially ad relevance. This reduces the need for additional hardware investment and allows for better use of available resources in data centers.