Click “Watch Now” to login or join the NVIDIA Developer Program.

WATCH NOW

CTR Inference Optimization on GPU

Chandler Zhou , NVIDIA | David Wu, NVIDIA

GTC 2020

The CTR (click-through-rate) model is one of the most important models in internet businesses such as search, recommendation, and advertising. The performance of CTR online service has become a critical impact on user experience/enterprise revenue. With the development of deep learning technology, the CTR model began to adopt deeper and more complex DNN structure, and the computation scale and complexity continued to rise, which demanded more computing power. With the evolution of the CTR model in recent years and the promotion of NVIDIA GPU computing platform, more and more companies have begun to use NVIDIA GPU to accelerate the CTR online inference model, and achieved significant acceleration and got commercial benefits. We'll introduce how to profile and locate the issues when doing optimization CTR inference model on NVIDIA GPU, and provide the general methods on how to solve the issues and get satisfying speedup.