Enhancing Neural Network Training at Yelp: Achieving 1,400x Speedup with WideAndDeep

Yelp drastically reduced ad revenue prediction model training time by optimizing data handling and distributed training. They developed ArrowStreamServer for efficient S3 data streaming and adopted Horovod for multi-GPU distributed training. These improvements led to a 1400x speedup, processing 2 billion samples in under 1 hour per epoch instead of 75 hours. This approach enhances scalability, reduces costs, and improves developer productivity for Yelp's machine learning processes.