Habitat: A Runtime-Based Computational Performance Predictor for Deep Neural Network Training
Published in USENIX Annual Technical Conference (ATC), 2021
Habitat is a runtime-based computational performance predictor for deep neural network training. It predicts a DNN’s training iteration execution time on a target GPU by leveraging execution time measurements from a source GPU. This enables users to make informed decisions about hardware selection for deep learning workloads without requiring expensive profiling on every target device.
Recommended citation: Geoffrey X. Yu, Yubo Gao, Pavel Golikov, Gennady Pekhimenko. (2021). "Habitat: A Runtime-Based Computational Performance Predictor for Deep Neural Network Training." USENIX ATC 2021. pp. 503-521.
Download Paper