Project Info
Accelerating Inference of Recurrent Neural Networks on GPUs
Bo Wu | bwu@mines.edu
Recurrent neural networks is widely used in domains, such as document understanding, machine translation, and sequence prediction. Although GPUs have shown their tremendous power in model training, inference of recurrent neural networks cannot efficiently utilize GPUs due to the limited parallelism. This project aims at designing a scheduling framework to accelerate inference on GPUs by more than 10 times.
More Information
http://www.wildml.com/2015/09/recurrent-neural-networks-tutorial-part-1-introduction-to-rnns/
https://www.nvidia.com/content/tegra/embedded-systems/pdf/jetson_tx1_whitepaper.pdf
Grand Engineering Challenge: Not applicable
Student Preparation
Qualifications
The student should be familiar with CUDA programming.
Time Commitment
60 hours/month
Skills/Techniques Gained
GPU programming
Deep learning
Scheduling
Locality optimization
Mentoring Plan
The student can attend our weekly group meetings. I’ll also have individual meetings with the student.