Overview
Explore a groundbreaking approach to optimizing Deep Neural Network (DNN) computation on hardware accelerators in this 17-minute conference talk from OSDI '20. Dive into the innovative Rammer compiler design, which generates efficient static spatio-temporal schedules for DNNs at compile time, minimizing scheduling overhead and maximizing hardware utilization. Learn how Rammer employs novel, hardware-neutral abstractions for computation tasks and accelerators, enabling holistic exploitation of parallelism through inter- and intra-operator co-scheduling. Discover the significant performance improvements achieved by Rammer over state-of-the-art compilers like TensorFlow XLA and TVM, as well as vendor-optimized libraries like NVIDIA's TensorRT. Gain insights into Rammer's implementation across multiple hardware backends, including NVIDIA GPUs, AMD GPUs, and Graphcore IPU, and understand its potential impact on efficient DNN execution in massively parallel environments.
Syllabus
OSDI '20 - Rammer: Enabling Holistic Deep Learning Compiler Optimizations with rTasks
Taught by
USENIX