backend.ai

It's commonly believed that hyperparameter tuning requires a large number of GPUs to get quick, optimal results. It's generally true that higher computation power delivers more accurate results quickly, but to what extent? We'll present our work and empirical results on finding a sweet spot to balance both costs and accuracy by exploiting partitioned GPUs with Backend.AI's container-level GPU virtualization. Our benchmark includes distributed MNIST, CIFAR-10 transfer learning, and TGS salt identification cases using AutoML with network morphism and ENAS tuner with NNI running on Backend.AI's NGC-optimized containers. Attendees will get a tangible guide to deploy their GPU infrastructure capacity in a more cost-effective way.