Explore a conference talk from OSDI '23 that introduces EINNET, a novel approach to optimizing tensor programs for deep neural networks. Delve into how EINNET leverages derivation-based transformations and automatically creates new operators to significantly expand the optimization space beyond traditional methods. Learn about the performance improvements achieved by EINNET, outperforming existing tensor program optimizers by up to 2.72x on NVIDIA A100 and 2.68x on NVIDIA V100 GPUs. Discover the potential impact of this research on boosting execution performance of DNNs in real-world applications. Access the publicly available EINNET implementation and gain insights into the future of tensor program optimization for machine learning and artificial intelligence.
EINNET - Optimizing Tensor Programs with Derivation-Based Transformations
Overview
Syllabus
OSDI '23 - EINNET: Optimizing Tensor Programs with Derivation-Based Transformations
Taught by
USENIX
Related Courses
-
PET - Optimizing Tensor Programs with Partially Equivalent Transformations and Automated Corrections
-
Ansor - Generating High-Performance Tensor Programs for Deep Learning
-
Hydro - Surrogate-Based Hyperparameter Tuning Service in Datacenters
-
Effectively Scheduling Computational Graphs of Deep Neural Networks toward Their Domain-Specific Accelerators
-
Rammer - Enabling Holistic Deep Learning Compiler Optimizations with rTasks
-
A Tensor Compiler Approach for One-size-fits-all ML Prediction Serving
